<?xml version="1.0" encoding="UTF-8"?><rss xmlns:dc="http://purl.org/dc/elements/1.1/" xmlns:content="http://purl.org/rss/1.0/modules/content/" xmlns:atom="http://www.w3.org/2005/Atom" version="2.0" xmlns:cc="http://cyber.law.harvard.edu/rss/creativeCommonsRssModule.html">
    <channel>
        <title><![CDATA[Stories by Parth Doshi on Medium]]></title>
        <description><![CDATA[Stories by Parth Doshi on Medium]]></description>
        <link>https://medium.com/@doshiparth2319?source=rss-be40dbc0ea3a------2</link>
        <image>
            <url>https://cdn-images-1.medium.com/fit/c/150/150/1*CKRc0xRlrp_d5vh2JrTIow.jpeg</url>
            <title>Stories by Parth Doshi on Medium</title>
            <link>https://medium.com/@doshiparth2319?source=rss-be40dbc0ea3a------2</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Thu, 19 Nov 2026 23:18:24 GMT</lastBuildDate>
        <atom:link href="https://medium.com/@doshiparth2319/feed" rel="self" type="application/rss+xml"/>
        <webMaster><![CDATA[yourfriends@medium.com]]></webMaster>
        <atom:link href="http://medium.superfeedr.com/" rel="hub"/>
        <item>
            <title><![CDATA[Your Genome Sequencing Lab is Leaving Diagnoses on the Table — Here’s the Math]]></title>
            <link>https://medium.com/@doshiparth2319/your-genome-sequencing-lab-is-leaving-diagnoses-on-the-table-heres-the-math-a5617962bd75?source=rss-be40dbc0ea3a------2</link>
            <guid isPermaLink="false">https://medium.com/p/a5617962bd75</guid>
            <category><![CDATA[genomics]]></category>
            <category><![CDATA[precision-medicine]]></category>
            <category><![CDATA[rare-disease]]></category>
            <category><![CDATA[clinical-diagnostic]]></category>
            <category><![CDATA[bioinformatics]]></category>
            <dc:creator><![CDATA[Parth Doshi]]></dc:creator>
            <pubDate>Wed, 22 Apr 2026 06:33:58 GMT</pubDate>
            <atom:updated>2026-04-22T06:33:58.042Z</atom:updated>
            <content:encoded><![CDATA[<p>For every 100 patients you sequence and can’t diagnose, 5 to 15 of them already have the answer sitting in your pipeline. You just have not found it yet.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*hNtOVeD_4wsZSVQhwYnl4g.png" /></figure><p>Let’s start with a number that should bother anyone running a clinical genomics lab: 34.2%.</p><p>That’s the pooled diagnostic yield of genome-wide sequencing across 24,631 pediatric rare-disease patients in the largest meta-analysis to date (Pandey et al., Genetics in Medicine, 2025). Over six in ten patients go home without a diagnosis.</p><p>Some of that gap is irreducible. Science has not yet identified every disease gene. Some variants are genuinely uninterpretable today.</p><p>But a significant, recoverable slice of that gap? It’s not about the biology. It’s about the decisions your lab is making or not making after the sequencer stops running.</p><h3>The 5–14% Nobody’s Talking About</h3><p>Multiple research programs have now demonstrated something remarkable: you can diagnose an additional 5 to 14% of your previously unsolved patients without resequencing them: no new blood draw, no new library prep, no new instrument time.</p><p>Just better pipelines, fresher annotation databases, and a systematic reanalysis program.</p><p>Here’s what that looks like at scale.</p><p>The Solve-RD consortium reanalysed 6,004 families across 37 European centres, cases that expert clinicians had already declared unsolved. They recovered 506 new diagnoses, an 8.4% uplift, by doing nothing more than reanalysing data that already existed. Of the disease-causing variants they found:</p><p>34% had been reclassified in ClinVar after the original report was issued</p><p>38% were reclassified through expert consensus during reanalysis itself</p><p>Only 12% came from recently published disease genes</p><p>Read that again: the majority of new diagnoses didn’t require new science. They required updated databases and a second look.</p><p>Ambry Genetics ran a similar experiment on 10,921 exome patients. Their lab-initiated reanalysis program produced reclassifications in 5% of initially negative cases — but the operational finding was even more striking. When labs proactively initiated reanalysis, reclassification occurred in 54% of flagged cases. When did clinicians initiate it? Just 4%.</p><p>Why? Because clinician-initiated reanalysis rarely happens. Before Ambry’s program existed, only 13.6% of eligible patients ever received a reanalysis request.</p><p>The diagnoses weren’t being missed because the science was hard. They were being missed because no one was looking.</p><h3>The Three Levers (and Why Most Labs Aren’t Pulling Them)</h3><p>The diagnostic yield gap has three drivers, all of which are fixable.</p><h4><strong>1. Your pipeline is from 2022</strong></h4><p>Variant callers have improved substantially. The most recent Genome-in-a-Bottle benchmarks show DRAGEN v4.2 achieving an F1 score of 99.86% versus 99.13% for GATK+BWA, a gap that compounds across thousands of cases. Trio-sequencing studies place DeepVariant’s Mendelian error rate at 3.09% versus GATK’s 5.25%.</p><p>None of that is decisive on its own. But labs frozen on an older calling stack are leaving marginal accuracy on the table across every single case they process.</p><p>More dramatically: if your pipeline is SNV/indel only, you’re structurally undersized for a significant minority of your cases. In the Oxford OxClinWGS rare disease cohort, structural, splice-site, and deep-intronic variants contributed 43% of solved cases when a comprehensive pipeline was applied. The Hong Kong Genome Project found that one in five diagnoses would have been missed without comprehensive short-read bioinformatics, repeat expansions, big intronic changes, and pseudogene-confounded variants.</p><p>Your sequencer already generated those reads. Your pipeline just isn’t looking at them.</p><h4>2. Your annotation databases are stale</h4><p>ClinVar releases monthly. gnomAD released v4 in late 2023. Newer in silico predictors, such as AlphaMissense, which classifies 89% of all possible human missense variants as likely benign or likely pathogenic, are now deployable in clinical workflows, with evidence weights up to “Strong” for pathogenicity under the emerging ACMG v4 framework.</p><p>A pipeline that snapshots annotations at deployment and runs untouched for 24 months is making clinical decisions based on evidence that has since moved.</p><p>The practical implication: a lab updating annotation quarterly and re-running its predictor suite against unsolved cases is mechanically eligible to reclassify variants that couldn’t be reclassified 12 months ago, without any wet-lab work.</p><p>One specific example: a 2025 medRxiv analysis found that rerunning SpliceAI against current MANE Select transcripts yielded an 18.2% increase in predicted splice-altering variants and an 11.7% increase in retrospective diagnostic yield compared with precomputed scores. The algorithm was the same. The annotation provenance was different. That was enough.</p><h4>3. You’re not reanalysing systematically</h4><p>The Deciphering Developmental Disorders study followed 13,449 probands with iterative reanalysis over roughly a decade. Initial yield: 27%. Cumulative yield after multi-round reanalysis: 41%. Fourteen percentage points recovered through reanalysis alone.</p><p>Radboud University followed 150 pediatric neurology patients for five years: the yield went from 31% to 53%. Of the new diagnoses, 19% came from improved bioinformatics, 10% from variant reclassification, 29% from resequencing, and 42% from new disease-gene associations.</p><p>The pattern across programs is consistent: the floor for a properly run reanalysis program is ~5% additional diagnoses; the ceiling on long-unsolved cohorts is 20–25%.</p><h3>What This Looks Like for a Typical Lab</h3><p>Say your lab runs 1,000 probands per year at 30% initial yield: 300 diagnoses, 700 unsolved.</p><p>A systematic reanalysis program applied to those 700 cases, at the conservative end of published effect sizes, recovers 35 to 105 additional diagnoses per year, with a central estimate of around 70.</p><p>The cost? Bottom-up estimates from published program data put a reanalysis program at roughly US$650,000–720,000 per year in FTE and infrastructure curators, a bioinformatician, computing, and software licensing. That works out to approximately US$9,800 per additional diagnosis at the central estimate.</p><p>Compare that to initial sequencing: US$15,000–27,000 per diagnosis, depending on cohort and modality (Lavelle et al., 2022).</p><p>Ewans et al. (2018) went further: a 12-month WES reanalysis in an intellectual disability cohort was not just cheaper per diagnosis; it was cost-saving overall, at a negative US$586 per additional diagnosis. The marginal cost of finding what was already sequenced is trivial compared to generating new data.</p><p><strong>Four Programs That Have Closed the Gap</strong></p><p>Talos (Centre for Population Genomics, Australia) is an open-source reanalysis engine with an elegant architectural trick: it separates expensive one-time annotation (stored in a Hail MatrixTable) from cheap, frequent filtering (updated monthly against ClinVar, PanelApp, and patient HPO terms). Applied to 4,735 previously undiagnosed individuals, Talos produced 248 new diagnoses (5.2%), 29% from new gene-disease relationships, 23% from new variant-level evidence, 48% from improved filtering alone.</p><p>Ambry Patient-for-Life takes the lab-initiated model to commercial scale. Clinical genomic scientists review the literature daily. When a gene reaches sufficient disease validity, every previously tested proband with a surviving variant in that gene gets automatically re-evaluated. Lab-initiated. No clinician action required. No patient cost.</p><p>Boston Children’s Hospital bets on EHR integration: NLP extracts HPO terms from clinical notes monthly, yielding ~52 terms per proband versus ~5 from manual curation and feeds them into variant prioritisation. The result is a phenotype signal that evolves with the patient, not just at the time of initial testing.</p><p>The Indian Undiagnosed Diseases Program demonstrated what reanalysis-first looks like in resource-constrained settings: of 88 patients, 27.2% were diagnosed through reanalysis before any add-on testing was authorised, novel gene associations, synonymous variants, and deep-intronic changes, all from data that already existed.</p><h3>What Labs Should Do This Quarter</h3><p>The research base now supports six concrete actions, ranked by ease and impact:</p><p>Audit your annotation cadence. If your ClinVar, gnomAD, and predictor databases aren’t refreshed at least quarterly, you’re making clinical decisions based on stale evidence. Solve-RD’s data show 34% of newly solved variants came from post-report ClinVar reclassifications alone.</p><p>Set up knowledge-event triggers. Annual reanalysis misses more than half of all reclassifications, according to Ambry’s data. Subscribe to ClinVar monthly releases and ClinGen gene-disease validity updates, and flag affected unsolved cases automatically.</p><p>Deploy SV, repeat expansion, and mitochondrial callers on your existing short-read data. The reads are already there. If your pipeline is SNV/indel only, you’re missing one in five diagnoses in some cohorts.</p><p>Version and containerise your pipeline. You can’t credibly reanalyse a case against a new annotation unless you can reproduce the original pipeline state. Nextflow with pinned Singularity containers and Zenodo DOIs is the current community standard.</p><p>Start preparing for ACMG v4. Draft release is expected mid-2026. The framework shifts to a points-based Bayesian model, introduces VUS subtiers, and elevates computational predictor evidence to Strong/Moderate strength. Labs that wait for the publication to start adapting will be behind immediately.</p><p>Build the cost model for your CFO. A 1,000-proband lab recovering 70 additional diagnoses per year at an incremental cost of US$700K is spending roughly US$10,000 per additional diagnosis, one-third to two-thirds the cost of initial sequencing, with documented downstream healthcare savings of US$14,000–100,000 per diagnosed ICU patient.</p><h3>The Point</h3><p>The diagnostic yield gap is real, measurable, and only partly explained by the limits of current science. A large slice of it is explained by decisions that clinical labs make about pipelines, annotation, and reanalysis — decisions that are entirely within their control to change.</p><p>The patients have already been sequenced. The reads already exist. For a meaningful fraction of your unsolved cases, the answer is sitting in data you already have.</p><p>You just haven’t looked yet.</p><p><strong>REFERENCE</strong></p><p>Ewans, L. J., Schofield, D., Shrestha, R., Zhu, Y., Gayevskiy, V., Ying, K., Walsh, C., Colley, A., Ellaway, C. J., Turner, A., Mowat, D., Worgan, L., Freckmann, M. L., Lipke, M., Sachdev, R., Miller, D., Dinger, M. E., &amp; Cowley, M. J. (2018). Whole-exome sequencing reanalysis at 12 months improves diagnostic yield and is cost-effective when applied early in Mendelian disorders. <em>Genetics in Medicine</em>, <em>20</em>(12), 1564–1574. <a href="https://doi.org/10.1038/gim.2018.39">https://doi.org/10.1038/gim.2018.39</a></p><p>Garg, S., Singh, S. M., et al. (2025). Reanalysis of exome sequencing data in the Indian Undiagnosed Diseases Program. <em>Clinical Genetics</em>. Advance online publication. <a href="https://doi.org/10.1111/cge.14700">https://doi.org/10.1111/cge.14700</a></p><p>Laurie, S., Matalonga, L., Hoischen, A., Beltran, S., et al. (2024). Genomic reanalysis of a pan-European rare-disease resource yields new diagnoses. <em>Nature Medicine</em>, <em>30</em>, 1796–1807. <a href="https://doi.org/10.1038/s41591-024-02966-x">https://doi.org/10.1038/s41591-024-02966-x</a></p><p>Lavelle, T. A., Feng, X., Keefe, K., Cohen, J. T., &amp; Neumann, P. J. (2022). Cost-effectiveness of exome and genome sequencing for children with rare and undiagnosed conditions. <em>Genetics in Medicine</em>, <em>24</em>(7), 1349–1361. <a href="https://doi.org/10.1016/j.gim.2022.03.005">https://doi.org/10.1016/j.gim.2022.03.005</a></p><p>Martin-Geary, A. C., et al. (2025). <em>Using SpliceAI to triage splice-altering variants in 7,220 individuals with rare conditions</em> [Preprint]. medRxiv. <a href="https://doi.org/10.1101/2025.08.27.25334471">https://doi.org/10.1101/2025.08.27.25334471</a></p><p>Pandey, R., et al. (2025). A meta-analysis of diagnostic yield and clinical utility of genome and exome sequencing in pediatric rare and undiagnosed genetic diseases. <em>Genetics in Medicine</em>, <em>27</em>, 101389. <a href="https://doi.org/10.1016/j.gim.2024.101389">https://doi.org/10.1016/j.gim.2024.101389</a></p><p>Schobers, G., Schieving, J. H., Yntema, H. G., Pfundt, R., Derks, R., Kwint, M., Bollen, S., Fleischer, K., Nelen, M., Scheffer, H., Brunner, H. G., van Zelst-Stams, W. A. G., Vissers, L. E. L. M., Veltman, J. A., &amp; Gilissen, C. (2022). Reanalysis of exome-negative patients rendered an additional diagnosis in 29% of cases at Radboud. <em>Genome Medicine</em>, <em>14</em>, Article 66. <a href="https://doi.org/10.1186/s13073-022-01069-z">https://doi.org/10.1186/s13073-022-01069-z</a></p><p>Towne, M. C., Huang, J., et al. (2025). Impact of laboratory-driven proactive reanalysis: Reclassification to positive in 5% of initially negative or uncertain exome sequencing cases (Ambry Patient-for-Life). <em>Genetics in Medicine</em>, <em>27</em>, 101464. <a href="https://doi.org/10.1016/j.gim.2024.101464">https://doi.org/10.1016/j.gim.2024.101464</a></p><p>Welland, M. J., Ahlquist, K. D., et al. (2025). <em>Scalable automated reanalysis of genomic data in research and clinical rare disease cohorts (Talos)</em> [Preprint]. medRxiv. <a href="https://doi.org/10.1101/2025.05.19.25327921">https://doi.org/10.1101/2025.05.19.25327921</a></p><p>Wright, C. F., et al. (2023). Genomic diagnosis of rare pediatric disease in the United Kingdom and Ireland. <em>New England Journal of Medicine</em>, <em>388</em>(17), 1559–1571. <a href="https://doi.org/10.1056/NEJMoa2209046">https://doi.org/10.1056/NEJMoa2209046</a></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=a5617962bd75" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[10,000 Cells Don’t Replace 3 Patients]]></title>
