<?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 Zongmei Gao on Medium]]></title>
        <description><![CDATA[Stories by Zongmei Gao on Medium]]></description>
        <link>https://medium.com/@zongmei08?source=rss-c0d277c14888------2</link>
        <image>
            <url>https://cdn-images-1.medium.com/fit/c/150/150/1*NoG5RnV99ZE2fJZez4asEw@2x.jpeg</url>
            <title>Stories by Zongmei Gao on Medium</title>
            <link>https://medium.com/@zongmei08?source=rss-c0d277c14888------2</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Tue, 17 Nov 2026 03:16:49 GMT</lastBuildDate>
        <atom:link href="https://medium.com/@zongmei08/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[Why Single-Cell Perturbation Models Should Predict Distributions, Not Just Average Cells]]></title>
            <link>https://medium.com/@zongmei08/why-single-cell-perturbation-models-should-predict-distributions-not-just-average-cells-c1ef05161495?source=rss-c0d277c14888------2</link>
            <guid isPermaLink="false">https://medium.com/p/c1ef05161495</guid>
            <dc:creator><![CDATA[Zongmei Gao]]></dc:creator>
            <pubDate>Fri, 06 Nov 2026 21:54:52 GMT</pubDate>
            <atom:updated>2026-11-06T21:54:52.019Z</atom:updated>
            <content:encoded><![CDATA[<p><em>GENESIS reframes perturbation prediction as generating realistic cell populations — and suggests that broader experimental coverage may beat deeper sampling of fewer conditions</em></p><p>Imagine treating the same cell type with the same drug at the same dose twice. The two experiments will not produce perfectly identical cells. Some cells may respond strongly, some weakly, and some not at all. The proportions may shift across replicates even when the underlying biology is reproducible.</p><p>Now imagine a model that predicts only one expression vector: the average treated cell.</p><p>That vector may score well under mean squared error. It may even recover the correct mean expression for every gene. Yet it can still describe a cell that does not exist. If half the cells occupy a low-response state and half occupy a high-response state, the average lies between the two populations. A model that collapses them into one central profile has erased the most interesting biology.</p><p>This is the problem addressed by <strong>GENESIS</strong>, short for <em>Generative ENErgy-based Single-cell Intervention Simulator</em>. In a <a href="https://doi.org/10.64898/2026.10.30.763081">bioRxiv preprint posted on November 1, 2026</a>, Riccardo Conci and colleagues propose a lightweight generative framework that predicts a distribution of possible single-cell states from experimental covariates such as perturbation, dose, and cellular context.</p><p>The important shift is conceptual: the target is no longer one point in gene-expression space. It is a population.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*wMhA-rIumx2fmHcY.png" /></figure><p><em>Figure 1. A model can predict the correct average while missing the response distribution. Distribution-aware prediction attempts to preserve distinct states and their relative frequencies.</em></p><h3>The average-cell trap</h3><p>Suppose a cytokine treatment produces two outcomes:</p><ul><li>60% of cells activate an interferon-response program.</li><li>40% remain close to the untreated state.</li></ul><p>A mean-only model can split the difference and predict moderate activation for every cell. Numerically, the population mean may be correct. Biologically, the model has invented a homogeneous population that was never observed.</p><p>This failure matters whenever heterogeneity is part of the scientific question. Examples include:</p><ul><li>sensitive and resistant tumor-cell states after drug treatment;</li><li>strong, weak, and non-responding immune cells after cytokine stimulation;</li><li>divergent cell-fate decisions after a genetic perturbation;</li><li>rare subpopulations that may drive relapse, toxicity, or adaptation;</li><li>donor-specific mixtures of response states.</li></ul><p>An average is not useless. It is often the right summary for bulk effects, differential expression, or an initial screen. The mistake is treating it as a complete description of a single-cell experiment.</p><h3>What GENESIS tries to learn</h3><p>GENESIS receives the covariates that define an experimental condition and generates multiple single-cell expression profiles for that condition. Instead of asking, “What is the expected cell?”, it asks, “What collection of cells could plausibly appear?”</p><p>The authors evaluate the framework on two large perturbation resources:</p><ol><li><strong>Tahoe-100M</strong>, a recently published atlas containing approximately 100 million transcriptomes from 50 cancer cell lines exposed to 1,100 drug-dose conditions. The <a href="https://doi.org/10.1016/j.cell.2026.08.035">Cell paper</a> describes a pooled “cell village” strategy designed to measure many cellular contexts with reduced batch variation.</li><li>A large peripheral-blood mononuclear cell dataset measuring cytokine responses.</li></ol><p>The preprint reports that GENESIS outperforms several generative models and statistical baselines on its retrospective evaluation splits. It also reports performance approaching the reproducibility of independently repeated conditions.</p><p>That last phrase deserves careful interpretation. “Near experimental reproducibility” does <strong>not</strong> mean that the model has recovered every causal mechanism or that simulation can replace measurement. It means that, under the authors’ evaluation scheme, the distance between prediction and experiment approaches the distance observed between repeated experiments. The apparent ceiling can change with sequencing depth, batch structure, cell composition, and the way cells are sampled.</p><h3>Why train with an energy score?