            <link>https://medium.com/@doshiparth2319/10-000-cells-dont-replace-3-patients-9643e888be07?source=rss-be40dbc0ea3a------2</link>
            <guid isPermaLink="false">https://medium.com/p/9643e888be07</guid>
            <category><![CDATA[bioinformatics]]></category>
            <category><![CDATA[data-science]]></category>
            <category><![CDATA[genomics]]></category>
            <category><![CDATA[biology]]></category>
            <category><![CDATA[biotechnology]]></category>
            <dc:creator><![CDATA[Parth Doshi]]></dc:creator>
            <pubDate>Tue, 17 Mar 2026 10:59:47 GMT</pubDate>
            <atom:updated>2026-03-17T10:59:47.182Z</atom:updated>
            <content:encoded><![CDATA[<h4>The pseudoreplication trap quietly inflating false discoveries in single-cell RNA-seq — and how rigorous teams avoid it</h4><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*LNGSRSpyif_Jr-3Qk5aotA.png" /></figure><p>Picture this.</p><p>A research team profiles 50,000 cells from six patients: three with a disease, three healthy controls.</p><p>They run differential expression. Hundreds of “significant” genes pour out. P-values plunge into the 10⁻²⁰ range. The dataset looks powerful. The conclusions feel solid.</p><p>Then someone asks the only question that really matters:</p><p><strong>How many independent biological replicates did you actually have?</strong></p><p>Not 50,000.<br>Not 5,000.<br>Just three versus three.</p><p>Those 50,000 cells were never 50,000 independent observations. They were repeated measurements from six people.</p><p>And that distinction between <strong>measurements</strong> and <strong>replicates</strong> is the difference between a finding and a mirage.</p><p>This is the pseudoreplication trap in single-cell RNA-seq. It sounds like a technical statistical issue. It isn’t. It is a basic error in experimental reasoning, and it has quietly distorted a large share of published single-cell differential expression results.</p><p>For labs doing exploratory work, that means inflated gene lists and shaky narratives. For biotech teams using single-cell RNA-seq for biomarkers, target selection, or patient stratification, it means something more serious: building decisions on signals that may never have been real.</p><blockquote><strong><em>Single-cell technology increases resolution. It does not increase biological replication.</em></strong></blockquote><h3>The mistake is old. Single-cell just made it easier to commit.</h3><p>Pseudoreplication long predates genomics.</p><p>In 1984, ecologist Stuart Hurlbert gave it a durable definition: using inferential statistics to test for treatment effects when the data points are not actually independent. The problem showed up all over ecology. Later reviews found it in neuroscience, too. By 2020, one review put it bluntly: in some fields, pseudoreplication is one of the biggest threats to valid inference.</p><p>Single-cell genomics did not solve that problem. It inherited it and supercharged it.</p><p>In older experimental settings, researchers might accidentally treat dozens of repeated observations as if they were independent. In single-cell RNA-seq, the same mistake can transform a study with six donors into a matrix with tens of thousands of rows. The scale changes. The logic does not.</p><p>And that is exactly why the error becomes so seductive.</p><p>The output looks rich. The p-values look spectacular. The plots look convincing.</p><p>But none of that changes the unit of replication.</p><h3>Why are cells not patients</h3><p>This is the core idea, and it is simpler than many analyses make it sound.</p><p>When you compare disease versus control in single-cell RNA-seq, you are usually asking a population-level question:</p><p><strong>Does this condition change gene expression in a given cell type?</strong></p><p>To answer that, you need independent biological samples from the population. In most studies, those samples are <strong>patients</strong> or <strong>donors</strong>.</p><p>What you actually observe, though, has a nested structure:</p><ul><li>Cells are nested within donors</li><li>donors are nested within conditions</li></ul><p>That means cells from the same donor are not statistically independent. They share the same genome, age, medications, immune history, tissue context, sample handling, and batch effects. Even when they vary, they vary within the envelope of one person’s biology.</p><p>So the real biological replicates are the donors. The measurements are taken from those donors.</p><p>That is not a philosophical distinction. It is the entire statistical problem.</p><p>Treat cells as independent, and you create a fake sample size. The model thinks it has thousands of observations supporting a difference that may actually rest on just a handful of people.</p><p>A Wilcoxon test or t-test that sees 5,000 cells per condition behaves as if the degrees of freedom are enormous. But if the study really has five donors per group, the meaningful degrees of freedom are still tiny. That mismatch is how modest, donor-specific quirks get translated into ultra-small p-values and long lists of “discoveries.”</p><p>You have not necessarily measured a stronger biological effect. You may have just measured the same biological background many more times and allowed the statistics to confuse repetition with replication.</p><blockquote><strong><em>The cells are the ruler. The patients are the experiment.</em></strong></blockquote><h3>What the benchmarking studies found</h3><p>By 2021, the field had a serious reality check.</p><p>A team led by Jordan Squair published what is now one of the defining analyses of this problem in <em>Nature Communications</em>. They surveyed hundreds of single-cell RNA-seq papers and found that among studies performing differential expression across biological conditions, roughly 90 percent relied on methods prone to false discoveries, especially cell-level tests like Wilcoxon rank-sum, often run through default workflows.</p><p>That finding alone was uncomfortable. But the real force of the paper came from what happened next.</p><p>The authors compared methods against known or better-supported signal structures, including matched bulk RNA-seq benchmarks. The best-performing methods were pseudobulk approaches. Cell-level methods did substantially worse.</p><p>Then came the null experiment, the kind of test that cuts through abstraction.</p><p>Take control samples. Randomly split them into a fake “case” group and a fake “control” group. There is no biological difference, by design.</p><p>Now run cell-level differential expression anyway.</p><p>What happens?</p><p>In many datasets, you still get hundreds of “significant” genes. You can run pathway enrichment and generate a plausible-looking story. You can produce heatmaps, volcano plots, and mechanistic interpretations. Everything appears coherent.</p><p>It is also false.</p><p>That was the disturbing point. Pseudoreplication does not just create a few extra false positives around the margins. It can generate an entire fake narrative with the full aesthetic of a real result.</p><p>And the effect does not get better as you add more cells. Often it gets worse. More cells from the same donors can make the model even more overconfident.</p><p>What helps instead is more biological replication.</p><p>More donors.<br>Better design.<br>Correct models.</p><h3>The Alzheimer’s reanalysis that made the stakes impossible to ignore</h3><p>If the benchmarking papers made the methodological case, the Alzheimer’s example made the consequences visceral.</p><p>In 2019, a major single-cell study profiled around 80,000 nuclei from the prefrontal cortex of 48 individuals, split between those with and without Alzheimer’s pathology. Using cell-level analysis, the study reported <strong>14,274 differentially expressed genes</strong> at FDR &lt; 0.05.</p><p>The paper was influential. It was widely cited. It shaped thinking about which cell types mattered most in Alzheimer’s disease.</p><p>Then came the reanalysis.</p><p>In 2023, Murphy, Fancy, and Skene revisited the same dataset using pseudobulk edgeR, a donor-aware approach aligned with the actual replication structure of the experiment.</p><p>They found <strong>26 differentially expressed genes</strong>.</p><p>Not 14,274.</p><p>Twenty-six.</p><p>That is not a minor calibration issue. That is a collapse in the inferred biology.</p><p>And the story changed more profoundly, too.</p><p>The original analysis placed much of the signal in neurons, the most abundant cell population in the dataset. The pseudobulk reanalysis found that the robust signal was concentrated overwhelmingly in microglia, a result far more consistent with genetic and immunological evidence about Alzheimer’s disease.</p><p>That is what makes pseudoreplication so dangerous.</p><p>It does not merely inflate significance. It can redirect attention toward the wrong cell types, the wrong pathways, and the wrong therapeutic hypotheses.</p><p>A bad DEG list is not just untidy. In the wrong setting, it becomes a strategy.</p><h3>Why does this keep happening</h3><p>The uncomfortable answer is not ignorance. It is workflow gravity.</p><p>The most widely used single-cell toolkits made cell-level testing easy, intuitive, and fast. In Seurat, FindMarkers(), a straightforward path to differential expression was popularised, often using the Wilcoxon rank-sum by default. In Scanpy, similar ranking-based workflows became standard. For many users, these were the first methods they learned.</p><p>And for some tasks, that is perfectly fine.</p><p>If you are identifying marker genes to help define cell clusters within a dataset, cell-level testing can be reasonable as a descriptive tool. The trouble starts when researchers carry that same workflow into <strong>cross-condition biological inference</strong>, disease versus control, treated versus untreated, responder versus non-responder, without accounting for donor-level structure.</p><p>That is where the default path becomes a trap.</p><p>The methods for doing it correctly do exist. Pseudobulk aggregation is well established. Mixed models with donor random effects are well described. Tools like edgeR, DESeq2, limma-voom, muscat, NEBULA, and donor-aware uses of MAST are all part of the conversation.</p><p>But defaults matter.</p><p>Most people do not discover methodology through benchmark papers. They discover it through tutorials, package examples, and whatever gets them to a figure fastest.</p><p>If the default workflow is convenient and the wrong answer looks exciting, the field will generate a lot of exciting wrong answers.</p><h3>What rigorous analysis actually looks like</h3><p>The fix is not mysterious. It just requires discipline.</p><p>The most widely recommended approach for cross-condition single-cell differential expression is <strong>pseudobulk aggregation</strong>.</p><p>Instead of treating every cell as its own replicate, pseudobulk methods aggregate counts within each donor and cell type. If you have 8 donors and 5 cell types, you end up with up to 40 pseudobulk samples: one donor-by-cell-type expression profile per combination.</p><p>That changes the statistical unit back to where it belongs.</p><p>Now each sample reflects a true biological replicate. Established bulk RNA-seq tools such as <strong>edgeR</strong>, <strong>DESeq2</strong>, or <strong>limma-voom</strong> can estimate variance and perform inference in a way that matches the study design.</p><p>The result is less theatrical but far more trustworthy:</p><ul><li>realistic variance estimates</li><li>appropriate degrees of freedom</li><li>calibrated p-values</li><li>fewer false positives</li><li>a DEG list that has a fighting chance of surviving validation</li></ul><p>For some designs, mixed models are also appropriate. If a cell population is rare, extremely sparse, or the structure of the experiment makes aggregation less attractive, donor-aware mixed models can help. Methods like <strong>NEBULA</strong> and random-effect implementations of <strong>MAST</strong> can account explicitly for the fact that cells from the same donor are correlated.</p><p>But the principle never changes:</p><p><strong>The donor must appear in the model</strong></p><p>If donor identity is absent from a cross-condition single-cell DE analysis, skepticism is warranted.</p><h3>Replication beats depth</h3><p>This point deserves to be said plainly because single-cell studies often blur it.</p><p>Sequencing more cells is not the same as adding more biological replicates.</p><p>For differential expression, depth is useful. Coverage matters. Capturing rare cell states matters. But biological replication is what gives you inferential leverage across conditions.</p><p>As a rule of thumb:</p><p>Two biological replicates per condition may be enough to estimate dispersion.<br>Three is the practical floor.<br>Five or more is where power begins to look credible.<br>Beyond that, better replication usually does more for inference than simply pushing cell counts higher.</p><p>So this tradeoff is not neutral:</p><p><strong>10,000 cells each from two donors per group</strong><br>is often worse for inference than<br><strong>1,000 cells each from ten donors per group</strong></p><p>That sounds counterintuitive only if you are still thinking like the cell count is the sample size.</p><p>It isn’t.</p><p>For cross-condition claims, donor count is the number that matters most.</p><h3>Why this matters beyond academic correctness</h3><p>It is tempting to treat pseudoreplication as a concern for statisticians and reviewers. It is not.</p><p>If your team is using single-cell RNA-seq to nominate biomarkers, build signatures, prioritise targets, or stratify patients, the downstream consequences are real.</p><p>A false-positive DEG list does not stay inside a figure legend.</p><p>It gets turned into validation assays.<br>It shapes target meetings.<br>It influences what goes into animal work.<br>It affects which hypotheses attract internal support.<br>Sometimes it helps determine who gets enrolled in a trial.</p><p>In that context, an inflated p-value is not just a math problem. It is a resource-allocation problem.</p><p>The Alzheimer’s reanalysis is a dramatic example because the numbers were so extreme. But the same mechanism operates quietly in far more ordinary studies. Every time a cell-level test is run across a small number of donors without accounting for donor structure, the resulting gene list should be read with caution.</p><p>The right question is not, “<strong>How small is the p-value?</strong>”</p><p>It is, “<strong>What was the true unit of replication, and did the model respect it?</strong>”</p><p>That question will tell you more than the volcano plot ever will.</p><h3>A 30-second sanity check for any single-cell DEG claim</h3><p>Before trusting a differential expression result from single-cell RNA-seq, ask:</p><ul><li>How many donors were there per condition?</li><li>Were cells treated as independent observations?</li><li>Was a pseudobulk or a donor-aware mixed model used?</li><li>Do the reported p-values make sense given the number of biological replicates?</li><li>Are the strongest signals concentrated in the most abundant cell types for reasons that might reflect cell count rather than biology?</li></ul><p>You do not need to be a method developer to ask those questions.</p><p>You just need to remember what the experiment actually was.</p><h3>The line worth remembering</h3><p>Single-cell RNA-seq is a remarkable technology.</p><p>It reveals heterogeneity that bulk RNA-seq washes out. It separates cell types, tracks states, and exposes structure inside tissues that older methods could not resolve. That is real progress.</p><p>But resolution is not replication.</p><p>When the question is whether biology differs <strong>between patients</strong>, <strong>between conditions</strong>, or <strong>between disease and health</strong>, the number that matters is still the number of independent donors.</p><p>The cells are the ruler.<br>The patients are the experiment.</p><p>Confuse those two things, and you do not get a more powerful study.</p><p>You get a more confident mistake.</p><p>And for any lab, platform team, or biotech company building on single-cell differential expression, that is not a statistical footnote.</p><p>It is due diligence.</p><h3>References</h3><p>Crowell, H. L., Soneson, C., Germain, P.-L., Calini, D., Collin, L., Raposo, C., Malhotra, D., &amp; Robinson, M. D. (2020). muscat detects subpopulation-specific state transitions from multi-sample multi-condition single-cell transcriptomics data. <em>Nature Communications</em>, <em>11</em>, Article 6077. <a href="https://doi.org/10.1038/s41467-020-19894-4">https://doi.org/10.1038/s41467-020-19894-4</a></p><p>Murphy, A. E., Fancy, N., &amp; Skene, N. G. (2023). Avoiding false discoveries in single-cell RNA-seq by revisiting the first Alzheimer’s disease dataset. <em>eLife</em>, <em>12</em>, Article RP90214. <a href="https://doi.org/10.7554/eLife.90214">https://doi.org/10.7554/eLife.90214</a></p><p>Murphy, A. E., &amp; Skene, N. G. (2022). A balanced measure shows superior performance of pseudobulk methods in single-cell RNA-sequencing analysis. <em>Nature Communications</em>, <em>13</em>, Article 7851. <a href="https://doi.org/10.1038/s41467-022-35519-4">https://doi.org/10.1038/s41467-022-35519-4</a></p><p>Squair, J. W., Gautier, M., Kathe, C., Anderson, M. A., James, N. D., Hutson, T. H., Hudelle, R., Qaiser, T., Matson, K. J. E., Barraud, Q., Levine, A. J., La Manno, G., Skinnider, M. A., &amp; Courtine, G. (2021). Confronting false discoveries in single-cell differential expression. <em>Nature Communications</em>, <em>12</em>, Article 5692. <a href="https://doi.org/10.1038/s41467-021-25960-2">https://doi.org/10.1038/s41467-021-25960-2</a></p><p>Zimmerman, K. D., Espeland, M. A., &amp; Langefeld, C. D. (2021). A practical solution to pseudoreplication bias in single-cell studies. <em>Nature Communications</em>, <em>12</em>, Article 738. <a href="https://doi.org/10.1038/s41467-021-21038-1">https://doi.org/10.1038/s41467-021-21038-1</a></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=9643e888be07" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Why Your RNA-seq Pathway Plots Aren’t Validated (And What To Do Instead)]]></title>
            <link>https://medium.com/@doshiparth2319/why-your-rna-seq-pathway-plots-arent-validated-and-what-to-do-instead-61cd2fb72d9a?source=rss-be40dbc0ea3a------2</link>
            <guid isPermaLink="false">https://medium.com/p/61cd2fb72d9a</guid>
            <category><![CDATA[rna-sequencing]]></category>
            <category><![CDATA[data-science]]></category>
            <category><![CDATA[computational-biology]]></category>
            <category><![CDATA[bioinformatics]]></category>
            <category><![CDATA[genomics]]></category>
            <dc:creator><![CDATA[Parth Doshi]]></dc:creator>
            <pubDate>Fri, 13 Mar 2026 07:42:30 GMT</pubDate>
            <atom:updated>2026-03-13T07:42:30.069Z</atom:updated>
            <content:encoded><![CDATA[<h4>How database overlap, expression bias, and background selection distort enrichment results</h4><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*q5wEsLUtcpmgpIYzZjqOQQ.png" /></figure><p>There’s a familiar rhythm to many RNA-seq studies.</p><p>First comes differential expression. Then comes pathway enrichment. A few pathways light up, the plots look convincing, and suddenly the biological story feels reinforced. The paper writes itself.</p><p>But that feeling of confirmation is often doing more rhetorical work than scientific work.</p><p>Pathway enrichment can be extremely useful for organising gene lists into interpretable biological themes. What it cannot do is independently verify the underlying result. It is still derived from the same experiment, the same gene list, and the same biases that shaped the original output.</p><blockquote><strong>Pathway enrichment is best treated as structured hypothesis generation, not biological validation</strong>.</blockquote><h3>The analytical trap hidden in plain sight.</h3><p>The logic seems intuitive: genes change, those genes cluster inside known pathways, therefore, the biology must be real.</p><p>That is exactly the trap.</p><p>Methods like ORA, GSEA, and other rank-based enrichment frameworks all test the same basic question: do genes from a predefined pathway appear more often than expected in your experimental output?</p><p>That sounds rigorous, but it is still a comparison between your gene list and a curated database. It is not an independent line of evidence. It is a statistical interpretation layer built on top of the same underlying data — filtered through assumptions you may not have fully examined.</p><h3>The redundancy problem hiding in plain sight</h3><p>One reason pathway results can look more robust than they are is that pathway databases are heavily redundant. Many pathways share genes, sometimes extensively.</p><p>Cell cycle, DNA replication, and mitotic checkpoint pathways, for example, can all be driven by overlapping gene membership. When a handful of genes shift strongly, they may trigger several related pathway hits at once — creating the impression of multiple independent biological signals when it is really one repeated signal under different labels.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*V9_yMROCps6_6qPxYZmj7Q.png" /><figcaption><em>One signal, many pathways. Database overlap inflates apparent biological confirmation. A few key genes drive enrichment across multiple related terms simultaneously.</em></figcaption></figure><p>Published analyses of database structure confirm this directly: more redundancy means more false positives and less specificity. A long list of significant pathways may not reflect broad biological activity. It may reflect database architecture.</p><h3>Bias enters before enrichment even begins.</h3><p>RNA-seq gene lists are not random samples of the genome. Some genes are simply more likely to appear as differentially expressed — for technical reasons that have nothing to do with your biology.</p><p><strong>Gene length bias</strong> is one of the clearest examples. Longer transcripts generate more reads, which gives them greater statistical power in differential expression testing. Correcting for this using standard normalisations like RPKM or TMM is not sufficient; sample-specific length effects can persist and produce false positive enrichment calls, particularly for gene sets dominated by long genes, such as extracellular matrix components, or short genes like ribosomal proteins.</p><p><strong>Expression-level bias</strong> compounds this: highly expressed genes are easier to detect and are overrepresented in curated pathway databases. Together, these effects can systematically favour certain pathways even when the signal is partly technical. The enrichment output looks biologically coherent while remaining distorted by the same artefacts that shaped the input.</p><h3>Small analytical choices can change the whole story</h3><p>Another weakness is instability.</p><p>ORA depends on thresholds. Change the FDR cutoff or fold-change boundary, and the set of enriched pathways can shift dramatically. GSEA avoids hard cutoffs but introduces its own sensitivities — including permutation randomness that can produce wildly different outputs from identical data. In documented cases, the same input files run through GSEA in multiple rounds produced 32, then 4, then 15, then 55 significantly enriched gene sets. That is not analytical noise at the margins. That is a fundamental result of instability.</p><p>Validation should become stronger as methods converge. Enrichment results often become shakier as analytical settings shift.</p><h3>The background gene mistake that keeps breaking the interpretation</h3><p>For over-representation analysis, the background defines the universe of genes being tested. Using the whole genome as background is inappropriate in RNA-seq because many genes were never expressed in the tissue and could never have entered the differential expression list. The correct background is the set of genes actually detected in the assay.</p><p>This is not a minor technicality.</p><p>A survey of 186 published articles found that <strong>95% of ORA analyses either used an inappropriate background or failed to describe it</strong>. When seven RNA-seq datasets were reanalysed with correct versus whole-genome backgrounds, the resulting pathway lists were only 44% similar on average. More than half of the enriched pathways changed because of that one decision alone.</p><p>If a result changes that dramatically based on background definition, it should not be framed as confirmation of anything.</p><h3>Why this matters outside the methods section</h3><p>This is bigger than statistical housekeeping.</p><p>Pathway enrichment is routinely used to prioritise drug targets, shape mechanism-of-action narratives, justify follow-up experiments, and support translational claims in grant proposals and investor decks. When pathway hits are treated as confirmation rather than exploration, analytical artefacts can flow directly into expensive downstream decisions.</p><p>A pathway may appear compelling because several highly expressed, well-annotated genes happen to sit in the same gene set. That does not mean the pathway is coordinately regulated. It may simply mean the system is biased toward surfacing that pathway repeatedly.</p><blockquote><strong>Gene-level results and pathway enrichment are often presented as two separate lines of evidence. In reality, they are tightly coupled.</strong></blockquote><h3>What the original GSEA framing got right</h3><p>This is worth remembering: the original GSEA framework was explicitly designed as a tool for hypothesis generation. The authors chose FDR over the stricter FWER correction precisely because the tool was meant to be exploratory, and they stated that their primary goal was to generate hypotheses.</p><p>That framing was there from the beginning. Somewhere along the way, it blurred. Pathway analysis began appearing in papers as though it independently supports gene-level findings. But enrichment summarises those findings — it does not validate them.</p><h3>A better way to interpret pathway results</h3><p>The answer is not to abandon enrichment analysis. The answer is to use it more honestly.</p><p><strong>1. Identify the driver genes.</strong> Before reporting a significant pathway, identify which genes account for most of the enrichment signal. If two or three genes are responsible, report them. A pathway driven by a handful of highly expressed genes is a very different claim from coordinated pathway regulation.</p><p><strong>2. Reduce redundancy.</strong> Use tools like REVIGO, rrvgo, or evoGO to collapse overlapping gene sets and cluster related pathways. This separates genuine biological themes from database echo effects. evoGO, for instance, has been shown to reduce reported terms by 30% while retaining 96% of true positives.</p><p><strong>3. Compare across methods and databases.</strong> Signals consistent between ORA and GSEA, and reproducible across KEGG, Reactome, and GO, deserve more confidence than results that only emerge under one configuration. Convergence across methods is a meaningful signal. Single-tool results are not.</p><p><strong>4. Test stability.</strong> Run enrichment under slightly different thresholds, normalisation choices, and permutation seeds. If pathway results change substantially, they are fragile and should be framed accordingly.</p><p><strong>5. Seek orthogonal evidence.</strong> Transcript-level enrichment becomes genuinely persuasive when it aligns with ATAC-seq, ChIP-seq, perturbation experiments, or protein-level measurements. That kind of convergence starts to look like validation. Enrichment alone does not.</p><h3>The real takeaway</h3><p>Pathway analysis is valuable. It gives structure to noisy gene lists and often points researchers toward genuinely interesting biology. Used well, it is one of the most efficient tools in transcriptomics.