</h3><p>GENESIS uses the <strong>energy score</strong>, a statistical scoring rule for probabilistic predictions. The score was developed in the broader forecasting literature; the foundational treatment by <a href="https://doi.org/10.1198/016214506000001437">Gneiting and Raftery</a> explains why strictly proper scoring rules reward a forecaster, in expectation, for reporting the true predictive distribution rather than a strategically distorted one.</p><p>In intuitive terms, the energy score balances two quantities:</p><ol><li><strong>Fidelity:</strong> generated cells should be close to observed cells.</li><li><strong>Diversity:</strong> generated cells should not all collapse onto the same point.</li></ol><p>The first term pulls predictions toward the data. The second compares generated samples with one another and rewards an appropriate spread. That combination makes the objective better aligned with distribution prediction than a loss based only on a mean profile.</p><p>But a proper score is not magic. In high-dimensional data, one numerical score may be relatively insensitive to some errors in dependency structure. A model can also match a transcriptomic distribution while getting pathway logic, causal direction, rare states, or out-of-distribution behavior wrong. Distributional scores should therefore be combined with biological checks, replicate-aware evaluation, and task-specific summaries.</p><h3>A better benchmark: compare predictions with replicates</h3><p>Single-cell prediction papers often report a model-versus-data distance without telling us how large that distance is relative to experimental variability. A number becomes much more interpretable when it is anchored to two reference points:</p><ul><li><strong>negative control:</strong> a simple baseline or mismatched condition;</li><li><strong>experimental reference:</strong> the difference between independent repeats of the same condition.</li></ul><p>If repeated measurements of a condition differ substantially, expecting a model to match one realization more closely than another replicate may be unrealistic. Conversely, “within replicate distance” is not proof of perfect prediction. A metric can ignore important biology, and repeats can share systematic biases.</p><p>The practical lesson is to design evaluation around experiments, not just around machine-learning splits. Keep replicate identity intact. Avoid letting nearly identical cells from the same experimental unit leak into both training and test sets. Report performance by perturbation, dose, cell context, and response strength rather than only as one global average.</p><h3>The experimental-design result may be the most important part</h3><p>The most provocative result in the preprint is not another leaderboard improvement. It concerns how to allocate a fixed training-cell budget.</p><p>At comparable numbers of training cells, the authors report better prediction when cells were spread across <strong>more experimental conditions at lower depth per condition</strong>, rather than concentrated deeply in fewer conditions.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*UxB7CIFjJQ8mzTeL.png" /></figure><p><em>Figure 2. GENESIS reports an advantage for broader condition coverage at a comparable training-cell budget. This is a study result, not a universal design rule.</em></p><p>Why might breadth help? A conditional generative model learns relationships across perturbations, doses, and cellular backgrounds. Additional conditions reveal how responses change along those axes. Once the model has seen enough structure, extra cells in an already well-sampled condition may add less information than a new condition.</p><p>This could change how teams think about perturbation atlases. Consider a study with 30 compounds, multiple doses, several donors, and multiple cell types. The full Cartesian product quickly becomes unaffordable. A traditional strategy may profile a small subset very deeply to stabilize mean estimates. A model-oriented design might instead sample more of the condition space, preserve genuine replicates, and use a generative model to augment sparse populations or prioritize the next experiments.</p><p>However, “broader is better” must not become a slogan. Some conditions need enough cells to detect rare states, estimate proportions, and separate biology from technical noise. Minimum viable depth depends on cell abundance, effect size, sequencing depth, donor variation, and the decision the experiment must support.</p><h3>What this does — and does not — justify</h3><p>The GENESIS study supports several useful ideas:</p><ul><li>Evaluate whole predicted populations, not only mean profiles.</li><li>Calibrate model error against independent experimental repeats.</li><li>Treat condition coverage as a design variable, not an afterthought.</li><li>Use models to guide where new experiments are most informative.</li></ul><p>It does <strong>not</strong> yet establish that a virtual experiment can replace a wet-lab experiment. The work is a preprint and the reported evidence is based on retrospective splits of existing atlases. A convincing next step would be prospective validation:</p><ol><li>Choose drug-dose-cell-context combinations whose outcomes are still unknown.</li><li>Freeze the training data, model, and evaluation metrics.</li><li>Register the predicted distributions before collecting new data.</li><li>Run independent experiments with genuine biological replicates.</li><li>Compare full distributions, rare-state recovery, gene programs, and decision-relevant outcomes.</li></ol><p>Generalization is another boundary. Tahoe-100M is extremely valuable, but it is dominated by immortalized cancer cell lines measured in a particular assay setting. Performance may change in primary tissue, organoids, diseased donors, different sequencing platforms, or perturbations with mechanisms absent from training.