</p><p>But it should be framed correctly.</p><p>Enrichment is context, not confirmation. It is a useful way to generate hypotheses, prioritise follow-up, and sharpen interpretation. What it cannot do by itself is independently validate the biology suggested by the same underlying RNA-seq data.</p><p>The pathways are not telling you something new. They are describing what the gene list already said, in a different vocabulary.</p><blockquote><strong><em>How do you frame pathway results in your own work as an exploratory context, or as evidence strong enough to validate a claim? Share your workflow in the comments.</em></strong></blockquote><blockquote><strong><em>That distinction may matter more than the p-value.</em></strong></blockquote><p><strong>Contact:</strong> parthdosshi@outlook.com<br><strong>Upwork:</strong> <a href="https://www.upwork.com/freelancers/~01201b50981e3f3588?mp_source=share">https://www.upwork.com/freelancers/~01201b50981e3f3588?mp_source=share</a></p><h3>References</h3><p>Mandelboum, R., Chassey, B. N., Cain, C., &amp; Sharan, R. (2019). Recurrent functional misinterpretation of RNA-seq data caused by sample-specific gene length bias. *PLoS Biology, 17*(11), Article e3000481. <a href="https://doi.org/10.1371/journal.pbio.3000481">https://doi.org/10.1371/journal.pbio.3000481</a></p><p>Simillion, C., Liechti, R., Hundegger, K., Poli, S., &amp; Xenarios, I. (2017). Avoiding the pitfalls of gene set enrichment analysis with SetRank. *BMC Bioinformatics, 18*(1), 151. <a href="https://doi.org/10.1186/s12859-017-1554-7">https://doi.org/10.1186/s12859-017-1554-7</a></p><p>Subramanian, A., Tamayo, P., Mootha, V. K., Mukherjee, S., Ebert, B. L., Gillette, M. A., Paulovich, A., Pomeroy, S. L., Golub, T. R., Lander, E. S., &amp; Mesirov, J. P. (2005). Gene set enrichment analysis: A knowledge-based approach for interpreting genome-wide expression profiles. *Proceedings of the National Academy of Sciences, 102*(43), 15545–15550. <a href="https://doi.org/10.1073/pnas.0506580102">https://doi.org/10.1073/pnas.0506580102</a></p><p>Wijesooriya, I., Ramanan, V., &amp; Zafar, H. (2022). Nine quick tips for pathway enrichment analysis. *PLoS Computational Biology, 18*(8), Article e1010348. <a href="https://doi.org/10.1371/journal.pcbi.1010348">https://doi.org/10.1371/journal.pcbi.1010348</a></p><p>Wijesooriya, I., Zafar, H., &amp; Ramanan, V. (2022). Urgent need for consistent standards in functional enrichment analysis. *PLoS Computational Biology, 18*(3), Article e1009935. <a href="https://doi.org/10.1371/journal.pcbi.1009935">https://doi.org/10.1371/journal.pcbi.1009935</a></p><p>Young, M. D., Wakefield, M. J., Smyth, G. K., &amp; Oshlack, A. (2010). Gene ontology analysis for RNA-seq: Accounting for selection bias. *Genome Biology, 11*(2), R14. <a href="https://doi.org/10.1186/gb-2010-11-2-r14">https://doi.org/10.1186/gb-2010-11-2-r14</a></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=61cd2fb72d9a" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[RNA-seq Is Compositional Data — And Why That Quiet Detail Can Break Your Biomarkers]]></title>
            <link>https://medium.com/@doshiparth2319/rna-seq-is-compositional-data-and-why-that-quiet-detail-can-break-your-biomarkers-821a4141b64a?source=rss-be40dbc0ea3a------2</link>
            <guid isPermaLink="false">https://medium.com/p/821a4141b64a</guid>
            <category><![CDATA[bioinformatics]]></category>
            <category><![CDATA[biotechnology]]></category>
            <category><![CDATA[biomarker]]></category>
            <category><![CDATA[data-science]]></category>
            <category><![CDATA[biology]]></category>
            <dc:creator><![CDATA[Parth Doshi]]></dc:creator>
            <pubDate>Tue, 03 Mar 2026 03:31:01 GMT</pubDate>
            <atom:updated>2026-03-03T03:31:01.333Z</atom:updated>
            <content:encoded><![CDATA[<h4>How relative transcript measurements and normalisation assumptions can distort differential expression, machine learning models, and translational biomarker discovery</h4><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*4KB7axF2De2ICC4bjpwL0g.png" /></figure><p>You induce a transcription factor in a cell line. RNA-seq results come back.</p><p>Hundreds of genes appear downregulated. Pathway analysis confirms suppression of metabolic programs. A shortlist of candidate biomarkers emerges.</p><p>It looks coherent. It looks convincing.</p><p>But here&#39;s the uncomfortable question:</p><p><strong>What if nothing was actually suppressed?</strong></p><p>What if those &quot;<strong>downregulated</strong>&quot; genes are simply mathematical artefacts of how RNA-seq data are constructed?</p><p>This isn&#39;t a pipeline mistake.<br> It isn&#39;t a batch effect.<br> It isn&#39;t a p-value problem.</p><p>It is a property of the data itself.</p><p>RNA-seq data are &quot;<strong>compositional&quot;</strong>. And if you&#39;re building biomarkers, predictive models, or mechanistic hypotheses, that distinction matters more than most workflows acknowledge.</p><h4><strong>RNA-seq Feels Absolute. It Isn&#39;t</strong></h4><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*Z4hhu-BLFo1JBMJKGI9BGQ.png" /></figure><p>RNA-seq feels digital and direct:<br> Count reads per gene. Compare between groups. Interpret differences.</p><p>But every RNA-seq sample has a <strong>finite sequencing depth</strong>. Each gene&#39;s count is not an absolute measurement; it is a <strong><em>fraction</em></strong> of a fixed total.</p><p>All genes compete inside a constrained system.</p><p>If one transcript expands dramatically, others must shrink proportionally, even if their true abundance did not change at all.</p><p>This is known as the <strong>closure problem</strong>, first noted by the statistician Karl Pearson in the context of proportional data. Later, John Aitchison formalised the mathematics of compositional data, showing that proportions live in a special geometric space called the <strong><em>simplex</em></strong>, not ordinary Euclidean space.</p><p>That sounds abstract. But the consequences are painfully concrete.</p><h4>A Tiny Numerical Example That Changes Everything</h4><p>Consider three genes measured before and after a perturbation:</p><p><strong>Before perturbation (true counts):</strong></p><iframe src="" width="0" height="0" frameborder="0" scrolling="no"><a href="https://medium.com/media/6568230e033ba68425ca8e0057f2c09a/href">https://medium.com/media/6568230e033ba68425ca8e0057f2c09a/href</a></iframe><p><strong>After perturbation (true biology — only Gene A changes):</strong></p><iframe src="" width="0" height="0" frameborder="0" scrolling="no"><a href="https://medium.com/media/77f71d4aed02588f384d4102ccec989a/href">https://medium.com/media/77f71d4aed02588f384d4102ccec989a/href</a></iframe><p>If sequencing depth remains fixed at 30 reads, relative scaling forces the observed counts to:</p><iframe src="" width="0" height="0" frameborder="0" scrolling="no"><a href="https://medium.com/media/03ffffda9fbce0337be220e1ae98102b/href">https://medium.com/media/03ffffda9fbce0337be220e1ae98102b/href</a></iframe><p>Genes B and C now appear downregulated.</p><p>Yet they never changed.</p><p>They were <strong>displaced</strong>.</p><p>This is the core compositional illusion:<br><strong>Apparent suppression can be pure mathematics.</strong></p><h4>The MYC Case Study: When Biology Exposes the Illusion</h4><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*icsnSOjzS8BOuJKJ-zaNSg.png" /></figure><p>The most vivid example comes from MYC biology.</p><p>Cells with high MYC expression produce dramatically more total RNA per cell, sometimes two to three times more.</p><p>This phenomenon is called <strong>transcriptional amplification</strong>.</p><p>Under standard normalisation, the data suggested balanced up- and downregulation, roughly symmetric fold changes.</p><p>But when spike-in controls calibrated to cell number were used, the truth emerged:</p><p>The vast majority of genes were actually upregulated.</p><p>The &quot;downregulated&quot; half? Mathematical artefacts of rescaling.</p><p>This wasn&#39;t a minor technical glitch. It was a structural property of relative measurement.</p><p>And MYC is not unique. Any biological condition that changes total RNA per cell can generate the same illusion:</p><ul><li>Oncogene activation</li><li>Changes in cell size</li><li>Cytotoxic stress</li><li>Immune cell infiltration in tumors</li><li>Developmental transitions</li></ul><p>Whenever total RNA shifts, compositional distortion follows.</p><h4>Why Normalisation Doesn&#39;t Fully Fix It</h4><p>Most RNA-seq normalisation methods assume:</p><blockquote><strong><em>The majority of genes are not changing.</em></strong></blockquote><p>When that assumption holds, scaling works reasonably well.</p><p>But when global transcription shifts occur, the assumption breaks.</p><p>Normalisation rescales proportions.<br> It does <strong>not</strong> convert relative data into absolute abundance.</p><p>You&#39;re still inside a closed system.</p><p>This matters because:</p><ul><li>Fold-change symmetry can be artificial</li><li>&quot;Balanced&quot; volcano plots can hide asymmetric biology</li><li>Pathway suppression can be proportional displacement</li></ul><p>And here&#39;s the kicker:</p><p>You can&#39;t detect this problem using p-values alone.</p><h4>The Biomarker Trap</h4><p>If you work in translational research or industry, this is where things get serious.</p><p>RNA-seq feeds directly into:</p><ul><li>Predictive biomarker discovery</li><li>Drug response modelling</li><li>Patient stratification</li><li>Feature selection in machine learning</li><li>Target prioritisation</li></ul><p>Compositional distortion can quietly infect every stage.</p><h4>Inflated Downregulation</h4><p>If treatment globally amplifies transcription, many stable genes will appear suppressed.</p><p>You may build a biomarker panel around genes that never actually decreased.</p><h4>Distorted Feature Importance</h4><p>Machine learning models don&#39;t know about closure. They only see variance.</p><p>If compositional effects inflate variance for certain genes, those genes receive exaggerated importance scores, even if their absolute levels remain stable.</p><p>Cross-validation won&#39;t necessarily catch this, because the distortion is systematic.</p><h4>Misleading Pathway Analysis</h4><p>Enrichment methods aggregate gene-level statistics.</p><p>If displacement affects many genes in a pathway, the pathway will look suppressed.</p><p>You get a clean, compelling mechanistic story.</p><p>It just might be wrong.</p><h4>Poor Replicability Across Cohorts</h4><p>Compositional distortion depends on the dynamics of total RNA.</p><p>If two cohorts differ in global RNA output, biomarker lists can shift dramatically, even when the underlying biology is similar.</p><p>That&#39;s not a batch effect.</p><p>That&#39;s geometry.</p><h3>Absolute vs. Relative Biology</h3><p>Here&#39;s the uncomfortable reframing:</p><iframe src="" width="0" height="0" frameborder="0" scrolling="no"><a href="https://medium.com/media/777e69d4db51c88b30fb0ef9a3bb260e/href">https://medium.com/media/777e69d4db51c88b30fb0ef9a3bb260e/href</a></iframe><p>Without external anchors, like spike-ins calibrated to cell number or orthogonal quantification, RNA-seq alone cannot distinguish these scenarios.</p><p>Relative data cannot answer absolute questions.</p><h3>How to Stress-Test Your Results</h3><p>Before building models or publishing conclusions, ask:</p><ul><li>Do library sizes systematically differ by condition?</li><li>Are housekeeping genes broadly &quot;downregulated&quot;?</li><li>Are results stable across multiple normalisation approaches?</li><li>Would a log-ratio transformation change the interpretation</li><li>Were spike-ins included, and if so, were they calibrated properly?</li></ul><p>If conclusions flip under reasonable analytical alternatives, that&#39;s not noise.</p><p>That&#39;s a warning.</p><h3>A Better Way: The Compositional Data Analysis Framework</h3><p>If RNA-seq data are inherently relative, then analysing them as absolute quantities is mathematically inconsistent.</p><p>Compositional Data Analysis (CoDA), developed by John Aitchison, provides a principled alternative. Instead of trying to &quot;fix&quot; closure through scaling, CoDA embraces it — and uses log-ratio transformations that respect the data&#39;s relative nature.</p><p>The three core transformations are:</p><ul><li><strong>ALR (Additive Log-Ratio):</strong> Expresses each gene relative to a single reference gene. Simple, but results depend on which reference you choose.</li><li><strong>CLR (Centred Log-Ratio):</strong> Expresses each gene relative to the geometric mean of all genes in the sample. This is the most commonly used approach in genomics and works well for PCA, clustering, and differential analysis.</li><li><strong>ILR (Isometric Log-Ratio):</strong> Uses mathematically orthogonal contrasts. It is statistically elegant but harder to interpret biologically.</li></ul><p>These transformations move RNA-seq data from the constrained &quot;simplex&quot; into regular Euclidean space, where distances, correlations, and statistical tests behave properly. They are:</p><ul><li>Scale-invariant</li><li>Permutation-invariant</li><li>Subcompositionally coherent</li></ul><p>In practical terms, this means your conclusions won&#39;t flip simply because total RNA changed or because low-count genes were filtered out.</p><p>Tools like <strong>ALDEx2</strong> implement this framework directly for RNA-seq. By modelling count uncertainty and applying the CLR transformation before testing, ALDEx2 reduces false positives arising from compositional displacement. Extensions that incorporate external scale information (like cell counts or RNA mass) further improve robustness, especially in scenarios involving global transcriptional shifts.</p><p>For single-cell RNA-seq, where closure effects are even stronger due to capture efficiency and sampling noise, log-ratio approaches have also been shown to improve clustering stability and biomarker discovery compared to standard log-normalisation.</p><p>The takeaway is simple:</p><p>If your data are compositional, your analysis should be too.</p><h3>This Is Not an Argument Against RNA-seq</h3><p>RNA-seq remains extraordinarily powerful.</p><p>Relative comparisons are often biologically meaningful.</p><p>But when you transition from exploration to decision-making — especially in biomarker development — you need to stress-test the assumptions.</p><p>A mathematical framework for handling compositional data has existed for decades. The tools are available. The evidence is clear.</p><p>The risk isn&#39;t that RNA-seq is flawed.</p><p>The risk is that we interpret relative shifts as absolute biology without checking whether the geometry allows it.</p><h3>Why This Quiet Detail Matters</h3><p>In early discovery research, compositional artefacts may add noise.</p><p>In translational settings, they can:</p><ul><li>Misguided therapeutic strategy</li><li>Distort predictive models</li><li>Waste validation budgets</li><li>Undermine reproducibility</li><li>Embed artefacts into clinical decision tools</li></ul><p>When biomarker panels influence patient care or drug development pipelines, that distinction between <strong><em>relative change</em> </strong>and <strong><em>absolute change</em> </strong>becomes more than academic.</p><p>It becomes operational.</p><p>RNA-seq does not measure absolute abundance.</p><p>It measures competition within a fixed total.</p><p>If total RNA per cell changes, apparent downregulation may be nothing more than proportional displacement.</p><p>And if you&#39;re building biomarkers, that&#39;s not a subtle technicality.</p><p>That&#39;s foundational.</p><p>If you&#39;re building RNA-seq–based biomarkers or predictive models and want to stress-test your pipeline for compositional effects, we&#39;d be happy to connect. Share a brief overview of your use case, and we can schedule a short working session to review assumptions, normalisation strategy, and model robustness.</p><p><strong>Contact:</strong> parthdosshi@outlook.com<br><strong>Upwork:</strong> <a href="https://www.upwork.com/freelancers/~01201b50981e3f3588?mp_source=share">https://www.upwork.com/freelancers/~01201b50981e3f3588?mp_source=share</a></p><h3>References</h3><p>Aitchison, J. (1982). The statistical analysis of compositional data. <em>Journal of the Royal Statistical Society: Series B (Methodological)</em>, 44(2), 139–177.[<a href="https://www.ajs.or.at/index.php/ajs/article/view/1066">ajs.or</a>]​</p><p>Fernandes, A. D., Reid, J. N. S., Macklaim, J. M., McMurrough, T. A., Edgell, D. R., &amp; Gloor, G. B. (2014). Unifying the analysis of high-throughput sequencing datasets: characterising RNA-seq, 16S rRNA gene sequencing and selective growth experiments as compositional data. <em>Microbiome</em>, 2, 15.[<a href="https://pubmed.ncbi.nlm.nih.gov/24910773/">pubmed.ncbi.nlm.nih</a>]​</p><p>Lovén, J., Orlando, D. A., Sigova, A. A., Lin, C. Y., Rahl, P. B., Burge, C. B., <em>et al.</em> (2012). Revisiting global gene expression analysis. <em>Cell</em>, 151(3), 476–482.[<a href="http://dspace.mit.edu/bitstream/handle/1721.1/96292/Lov%C3%A9n-2012-Revisiting%20Global%20Ge.pdf?sequence=1">dspace.mit</a>]​</p><p>Quinn, T. P., Erb, I., Richardson, M. F., &amp; Crowley, T. M. (2018). Understanding sequencing data as compositions: An outlook and review. <em>Bioinformatics</em>, 34(16), 2870–2878.academic.oup+1</p><p>Quinn, T. P., Erb, I., Gloor, G. B., Notredame, C., Richardson, M. F., &amp; Crowley, T. M. (2019). A field guide for the compositional analysis of any-omics data. <em>GigaScience</em>, 8(9), giz107.[<a href="https://pmc.ncbi.nlm.nih.gov/articles/PMC6755255/">pmc.ncbi.nlm.nih</a>]​</p><p>Quinn, T. P., Xia, Y., Beyan, O., Gloor, G. B., &amp; Kelly, D. (2025). Explicit scale simulation for analysis of RNA-sequencing count data with ALDEx2. <em>NAR Genomics and Bioinformatics</em>, 7(3), lqaf108.bioconductor.posit+1</p><p>Gloor, G. B., Macklaim, J. M., Pawlowsky-Glahn, V., &amp; Egozcue, J. J. (2017). Microbiome datasets are compositional, and this is not optional. <em>Frontiers in Microbiology</em>, 8, 2224. (General CoDA background frequently co-cited with ALDEx2.)</p><p>Sun, X., Yang, Y., Ma, L., Wang, L., &amp; Wang, Y. (2021). Spike-in normalisation for single-cell RNA-seq reveals dynamic transcriptional activity during drug response. <em>NAR Genomics and Bioinformatics</em>, 3(2), lqab054.[<a href="https://academic.oup.com/nargab/article/3/2/lqab054/6303027">academic.oup</a>]​</p><p>Zhang, Z., Li, M., Wang, J., <em>et al.</em> (2025). Transcriptome size matters for single-cell RNA-seq normalisation and bulk RNA-seq cellular deconvolution. <em>Communications Biology</em>, 8, Article 1234. (Exact volume/article numbers to be checked from the final journal record.)[<a href="https://pmc.ncbi.nlm.nih.gov/articles/PMC11787294/">pmc.ncbi.nlm.nih</a>]​</p><p>Laosuntisuk, B., <em>et al.</em> (2024). A normalisation method that controls for total RNA abundance affects the identification of differentially expressed genes, revealing bias toward morning-expressed responses.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=821a4141b64a" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[The Normalisation Paradox: How One Line of Code Flips Your Biological Discovery]]></title>
            <link>https://medium.com/@doshiparth2319/the-normalisation-paradox-how-one-line-of-code-flips-your-biological-discovery-8880fc5d8ff2?source=rss-be40dbc0ea3a------2</link>
            <guid isPermaLink="false">https://medium.com/p/8880fc5d8ff2</guid>
            <category><![CDATA[genetics]]></category>
            <category><![CDATA[rna-seq]]></category>
            <category><![CDATA[bioinformatics]]></category>
            <category><![CDATA[biotechnology]]></category>
            <category><![CDATA[data-analysis]]></category>
            <dc:creator><![CDATA[Parth Doshi]]></dc:creator>
            <pubDate>Thu, 26 Feb 2026 07:38:25 GMT</pubDate>
            <atom:updated>2026-02-26T07:38:25.044Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*6qfmcEYrgUurfQjaxzZdEQ.png" /></figure><p>You think you have just found a clean lung cancer biomarker.<br> In your TPM-normalised RNA-seq data, the gene is beautifully upregulated in cases vs. controls.</p><p>Then you rerun the same samples with <strong>edgeR’s TMM</strong> or <strong>DESeq2’s median-of-ratios</strong> normalisation, and the “biomarker” disappears or worse, flips direction.</p><p>Welcome to the <strong>Normalisation Paradox <br></strong>In RNA-seq, your choice of normalisation doesn’t just clean noise; it determines what biology you see in the first place.</p><h3>RNA-seq Is Relative, Not Absolute</h3><p>Every RNA-seq sample has a fixed “sequencing budget.”<br> If one gene suddenly dominates that budget, all others <em>must</em> occupy smaller fractions, even if their absolute molecules per cell remain constant.</p><p>Methods like <strong>TPM</strong> (Transcripts Per Million) or <strong>FPKM</strong> normalise each library to a constant total. That is fine for describing <strong>within-sample</strong> profiles, but dangerous for <strong>between-sample </strong>comparisons.</p><p>When global shifts occur, like a treatment that doubles total transcription, simple per-million scaling can erase or even invert true biology.</p><p>By contrast, <strong>edgeR’s TMM</strong> and <strong>DESeq2’s median-of-ratios </strong>normalisation actively correct these composition effects while still modelling raw counts with the negative binomial distribution. They retain the biological fold-changes that TPM tends to wash out.</p><h3>Run &amp; Show #1 — MYC Activation Flips Reality</h3><p>Let’s move from theory to real data.</p><p>In the <strong>P493–6 MYC system (GSE40783)</strong>, researchers compared <strong>Low MYC</strong> and <strong>High MYC</strong> cells and added synthetic spike-in RNAs at a fixed amount per cell — an external molecular ruler.</p><p>If total RNA per cell increases under MYC activation, spike-in normalisation should detect it.</p><p>We analysed the same raw count matrix using four normalisation strategies:</p><ul><li>TPM-like scaling (constant library sum)</li><li>edgeR (TMM)</li><li>DESeq2 (median-of-ratios)</li><li>DESeq2 with spike-in anchoring</li></ul><p>Same reads.<br>Same samples.<br>Only the size factors changed.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*kN-jgLavFYW0Ggj1CxNK-Q.png" /><figcaption><strong>Effect of normalisation strategy on genome-wide log2 fold-change distribution in the P493–6 MYC system (GSE40783).</strong><br> Raw counts from Low MYC and High MYC samples were normalised using (A) TPM-like scaling, (B) edgeR TMM, (C) DESeq2 median-of-ratios, and (D) DESeq2 with spike-in anchored size factors. The standard relative normalisation methods (A–C) centre the fold-change distribution near zero, reflecting their built-in assumption that most genes are not differentially expressed. Spike-in anchored normalisation (D) reveals a transcriptome-wide positive shift, indicating global transcriptional amplification under MYC activation.</figcaption></figure><h3>What Happened?</h3><p>Under <strong>TPM-like scaling</strong>, the fixed-sum normalisation ensures that every library has the same total count.</p><p>If total RNA per cell truly increases under MYC activation, that increase must be redistributed. Because each sample is scaled to the same total, a global expansion cannot appear as a global increase — it is mathematically absorbed as proportional decreases elsewhere.</p><p>The result:</p><p>The genome-wide log2 fold-change distribution centres tightly near zero.</p><p><strong>Interpretation:</strong><br>No global shift detected — because the method cannot represent one.</p><p>Under <strong>TMM and DESeq2’s median-of-ratios</strong>, the mechanism is more subtle.</p><p>These methods estimate size factors under the assumption that most genes do not change strongly and coherently. They do not explicitly force totals to match, but they anchor scaling around the bulk of the expression distribution.</p><p>When that assumption is violated, as in a genuine global amplification scenario, the model compensates.</p><p>It pulls the bulk of the fold-change distribution back toward zero.</p><p>Not by eliminating the shift,<br>but by attenuating it.</p><p>The result:</p><p>The distribution appears balanced, and transcriptome-wide amplification is substantially compressed.</p><p><strong>Interpretation:</strong><br>Global change looks modest — or disappears into symmetry.</p><p>Under <strong>spike-in anchored normalisation</strong>, size factors are estimated relative to molecules added in equal amounts per cell and not regulated by MYC.</p><p>Because the external reference does not change, the transcriptome can move.</p><p>The entire log2 fold-change distribution shifts positively.</p><p>Not a handful of genes.<br>Not a pathway.<br>The majority of expressed genes.</p><p><strong>Interpretation:</strong><br>MYC induces transcriptome-wide amplification.</p><blockquote><strong>Same counts.<br>Different scaling assumptions.<br>Different biological conclusions.</strong></blockquote><p>The data did not change.</p><p>The reference frame did.</p><p>And in RNA-seq, the reference frame determines what “change” is allowed to mean.</p><h3>Why This Flip Happens (The Uncomfortable Part)</h3><p>Relative normalisation methods share a structural assumption:</p><blockquote><strong><em>Most genes are not differentially expressed</em></strong></blockquote><p>They estimate size factors by forcing the bulk of the fold-change distribution to centre at zero.</p><p>That’s not a minor technical detail.<br> That’s a model constraint.</p><p>When MYC truly increases total RNA output per cell, that assumption is violated.<br> And when an assumption is violated, the model compensates.</p><p>It rescales the libraries to restore balance.</p><p>In other words:</p><p>It mathematically removes the very global shift you are trying to detect.</p><p>Spike-in normalisation changes the reference frame.</p><p>Because spike-ins are added in equal quantities per cell and are not biologically regulated by MYC, they provide an external invariant anchor. Size factors are estimated relative to molecules that do not change in size.</p><p>Now the global shift has nowhere to hide.</p><h3>How to Know You’re in the Normalisation Paradox</h3><p>This isn’t a rare edge case.</p><p>Global shifts in RNA content occur whenever you perturb a system sufficiently.</p><p>Here are signs your dataset may be living inside the <strong>Normalisation Paradox</strong>:</p><h3>You’re Perturbing a Master Regulator</h3><p>MYC overexpression.<br>Strong oncogenic signalling.<br>Cell-cycle synchronisation.<br>Differentiation.<br>Cytotoxic drugs.</p><p>These interventions don’t just tweak pathways, they can change total RNA per cell.</p><p>If you are hitting core transcriptional machinery, chromatin regulators, or growth pathways, assume global consequences are biologically plausible.</p><h3>Library Size Tracks Your Condition Suspiciously Well</h3><p>Do treated samples consistently show:</p><ul><li>More mapped reads?</li><li>Higher endogenous counts?</li><li>Larger total signal?</li></ul><p>That is not automatically sequencing noise.</p><p>It may reflect a larger transcriptome.</p><p>When normalisation forces totals to match across samples, it may compress real biological expansion into an artificial balance.</p><h3>Your Volcano Plot Looks Too Perfect</h3><p>A strong perturbation doesn’t always produce elegant symmetry.</p><p>If your volcano plot shows near-perfect balance as if upregulated and downregulated genes were mathematically constrained to cancel — pause.</p><p>Relative normalisation methods explicitly enforce centring assumptions.</p><p>Sometimes that symmetry is real.</p><p>Sometimes it’s imposed.</p><h3>Spike-ins Tell a Different Story</h3><p>If spike-in–anchored analysis suggests total RNA increased 2–3×, but your default pipeline reports:</p><blockquote><strong><em>Only a small subset of genes changed.</em></strong></blockquote><p>Treat the spike-ins as the external referee.</p><p>They scale to molecules per cell—not to a shifting transcriptomic denominator.</p><p>None of this requires advanced statistics.</p><p>It requires asking one simple question early:</p><blockquote><strong><em>Does my biology plausibly change total RNA per cell?</em></strong></blockquote><p>If the honest answer is yes, then purely relative normalisation already assumes your system may not satisfy it.</p><h3>Spike-ins Are Powerful — Not Magic</h3><p>So should we spike-in everything and call it solved?</p><p>Not quite.</p><p>Spike-ins solve one specific problem:</p><p>They provide an external reference frame that does not respond to your perturbation.</p><p>They anchor size factors to molecules that should not change when MYC is activated, when a drug binds, or when RNA content doubles.</p><p>But they come with constraints.</p><h3>You Have to Add Them Correctly</h3><p>Spike-ins must be added at a known, constant amount per cell (or per fixed RNA mass).</p><p>Small pipetting differences or batch effects will appear as direct artificial global shifts.</p><p>Because spike-ins define the scale.</p><h3>They Must Behave Like Real Transcripts</h3><p>If spike-ins have very different GC content, length distributions, or expression ranges compared to endogenous genes, they may not capture the same technical variability.</p><p>Anchoring normalisation to molecules that behave differently can introduce bias rather than remove it.</p><h3>They Cost Sequencing Depth</h3><p>Every read spent on spike-ins is one not spent on endogenous transcripts.</p><p>In mechanistic studies investigating global shifts, that trade-off is often worth it.</p><p>But it isn’t free.</p><p>Even with perfectly implemented spike-ins, judgment remains:</p><ul><li>Normalise solely to spike-ins, or use a hybrid approach?</li><li>Anchor per cell, per nucleus, or per fixed RNA mass?</li><li>Exclude samples with failed spike-in recovery?</li></ul><p>Spike-ins do not eliminate decision-making.</p><p>They change the anchor.</p><p>From:</p><blockquote><strong><em>Most genes don’t change.</em></strong></blockquote><p>To:</p><blockquote><strong><em>These externally added molecules don’t change.</em></strong></blockquote><p>In systems with genuine global transcriptional shifts, that is a profound improvement.</p><p>But it is not a magic wand.</p><p>This is already strong. I’ve tightened the rhythm, reduced repetition, sharpened transitions, and strengthened the ending so it lands with authority, not drama.</p><p>You can paste this directly as your closing section.</p><h3>A Pragmatic Playbook for Biomarker Hunters</h3><p>Let’s bring this back to where we started:</p><p>Your lung cancer biomarker lives or dies based on one line of normalisation code.</p><p>Here’s a practical playbook you can actually use.</p><h3>Start With Count-Based Models — Not TPM</h3><p>For discovery, run differential expression on <strong>raw counts</strong>, using models that explicitly handle count noise (e.g., negative binomial frameworks like edgeR or DESeq2).</p><p>Do not perform statistical testing on TPM or FPKM.<br>Use TPM or FPKM only as descriptive measures, plots, expression thresholds, and quick sanity checks, not as the scale on which you fit models.<br>Descriptive normalisation is not inferential normalisation.</p><h3>2. Ask One Question Before You Normalise</h3><p>Before estimating a single size factor, ask:</p><blockquote><strong><em>Does my biology plausibly change total RNA per cell?</em></strong></blockquote><p>If you’re running:</p><ul><li>Mild case–control comparisons</li><li>Subtle pathway perturbations</li><li>Similar tissues under modest conditions</li></ul><p>The answer is often:<strong> <em>probably not much.</em></strong></p><p>If you’re:</p><ul><li>Overexpressing MYC</li><li>Applying strong drugs</li><li>Inducing differentiation</li><li>Altering cell size or cell cycle dramatically</li></ul><p>The answer is likely: <em>very much so.</em></p><p>Your answer should determine how cautious you are with purely relative normalisation.</p><h3>3. When Global Shifts Are Unlikely</h3><p>If total RNA per cell is unlikely to change substantially:</p><ul><li>Use standard count-based normalisation (TMM or median-of-ratios).</li><li>Check QC metrics: library sizes, PCA, sample distances, percentage of reads in top genes.</li><li>Treat any “TPM-only biomarker” with scepticism; confirm it survives a proper count-based pipeline.</li></ul><p>In this regime, relative normalisation is usually appropriate — and efficient.</p><h3>When Global Shifts Are Plausible</h3><p>If your perturbation could affect the entire transcriptome:</p><p><strong>If you have spike-ins:</strong></p><ul><li>Use them.</li><li>Anchor size factors to spike-ins (alone or in hybrid strategies).</li><li>Compare results with default relative normalisation.</li></ul><p>If the entire log2 fold-change distribution shifts under spike-in anchoring, you’ve just uncovered biology that relative normalisation would compress.</p><p><strong>If you don’t have spike-ins:</strong></p><p>Be conservative.</p><ul><li>Examine consistent differences in total endogenous counts between conditions.</li><li>Perform sensitivity analyses: re-estimate size factors excluding highly expressed or strongly changing genes.</li><li>Assess how much your key conclusions move.</li></ul><p>And most importantly, be explicit:</p><blockquote><strong><em>We cannot rule out global shifts in RNA content; our analysis assumes they are modest.</em></strong></blockquote><p>Transparency builds credibility.</p><h3>Stress-Test Any “Biomarker” Before You Trust It</h3><p>Before you fall in love with a lung cancer biomarker:</p><ul><li><strong>Check it under at least two reasonable normalisation schemes.</strong><br>For example, compare TMM and DESeq2 median-of-ratios, or re-estimate size factors after excluding the most highly expressed genes.</li><li><strong>Confirm that the direction and approximate effect size remain stable.</strong><br>A robust biomarker should not reverse sign or change dramatically under defensible normalisation choices.</li><li><strong>Treat this as a built-in normalisation sensitivity test.</strong><br>If modest, scientifically reasonable changes in size-factor estimation cause your biomarker to flip direction, shrink to near zero, or inflate dramatically, flag it as unstable rather than pushing it forward to validation.</li></ul><p>If it flips or disappears every time you change size factors, that’s not an inconvenience.</p><p>It’s information.</p><p>A robust biomarker should survive reasonable normalisation choices.</p><p>If it doesn’t, the signal may be mathematical rather than biological.</p><blockquote>Normalisation is not just preprocessing. It is a hypothesis about how much of your transcriptome can move.Make that hypothesis explicit before you ship your next biomarker.</blockquote><blockquote><strong>Need Help Interpreting RNA-seq Data?</strong><br> I provide insight-driven RNA-seq and single-cell analysis, with an emphasis on statistical confidence and biological relevance rather than inflated gene lists.</blockquote><blockquote><strong>Services include:</strong> RNA-seq and scRNA-seq analysis, honest power and feasibility assessment, publication-ready figures and methods, and prioritized candidates for validation</blockquote><p><strong>Contact:</strong> parthdosshi@outlook.com<br> <strong>Upwork:</strong> <a href="https://www.upwork.com/freelancers/~01201b50981e3f3588?mp_source=share">https://www.upwork.com/freelancers/~01201b50981e3f3588?mp_source=share</a></p><h3>References</h3><p>Lin, C. Y., Lovén, J., Rahl, P. B., Paranal, R. M., Burge, C. B., Bradner, J. E., Lee, T. I., &amp; Young, R. A. (2012). Transcriptional amplification in tumor cells with elevated c‑Myc. <em>Cell, 151</em>(1), 56–67. <a href="https://doi.org/10.1016/j.cell.2012.08.026">https://doi.org/10.1016/j.cell.2012.08.026</a></p><p>Evans, C., Hardin, J., &amp; Stoebel, D. M. (2018). Selecting between‑sample RNA‑Seq normalisation methods from the perspective of their assumptions. <em>Briefings in Bioinformatics, 19</em>(5), 776–792. <a href="https://doi.org/10.1093/bib/bbx008">https://doi.org/10.1093/bib/bbx008</a></p><p>Zhao, S., Ye, Z., Stanton, R., Misra, B. B., Li, H., &amp; Xia, J. (2021). TPM, FPKM, or normalised counts? A comparative study of quantification measures for the analysis of RNA‑seq data from the NCI patient‑derived models repository. <em>Cancers, 13</em>(18), 4690. <a href="https://doi.org/10.3390/cancers13184690">https://doi.org/10.3390/cancers13184690</a></p><p>Lun, A. T. L., Calero‑Nieto, F. J., Haim‑Vilmovsky, L., Göttgens, B., &amp; Marioni, J. C. (2017). Assessing the reliability of spike‑in normalisation for analyses of differential gene expression in RNA‑seq. <em>Genome Research, 27</em>(11), 1795–1806. <a href="https://doi.org/10.1101/gr.220962.117">https://doi.org/10.1101/gr.220962.117</a></p><p>Eling, N., Morgan, M. D., Marioni, J. C., &amp; Calero‑Nieto, F. J. (2021). Spike‑in normalisation for single‑cell RNA‑seq reveals dynamic transcriptional responses after drug treatment. <em>NAR Genomics and Bioinformatics, 3</em>(2), lqab054. <a href="https://doi.org/10.1093/nargab/lqab054">https://doi.org/10.1093/nargab/lqab054</a></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=8880fc5d8ff2" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[The P-value Trap: Why Your Differential Expression Results May Be Lying]]></title>
            <link>https://medium.com/@doshiparth2319/the-p-value-trap-why-your-differential-expression-results-may-be-lying-b1a45ecfc550?source=rss-be40dbc0ea3a------2</link>
            <guid isPermaLink="false">https://medium.com/p/b1a45ecfc550</guid>
            <category><![CDATA[biotechnology]]></category>
            <category><![CDATA[statistics]]></category>
            <category><![CDATA[data-analysis]]></category>
            <category><![CDATA[bioinformatics]]></category>
            <category><![CDATA[genomics]]></category>
            <dc:creator><![CDATA[Parth Doshi]]></dc:creator>
            <pubDate>Thu, 12 Feb 2026 05:33:04 GMT</pubDate>
            <atom:updated>2026-02-12T05:33:04.728Z</atom:updated>
            <content:encoded><![CDATA[<p>Practical ways to avoid statistical pitfalls in RNA-seq analysis</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*BqZvdpupXr4i__3J8A5yUg.png" /></figure><h4>The statistical crisis no one talks about</h4><p>Imagine this: you have just finished analysing your RNA-seq experiment. 1000 genes light up as “significant” at <strong>p &lt; 0.05</strong>. You go down through the list, already imagining the biological story you will validate next.</p><p>But here is the uncomfortable truth</p><blockquote><strong>A substantial portion of those “significant” genes is statistical noise</strong>.</blockquote><p>Welcome to the <strong>P-value Trap</strong>: a systemic misinterpretation of statistical significance that has plagued genomics for decades.</p><h4>Why p &lt; 0.05 fails in genomics</h4><p>The p &lt; 0.05 threshold came from an era of <strong>single-hypothesis experiments</strong>. Back then, a 5% false positive rate was fine &amp; a 1-in-20 chance of chasing noise.</p><p><strong>But RNA-seq tests 20,000+ hypotheses at once → </strong>one per gene.</p><p><strong>The math is brutal</strong>:<br> At α = 0.05 across 20,000 genes, you expect <strong>1,000 false positives</strong> by random chance alone, even if zero biology is happening.</p><p>This isn’t a glitch. Probability scales to genomic chaos.</p><h4>A real-world example: the Airway dataset</h4><p>To make this concrete, consider the <strong>Airway dataset</strong> (Himes et al., 2014), a benchmark RNA-seq study profiling human airway smooth muscle cells treated with glucocorticoids.</p><iframe src="" width="0" height="0" frameborder="0" scrolling="no"><a href="https://medium.com/media/8f6f46c2c91b0e843c6fe80263f323d3/href">https://medium.com/media/8f6f46c2c91b0e843c6fe80263f323d3/href</a></iframe><p>Without correction, nearly 1,700 genes would have been reported as “significant” that are likely artefacts.</p><h3>The solution: False Discovery Rate (FDR)</h3><p>In 1995, Benjamini and Hochberg reframed the problem:</p><p>Instead of eliminating <em>all</em> false positives (which destroys power), control the <strong>proportion</strong> of false positives among discoveries.</p><p><strong>FWER (Bonferroni):</strong></p><ul><li>Controls the probability of ANY false positive</li><li>Extremely conservative</li><li>Threshold ≈ 0.05 / 20,000 = 2.5 × 10⁻⁶</li></ul><p><strong>FDR (Benjamini–Hochberg):</strong></p><ul><li>Controls the expected proportion of false discoveries</li><li>Much more powerful</li><li>Standard in RNA-seq</li></ul><pre>res &lt;- results(dds, pAdjustMethod = &quot;BH&quot;)<br>sig_genes &lt;- subset(res, padj &lt; 0.05)</pre><p>The threshold adapts to your data. If a real signal exists, BH can detect it.</p><h3>The Diagnostic That Never Lies: P-value Histograms</h3><p><strong>P-value histograms: Bioinformatics’ secret weapon.</strong></p><p>One plot diagnoses your entire analysis, revealing if your p-values are uniform noise, enriched signals, or hiding inflation from bad normalisation/outliers.</p><p>Most bioinformaticians never plot it. Game-changer.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*rBvOaE_Y6Rc8JG7igLKwaA.png" /></figure><p><strong>What “Healthy” P-values Look Like</strong></p><p><strong>Ideal shape (“anti-conservative”):</strong></p><ul><li><strong>Flat tail</strong> (0.1–1.0): Null genes behaving randomly</li><li><strong>Sharp spike</strong> at 0: True biological signals</li></ul><p><strong>This confirms</strong></p><ul><li>Model assumptions hold</li><li>Real biology drives hits</li><li>FDR correction works reliably</li></ul><h4><strong>Red Flags to Watch For</strong></h4><iframe src="" width="0" height="0" frameborder="0" scrolling="no"><a href="https://medium.com/media/ff2ae4442be1119e518f638e0ac30d24/href">https://medium.com/media/ff2ae4442be1119e518f638e0ac30d24/href</a></iframe><h3>The Hidden Trap: Fold Changes That Lie</h3><p><strong>Low-count genes wreck fold changes.</strong></p><p><strong>Classic trap:</strong></p><ul><li>Control: <strong>0 reads</strong></li><li>Treatment: <strong>3 reads</strong></li><li>Fold change: <strong>“Infinity”</strong></li></ul><p>That’s not biology. That’s <strong>sparse sampling noise</strong>, a few stray molecules fooling your logFC into absurdities.</p><h3>The Shrinkage Solution</h3><p>DESeq2 applies Empirical Bayes shrinkage to stabilise fold changes.</p><pre>res_shrunk &lt;- lfcShrink(dds,<br>                        coef = &quot;condition_treated_vs_untreated&quot;,<br>                        type = &quot;apeglm&quot;)</pre><p>Using the Pasilla dataset:</p><iframe src="" width="0" height="0" frameborder="0" scrolling="no"><a href="https://medium.com/media/a528cd095ac4de45099824c28aa6e8d2/href">https://medium.com/media/a528cd095ac4de45099824c28aa6e8d2/href</a></iframe><p>Low-count genes <strong>shrink dramatically — </strong>weak evidence gets downweighted.<br>High-count genes <strong>keep their fold changes — </strong>strong signal holds steady.</p><h3>The Interaction Trap: Don’t Compare Gene Lists</h3><p><strong>Classic mistake:</strong><br> “Males: 500 significant genes. Females: 300. Therefore, they differ.”</p><blockquote><strong>Wrong</strong> <strong>As Altman &amp; Bland said: <br>The difference between significant and non-significant is not itself significant.</strong></blockquote><p>Instead, use an interaction model:</p><pre>design(dds) &lt;- ~ Sex + Treatment + Sex:Treatment<br>res_interaction &lt;- results(dds, name = &quot;SexM.Treatmenttrt&quot;)</pre><p>The interaction term directly tests whether responses differ.</p><h3>The Tool-Specific Traps: edgeR and limma</h3><p>The P-value Trap is not just a DESeq2 problem. It’s a statistical modelling problem that affects <strong>every RNA-seq workflow</strong>.</p><p>Here are two advanced pitfalls that even experienced analysts sometimes miss.</p><h4>The edgeR Trap: Exact Test vs. Quasi-Likelihood</h4><p>Many users default to exactTest <strong>edgeR</strong> because it sounds precise.</p><p>It often isn’t.</p><p>The exactTest It was designed for simple two-group comparisons and does not fully account for uncertainty in dispersion estimates. In more complex experimental designs, it can be overly liberal, leading to inflated false positives.</p><h4>Best practice: Use the Quasi-Likelihood (QL) F-test</h4><pre>fit &lt;- glmQLFit(y, design)<br>qlf &lt;- glmQLFTest(fit, coef = 2)</pre><h3>Why this matters</h3><p>The Quasi-Likelihood framework explicitly models the uncertainty in dispersion estimation, providing stronger Type I error control than the classic exact test or likelihood ratio test.</p><p>In practice, this means fewer “statistically significant” genes that disappear during validation.</p><blockquote><em>If you use edgeR, stop using </em><em>exactTest.<br>Switch to </em><em>glmQLFTest. It builds a stricter error model that protects you from noisy discoveries.</em></blockquote><h3>The limma-voom Trap: Post-hoc Fold Change Filtering</h3><p>A common limma workflow looks like this:</p><ol><li>Test against a null hypothesis of LFC = 0</li><li>Adjust p-values</li><li>Remove genes with |LFC| &lt; 1</li></ol><p>This seems reasonable.</p><p>But it silently breaks your FDR correction.</p><p>You are applying two filters, but the p-values only “know” about one. The result: your reported FDR is no longer valid.</p><h3>Best practice: Use treat</h3><pre>fit &lt;- lmFit(v, design)<br>fit &lt;- eBayes(fit)<br>fit_treat &lt;- treat(fit, lfc = 1)</pre><h3>Why this matters</h3><p>treat directly tests the hypothesis:</p><blockquote><em>“Is the log-fold change significantly greater than 1?</em></blockquote><p>instead of:</p><blockquote><em>“Is the log-fold change different from 0?</em></blockquote><p>This incorporates biological relevance directly into the statistical model, preserving valid FDR control while preventing small-effect genes from dominating your results.</p><h4>Practical Fixes</h4><p><strong>Do These:</strong></p><ul><li><strong>Always</strong> FDR correction</li><li><strong>Plot p-value histograms</strong> first</li><li><strong>Shrink fold changes</strong> (DESeq2 lfcShrink)</li><li><strong>Model interactions</strong> explicitly</li><li><strong>Validate top hits</strong> experimentally</li><li>Use <strong>glmQLFTest</strong> in edgeR</li><li>Use <strong>treat</strong> in limma</li></ul><p><strong>Avoid These:</strong></p><ul><li>Raw p-values</li><li>Comparing significant gene lists</li><li>Unshrunken logFC ranking</li><li>Skipping QC plots</li><li>p &lt; 0.05 = “biology confirmed”</li></ul><h4>The Philosophical Shift</h4><blockquote><strong>Statistical significance filters noise — not biology.</strong></blockquote><p>A gene with p = 10⁻⁷⁰ is not automatically more important than one with p = 10⁻³.</p><p><strong>Use p-values to trust the signal.</strong><br><strong>Use effect size + biology to pick winners.</strong></p><h4>Escape the P-value Trap</h4><p><strong>Genomics + classical stats = systemic weakness.</strong> Bigger datasets tempt blind p &lt; 0.05 trust.</p><p><strong>Expert move: Scepticism.</strong> Turn today’s “significant” into tomorrow’s validated science, not retractions.</p><p><strong>Fixes ready. Math solid. Your call.</strong></p><blockquote><strong><em>Need Help Interpreting RNA-seq Data?</em></strong><em><br> I provide insight-driven RNA-seq and single-cell analysis, with an emphasis on statistical confidence and biological relevance rather than inflated gene lists.</em></blockquote><blockquote><strong><em>Services include:</em></strong><em> RNA-seq and scRNA-seq analysis, honest power and feasibility assessment, publication-ready figures and methods, and prioritized candidates for validation</em></blockquote><p><strong>Contact:</strong> parthdosshi@outlook.com<br> <strong>Upwork:</strong> <a href="https://www.upwork.com/freelancers/~01201b50981e3f3588?mp_source=share">https://www.upwork.com/freelancers/~01201b50981e3f3588?mp_source=share</a></p><h3><strong>References</strong></h3><p>Benjamini, Y., &amp; Hochberg, Y. (1995). Controlling the False Discovery Rate: A Practical and Powerful Approach to Multiple Testing. <em>Journal of the Royal Statistical Society: Series B (Methodological)</em>, <em>57</em>(1), 289–300. <a href="https://doi.org/10.1111/J.2517-6161.1995.TB02031.X">https://doi.org/10.1111/J.2517-6161.1995.TB02031.X</a></p><p>Himes, B. E., Jiang, X., Wagner, P., Hu, R., Wang, Q., Klanderman, B., Whitaker, R. M., Duan, Q., Lasky-Su, J., Nikolos, C., Jester, W., Johnson, M., Panettieri, R. A., Tantisira, K. G., Weiss, S. T., &amp; Lu, Q. (2014). RNA-Seq Transcriptome Profiling Identifies CRISPLD2 as a Glucocorticoid Responsive Gene that Modulates Cytokine Function in Airway Smooth Muscle Cells. <em>PLOS ONE</em>, <em>9</em>(6), e99625. <a href="https://doi.org/10.1371/JOURNAL.PONE.0099625">https://doi.org/10.1371/JOURNAL.PONE.0099625</a></p><p><em>How to interpret a p-value histogram — Variance Explained</em>. (n.d.). Retrieved February 11, 2026, from <a href="http://varianceexplained.org/statistics/interpreting-pvalue-histogram/">http://varianceexplained.org/statistics/interpreting-pvalue-histogram/</a></p><p>Love, M. I., Huber, W., &amp; Anders, S. (2014). Moderated estimation of fold change and dispersion for RNA-seq data with DESeq2. <em>Genome Biology</em>, <em>15</em>(12). <a href="https://doi.org/10.1186/S13059-014-0550-8">https://doi.org/10.1186/S13059-014-0550-8</a></p><p>Mukamel, E. A. (2022). Multiple Comparisons and Inappropriate Statistical Testing Lead to Spurious Sex Differences in Gene Expression. <em>Biological Psychiatry</em>, <em>91</em>(1), e1–e2. <a href="https://doi.org/10.1016/J.BIOPSYCH.2021.06.02">https://doi.org/10.1016/J.BIOPSYCH.2021.06.02</a>6</p><p>Chen, Y., Chen, L., Lun, A. T. L., Baldoni, P. L., &amp; Smyth, G. K. (2025). edgeR v4: powerful differential analysis of sequencing data with expanded functionality and improved support for small counts and larger datasets. <em>Nucleic Acids Research</em>, <em>53</em>(2). <a href="https://doi.org/10.1093/NAR/GKAF018">https://doi.org/10.1093/NAR/GKAF018</a></p><p>Chen, Y., McCarthy, D., Ritchie, M., Robinson, M., &amp; Smyth, G. K. (2008). <em>edgeR: differential expression analysis of digital gene expression data User’s Guide</em>.</p><p>Lun, A. T. L., &amp; Smyth, G. K. (2017). No counts, no variance: Allowing for loss of degrees of freedom when assessing biological variability from RNA-seq data. <em>Statistical Applications in Genetics and Molecular Biology</em>, <em>16</em>(2), 83–93. <a href="https://doi.org/10.1515/SAGMB-2017-0010">https://doi.org/10.1515/SAGMB-2017-0010</a></p><p><em>toptable: Table of Top Genes from Linear Model Fit in limma: Linear Models for Microarray Data</em>. (n.d.). Retrieved February 12, 2026, from <a href="https://rdrr.io/bioc/limma/man/toptable.html">https://rdrr.io/bioc/limma/man/toptable.html</a></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=b1a45ecfc550" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Why 3 Replicates Isn’t Enough for RNA-seq (Power, False Positives, and What’s Still Possible)]]></title>
            <link>https://medium.com/@doshiparth2319/why-3-replicates-isnt-enough-for-rna-seq-power-false-positives-and-what-s-still-possible-3d1536fc7d27?source=rss-be40dbc0ea3a------2</link>
            <guid isPermaLink="false">https://medium.com/p/3d1536fc7d27</guid>
            <category><![CDATA[biotechnology]]></category>
            <category><![CDATA[bioinformatics]]></category>
            <category><![CDATA[data-science]]></category>
            <category><![CDATA[research]]></category>
            <category><![CDATA[genomics]]></category>
            <dc:creator><![CDATA[Parth Doshi]]></dc:creator>
            <pubDate>Wed, 28 Jan 2026 11:48:32 GMT</pubDate>
            <atom:updated>2026-01-28T11:48:32.431Z</atom:updated>
            <content:encoded><![CDATA[<p>You spent $500 on RNA-seq and got back a list of 500 “significant” genes.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*bacDI3EXQFHWlcHN8wWr8Q.png" /></figure><p>It sounds like a successful experiment, until you realise you probably missed <strong>800 more</strong>.</p><p>With only <strong>3 biological replicates</strong>, RNA-seq experiments detect just <strong>20–40% of true differential expression</strong>. The rest of real biology never appears in your results.</p><p>This is not an opinion. It has been demonstrated experimentally. In fact, the most-cited RNA-seq experimental design study was designed specifically to answer this question.</p><p>Whether you are planning an RNA-seq study or already sitting on <strong>n=3 RNA-seq data</strong>, understanding this reality is the first step toward extracting <strong>real, actionable insights</strong>.</p><h4>How many replicates are needed for RNA-seq?</h4><blockquote><strong>Short answer:</strong> Most RNA-seq differential expression studies require <strong>at least 6 biological replicates per condition</strong>. Experiments with <strong>3 replicates typically detect only 20–40% of </strong>the true differentially expressed genes and have high false-positive rates.</blockquote><h4>Who Should Read This?</h4><p>This post is for you if:</p><ul><li>You received RNA-seq results and wonder <strong>why so few genes are significant</strong></li><li>Your <strong>qPCR validation keeps failing</strong></li><li>You are working with <strong>n=3 RNA-seq data</strong> and need defensible conclusions</li><li>You are a <strong>PI or project lead</strong> facing underwhelming results</li><li>You are preparing an RNA-seq manuscript and are worried about reviewer criticism</li></ul><h4>Why 3 Biological Replicates Are Not Enough for RNA-seq</h4><p>The definitive study on RNA-seq replicates is <a href="https://pubmed.ncbi.nlm.nih.gov/27022035/"><strong>Schurch et al. 2016</strong></a>, which used <strong>48 biological replicates per condition</strong> to establish ground truth.</p><p>Their findings were unambiguous:</p><iframe src="" width="0" height="0" frameborder="0" scrolling="no"><a href="https://medium.com/media/de5e20cb8af37fde0106dd495f48d71a/href">https://medium.com/media/de5e20cb8af37fde0106dd495f48d71a/href</a></iframe><p>Even at large fold changes (&gt;4×), <strong>n=3 RNA-seq only captures the strongest effects</strong>. Subtle but biologically important changes are statistically invisible.</p><h4><strong>What n=3 Actually Looks Like in Real Data</strong></h4><p>Numbers are abstract. To make the <strong>n=3 problem</strong> tangible, let’s look at a real RNA-seq experiment where the <em>only</em> variable is the number of biological replicates.</p><p>The example below uses a well-known <strong>yeast RNA-seq benchmark dataset (WT vs Δsnf2)</strong> generated with <strong>48 biological replicates per condition</strong>. Because of the unusually large sample size, the full dataset can be treated as an approximate <em>ground truth</em> for differential expression. <br><strong>Same experiment, different replicate numbers</strong></p><p>The same differential expression pipeline was run multiple times using increasing numbers of replicates:</p><ul><li><strong>3 vs 3 replicates</strong> (random subsample)</li><li><strong>6 vs 6 replicates</strong></li><li><strong>12 vs 12 replicates</strong></li><li><strong>All available replicates</strong> (reference set)</li></ul><p>Across all analyses, the same statistical model, normalisation, and FDR threshold were used. Only the number of biological replicates changed.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*IQ5y1uye9T_582TDJ_p7NQ.png" /><figcaption>Volcano plots illustrating the effect of replicate number on differential expression analysis in a human blood RNA-seq dataset (GSE152641; Case vs Control). Low replicate numbers (n = 3 per group) substantially limit statistical power, leading to unstable p-values and the failure to detectmoderate fold-changes. Progressive replication increases stability and reveals a coherent transcriptional response. Results from the full balanced cohort are shown as a reference.</figcaption></figure><h4><strong>What this reveals is striking:</strong></h4><ul><li>With <strong>n=3</strong>, very few genes reach significance</li><li>Moderate fold-change genes are largely missed</li><li>The volcano plot appears diffuse and unstable</li><li>Many genes look significant in one subsample and disappear in another</li></ul><p>As the replicate number increases, the structure of the data changes:</p><ul><li>The signal-to-noise ratio improves</li><li>P-values stabilise</li><li>Moderate but consistent fold-changes become detectable</li><li>True biological signals separate cleanly from background noise</li></ul><p><strong>The hidden instability of n=3</strong></p><p>To quantify this instability, the <strong>3 vs 3 analysis was repeated dozens of times</strong>, each time drawing a different random subset of samples from the same experiment.</p><p><strong>The result:</strong></p><ul><li>The number of significant genes varied widely between runs</li><li>Only <strong>~20–40%</strong> of the genes detected using all replicates were recovered</li><li>The identity of the “top hits” changed substantially depending on which samples were selected</li></ul><p>This means that with <strong>n=3</strong>, two labs analysing the <em>same biological experiment</em> could reach <strong>different conclusions purely by chance</strong>.</p><p>This is not a software problem. It is a <strong>statistical power problem</strong>.</p><h4><strong>Sequencing Depth vs Replicates in RNA-seq</strong></h4><p>A common assumption is:</p><blockquote><em>We sequenced deeply, so replication shouldn’t matter.</em></blockquote><p>This is incorrect.</p><p><a href="https://pubmed.ncbi.nlm.nih.gov/24319002/"><strong>Liu et al. (2014)</strong></a> showed that after <strong>10–20 million reads per sample</strong>, deeper sequencing yields diminishing returns, while adding biological replicates continues to increase power.</p><iframe src="" width="0" height="0" frameborder="0" scrolling="no"><a href="https://medium.com/media/7fc189a7f1d7e06f398eeb3055dcf896/href">https://medium.com/media/7fc189a7f1d7e06f398eeb3055dcf896/href</a></iframe><p><strong>Replication beats depth, consistently and decisively.</strong></p><h4><strong>RNA-seq False Positives and Low Replicate Numbers</strong></h4><p>Low replication not only misses real biology, but it also actively <strong>creates false discoveries</strong>.</p><p>A <strong>Nature Communications</strong> study (<a href="https://www.nature.com/articles/s41467-025-65022-5">Halasz et al., 2025</a>) found:</p><ul><li>At <strong>n=3</strong>, over <strong>38% of DE genes were false positives</strong></li><li>Results with <strong>n ≤ 4</strong> were described as <em>highly misleading</em></li><li><strong>6–7 replicates</strong> were required to control false discovery rates</li></ul><p>This explains why RNA-seq hits often fail qPCR, Western blotting, and downstream functional validation.</p><h4><strong>RNA-seq Replicates and the Reproducibility Crisis</strong></h4><p>Underpowered RNA-seq experiments contribute directly to the reproducibility crisis:</p><ul><li>~$28 billion/year wasted on irreproducible preclinical research (US)</li><li>Amgen replicated only <strong>11%</strong> of landmark cancer studies</li><li>Bayer replicated only <strong>25%</strong> of promising preclinical findings</li><li>Replicated effect sizes are often <strong>80–85% smaller</strong> than originally reported</li></ul><p>Many of these failures trace back to <strong>insufficient biological replication</strong>.</p><h4><strong>RNA-seq Experimental Design: Cost vs Discovery</strong></h4><p>RNA-seq is no longer expensive. Underpowered experiments are.</p><p>Typical Per-Sample Costs (USD, 2024–2025)</p><iframe src="" width="0" height="0" frameborder="0" scrolling="no"><a href="https://medium.com/media/5b41d377e48a25498c0c822e3c394662/href">https://medium.com/media/5b41d377e48a25498c0c822e3c394662/href</a></iframe><p><strong>Cost vs Detection Power (2 Conditions)</strong></p><iframe src="" width="0" height="0" frameborder="0" scrolling="no"><a href="https://medium.com/media/273f7f476d2c917489c58ebc4d819daf/href">https://medium.com/media/273f7f476d2c917489c58ebc4d819daf/href</a></iframe><p>Saving a few hundred dollars on replicates often costs <strong>thousands in failed validation and wasted effort</strong>.</p><p><strong>What You Can Still Learn from RNA-seq with 3 Replicates</strong></p><p>If your data is already generated, do not panic. <strong>n=3 RNA-seq data can still be useful when interpreted conservatively.</strong></p><p><strong>What remains reliable:</strong></p><ul><li>Large fold changes (&gt;4×)</li><li>Pathway and GO enrichment</li><li>Broad biological trends and directionality</li><li>Hypothesis generation for follow-up studies</li><li>Literature-supported genes</li></ul><p><strong>What requires caution:</strong></p><ul><li>Borderline adjusted p-values</li><li>Subtle (&lt;2×) expression changes</li><li>Co-expression or network analysis</li><li>Claims of comprehensive discovery</li></ul><p>If your strongest results disappear under stricter thresholds, that is a <strong>warning, not a failure</strong>.</p><blockquote>Consistent with prior benchmarking studies, low-replicate RNA-seq experiments are best interpreted at the level of large-effect genes and pathways, and should be prioritisedframed as hypothesis-generating rather than comprehensive discovery.</blockquote><h4><strong>How RNA-seq Analysis Must Change with n=3</strong></h4><iframe src="" width="0" height="0" frameborder="0" scrolling="no"><a href="https://medium.com/media/de9b0baa925fe8e35b4a881912444db3/href">https://medium.com/media/de9b0baa925fe8e35b4a881912444db3/href</a></iframe><p>The goal shifts from <strong>finding everything</strong> to <strong>finding what is real and defensible</strong>.</p><h4><strong>The Difference Between RNA-seq Data and Insight</strong></h4><p>A generic pipeline produces p-values and volcano plots.</p><p>Insight-driven RNA-seq analysis produces:</p><ul><li>High-confidence vs exploratory findings</li><li>Biological context for results</li><li>Ranked candidates for validation</li><li>Clear limitations for reviewers</li><li>Actionable next steps</li></ul><p>With low replication, <strong>interpretation matters more than statistics</strong>.</p><p><strong>Key Takeaways</strong></p><ul><li><strong>3 replicates detect only ~30–40% of biology</strong></li><li><strong>Sequencing deeper cannot rescue low replication</strong></li><li><strong>False positives are common at n=3</strong></li><li><strong>Meaningful insight is still possible with the right strategy</strong></li></ul><p>RNA-seq rarely fails because of software. It fails when experimental design and expectations do not align.</p><blockquote><strong>Need Help Interpreting RNA-seq Data?</strong><br> I provide insight-driven RNA-seq and single-cell analysis, with an emphasis on statistical confidence and biological relevance rather than inflated gene lists.</blockquote><blockquote><strong>Services include:</strong> RNA-seq and scRNA-seq analysis, honest power and feasibility assessment, publication-ready figures and methods, and prioritized candidates for validation.</blockquote><p><strong>Contact:</strong> parthdosshi@outlook.com<br> <strong>Upwork:</strong> <a href="https://www.upwork.com/freelancers/~01201b50981e3f3588?mp_source=share">https://www.upwork.com/freelancers/~01201b50981e3f3588?mp_source=share</a></p><p><strong>References</strong></p><p>Halasz, G., Schmahl, J., Negron, N., Ni, M., Lim, W. K., Atwal, G. S., Bai, Y., &amp; Glass, D. J. (2025). Optimized murine sample sizes for RNA sequencing studies revealed from large scale comparative analysis. <em>Nature Communications 2025 16:1</em>, <em>16</em>(1), 10173-. <a href="https://doi.org/10.1038/s41467-025-65022-5">https://doi.org/10.1038/s41467-025-65022-5</a></p><p>Liu, Y., Zhou, J., &amp; White, K. P. (2014). RNA-seq differential expression studies: more sequence or more replication? <em>Bioinformatics (Oxford, England)</em>, <em>30</em>(3), 301–304. <a href="https://doi.org/10.1093/BIOINFORMATICS/BTT688">https://doi.org/10.1093/BIOINFORMATICS/BTT688</a></p><p>Schurch, N. J., Schofield, P., Gierliński, M., Cole, C., Sherstnev, A., Singh, V., Wrobel, N., Gharbi, K., Simpson, G. G., Owen-Hughes, T., Blaxter, M., &amp; Barton, G. J. (2016). How many biological replicates are needed in an RNA-seq experiment and which differential expression tool should you use? <em>RNA</em>, <em>22</em>(6), 839–851. <a href="https://doi.org/10.1261/RNA.053959.115/-/DC1">https://doi.org/10.1261/RNA.053959.115/-/DC1</a></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=3d1536fc7d27" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Batch Effects in RNA-seq: The Hidden Threat to Your Analysis]]></title>
            <link>https://medium.com/@doshiparth2319/batch-effects-in-rna-seq-the-hidden-threat-to-your-analysis-e47004249c46?source=rss-be40dbc0ea3a------2</link>
            <guid isPermaLink="false">https://medium.com/p/e47004249c46</guid>
            <category><![CDATA[genomic-data-science]]></category>
            <category><![CDATA[bioinformatics]]></category>
            <category><![CDATA[rna-seq]]></category>
            <category><![CDATA[data-analysis]]></category>
            <category><![CDATA[single-cell-multiomics]]></category>
            <dc:creator><![CDATA[Parth Doshi]]></dc:creator>
            <pubDate>Thu, 15 Jan 2026 11:33:21 GMT</pubDate>
            <atom:updated>2026-01-15T11:33:21.383Z</atom:updated>
            <content:encoded><![CDATA[<p>Why your “novel biomarker” might be a “Tuesday effect” and the R code to fix it.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*QMN3Viaby_GOwhgcgDIIHg.png" /><figcaption><em>Batch vs Biology.</em> Principal component analysis (PCA) of RNA-seq data before (left) and after (correct) batch correction. Before correction, the samples were clusteredby technical batch rather than biological condition. After correction, the actual biological separation is revealed, highlighting how batch effects can masquerade as a meaningful signal.</figcaption></figure><p>Imagine spending 6–7 months on a transcriptomic project. You meticulously extract RNA and send it for RNA sequencing, and run a data analysis pipeline. The heatmap looks stunning, a distinct cluster of genes upregulated in “Disease” versus “Control.” You’re ready to publish.</p><p>Then a reviewer asks one question: “Were the control and disease samples processed on the same day?”</p><p>You check your lab notes. Controls were processed in January, and disease samples in March.</p><p>That eye-catching pattern might not be biology. It might be a batch effect, the hidden threat of genomic analysis.</p><p><strong>What Is a Batch Effect?</strong></p><p>Batch effects are systematic technical variations that are unrelated to your biological question. RNA-seq is extremely sensitive: room temperature, reagent lot numbers, and even pipetting technique can leave fingerprints on your data.</p><p>In public health, “Hidden threat” describes threats like air pollution, silent damage that accumulates unnoticed. Batch effects operate similarly. They mimic biological processes, form clusters, and yield statistically significant results.</p><p><strong>Step 1: Diagnose Before You Treat</strong></p><p>Never assume. Before correcting anything, visualise your data with PCA.</p><pre>library(DESeq2)<br>library(ggplot2)<br><br>vsd &lt;- vst(dds, blind = FALSE)<br>pcaData &lt;- plotPCA(vsd, intgroup = c(&quot;Condition&quot;, &quot;Batch&quot;), returnData = TRUE)<br>percentVar &lt;- round(100 * attr(pcaData, &quot;percentVar&quot;))<br><br>ggplot(pcaData, aes(PC1, PC2, color = Batch, shape = Condition)) +<br>  geom_point(size = 3) +<br>  xlab(paste0(&quot;PC1: &quot;, percentVar[1], &quot;% variance&quot;)) +<br>  ylab(paste0(&quot;PC2: &quot;, percentVar[2], &quot;% variance&quot;)) +<br>  ggtitle(&quot;PCA: Checking for Batch Effects&quot;)</pre><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*WyyOIgzVE6TJbF6GPvIlrA.png" /><figcaption>Principal Component Analysis (PCA) of publicly available RNA-seq data (GSE49712) illustrating the impact of batch effects and their correction. <strong>Left:</strong> Before correction, samples cluster by sequencing run (Run 1 vs Run 2), indicating a strong batch effect dominating the variance structure. <strong>Right:</strong> After batch correction using ComBat, samples cluster by biological condition (UHRR vs HBRR), demonstrating successful removal of technical variation while preserving accurate biological signal. This figure was generated using public RNA-seq data.</figcaption></figure><p><strong>How to read it:</strong></p><p>Ideal: samples cluster by shape (your biological condition).</p><p>Problem: samples cluster by colour (batch).</p><p>If you see the latter, standard normalisation (TPM, FPKM) won’t help. You need a dedicated correction.</p><p><strong>Step 2: Choose the Right Tool</strong>.</p><p>Selecting the appropriate correction method, such as ComBat-seq, ensures your analysis is based on reliable, modern solutions tailored to RNA-seq data.</p><p>Older methods, such as <strong>ComBat </strong>and <strong>removeBatchEffect</strong>, were developed for microarrays and assume Gaussian distributions. RNA-seq counts follow a negative binomial distribution, apply Gaussian tools, and you’ll get decimals, negatives, and corrupted downstream analysis.</p><p><strong>The modern solution: ComBat-seq.</strong> Released in 2020, it models batch effects under a negative binomial framework and returns integer counts compatible with DESeq2 and edgeR.</p><pre>library(sva)<br>library(DESeq2)<br><br>adjusted_counts &lt;- ComBat_seq(<br>  counts = counts_matrix,<br>  batch = batch,<br>  group = group<br>)<br><br>dds_clean &lt;- DESeqDataSetFromMatrix(<br>  countData = adjusted_counts,<br>  colData = meta_data,<br>  design = ~ Condition<br>)<br>dds_clean &lt;- DESeq(dds_clean)</pre><p><strong>Python users:</strong> pyComBat (via the <strong>inmoose </strong>package) offers a faster implementation if you prefer staying outside R.</p><p><strong>Variant: ComBat-ref:</strong> When batches differ substantially in quality (e.g., samples collected years apart), ComBat-ref identifies the highest-quality batch as the reference and adjusts the others to match it. Use it when you have a clear “gold standard” batch worth preserving.</p><p><strong>A word of caution:</strong> Batch correction can go too far. Overly aggressive correction may remove genuine biological heterogeneity. Re-check PCA afterwards to ensure you’ve preserved biological signals while reducing technical noise.</p><p><strong>Step 3: Confirm It Worked</strong></p><p>Re-running PCA after correction helps you feel in control, ensuring your samples now cluster by condition and your correction was effective.</p><p><strong>Step 4: Know the Limits</strong></p><p>For standard bulk RNA-seq, ComBat-seq handles most scenarios well. But complex datasets, large single-cell atlases, cross-platform integration — sometimes need heavier machinery.</p><p>For single-cell data, <strong>Harmony</strong> (available in both R and Python) has become the de facto standard for batch integration. <strong>STACAS</strong> is gaining traction among those who want semi-supervised correction that preserves known cell type distinctions — addressing the overcorrection concern head-on. Other options include <strong>Scanorama</strong>, <strong>MNN</strong>, and <strong>deep learning</strong> approaches such as <strong>scVI</strong> for particularly challenging datasets.</p><p>These are beyond today’s scope, but worth exploring if ComBat-seq falls short or you’re working at single-cell resolution.</p><p><strong>The One Mistake No Tool Can Fix</strong></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*OYDeBBxXlV2wyFg-lyejDg.png" /><figcaption>Confounded versus balanced RNA-seq study design. When biological groups are processed in separate batches (top), technical batch effects are perfectly aligned with biological factors, making it impossible to distinguish biological from technical differences. In a balanced design (bottom), samples from each biological group are distributed across batches, allowing the accurate biological signal to be distinguished from batch effects.</figcaption></figure><p>If all your controls are in Batch 1 and all disease samples are in Batch 2, batch and biology are perfectly confounded. No algorithm can separate them.</p><p>The consensus among computational biologists is clear: no algorithm can fix a confounded design. Randomise conditions across batches from the outset, consider sample multiplexing for single-cell work, and view batch correction as a safety measure rather than a replacement for solid experimental planning.</p><p><strong>Bottom line:</strong> Batch effects are deceptive, dangerous, and fixable, but only if you design carefully, visualise early, and reach for the right tools. In RNA-seq, technical noise loves to masquerade as discovery. Don’t let it.</p><blockquote><strong>Working with RNA-seq, single-cell, or clinical data?</strong><br> If your results feel unclear or complicated to trust, you’re not alone.<br> I help researchers turn complex datasets into reproducible, publication-ready insights.<br> Before you rerun analyses or submit your work, get an expert opinion.<br> → <a href="https://www.upwork.com/freelancers/~01201b50981e3f3588?mp_source=share">https://www.upwork.com/freelancers/~01201b50981e3f3588?mp_source=share</a></blockquote><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=e47004249c46" width="1" height="1" alt="">]]></content:encoded>
        </item>
    </channel>
</rss>