</p><h3>A practical checklist for distribution-aware perturbation prediction</h3><p>Before trusting a model, ask:</p><h3>1. What distribution is being predicted?</h3><p>Is the target raw counts, normalized expression, a latent embedding, pathway scores, or discrete cell states? A realistic latent distribution does not automatically imply realistic gene-level counts.</p><h3>2. Are experimental units separated correctly?</h3><p>Cells from one well are not independent biological replicates. Split by donor, batch, plate, or condition replicate where appropriate; otherwise the model may learn technical fingerprints.</p><h3>3. Does the evaluation preserve heterogeneity?</h3><p>Report means, variances, state proportions, differential programs, and rare-population behavior. Visual agreement in UMAP space is not enough.</p><h3>4. Is uncertainty calibrated?</h3><p>A model should know when it is extrapolating. Test held-out doses, perturbations, cell backgrounds, and combinations separately.</p><h3>5. What experimental decision will the model change?</h3><p>Prediction is most valuable when it selects the next perturbation, narrows a screen, identifies uncertain regions, or allocates replication. A good metric without a decision pathway may not save time or money.</p><h3>The takeaway</h3><p>Single-cell perturbation prediction should not be framed as filling missing entries in an average expression matrix. The more faithful goal is to estimate which cellular states may occur under a condition, in what proportions, and with what uncertainty.</p><p>GENESIS offers a clean formulation of that goal: generate condition-specific cell populations, train with a distribution-aware score, and judge predictions relative to experimental reproducibility. Its early results also suggest a potentially important design principle — when the total cell budget is fixed, broader coverage of the experimental condition space may sometimes be more valuable than measuring a few conditions ever more deeply.</p><p>The model is not a substitute for experiments. Its highest-value role may be deciding <strong>which experiment should come next</strong>.</p><h3>Sources</h3><ul><li>Conci R, D’Ascenzo D, Cultrera di Montesano S, et al. <a href="https://doi.org/10.64898/2026.10.30.763081">Single-cell perturbation prediction near experimental reproducibility with GENESIS</a>. bioRxiv. 2026. Preprint.</li><li>Zhang J, Ubas AA, Svensson V, et al. <a href="https://doi.org/10.1016/j.cell.2026.08.035">Tahoe-100M: Mapping drug-induced molecular phenotypes at single-cell resolution</a>. <em>Cell</em>. 2026.</li><li>Gneiting T, Raftery AE. <a href="https://doi.org/10.1198/016214506000001437">Strictly Proper Scoring Rules, Prediction, and Estimation</a>. <em>Journal of the American Statistical Association</em>. 2007.</li></ul><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=c1ef05161495" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[From One SNP to One Candidate Gene: How GWAS, eQTL, and TWAS Fit Together]]></title>
            <link>https://medium.com/@zongmei08/from-one-snp-to-one-candidate-gene-how-gwas-eqtl-and-twas-fit-together-8cfc0fdaa4e7?source=rss-c0d277c14888------2</link>
            <guid isPermaLink="false">https://medium.com/p/8cfc0fdaa4e7</guid>
            <dc:creator><![CDATA[Zongmei Gao]]></dc:creator>
            <pubDate>Thu, 29 Oct 2026 14:09:25 GMT</pubDate>
            <atom:updated>2026-10-29T14:09:25.566Z</atom:updated>
            <content:encoded><![CDATA[<h3>A practical guide to moving from a statistical association to a defensible biological hypothesis — without confusing proximity, correlation, or prediction with causality</h3><p>A genome-wide association study identifies a significant SNP. Several genes sit nearby. One of them has an eQTL in the same region, and a transcriptome-wide association study later reports that its genetically predicted expression is associated with the trait.</p><p>It is tempting to summarize the result in one sentence:</p><blockquote><em>This SNP regulates this gene, and the gene causes the phenotype.</em></blockquote><p>That sentence may eventually prove correct. But none of the individual analyses establishes the complete chain on its own.</p><p>GWAS asks whether genetic variants are associated with a phenotype. eQTL analysis asks whether variants are associated with gene expression. TWAS asks whether the genetically predicted component of expression is associated with the phenotype. Colocalization asks whether two association patterns are consistent with a shared causal variant. Functional experiments ask whether perturbing the proposed mechanism changes the biological outcome.</p><p>These are related questions, not interchangeable answers.</p><p>This guide connects the methods into one coherent workflow and explains what each step contributes, what it cannot prove, and which checks make a candidate gene genuinely worth following.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*61XvOoCgYN45H1xg4H6qHg.png" /></figure><p><em>Figure 1. A practical evidence chain from variant-level association to a testable</em></p><p><em>gene hypothesis. Each step narrows the interpretation, but no statistical step alone proves mechanism.</em></p><h3>Start with the coordinate system</h3><p>Every post-GWAS analysis depends on a shared genomic coordinate system. A variant such as chr1:160000:C:T only has an unambiguous meaning when the genome build, chromosome naming convention, reference allele, alternate allele, and coordinate convention are known.</p><p>This matters because the same locus may move between genome assemblies, while gene models and transcript annotations may change between releases. In plants and other organisms with complex genomes, homeologs, paralogs, polyploidy, and pangenome variation can make one-to-one mapping even less straightforward.</p><p>Before integrating datasets, verify:</p><ul><li>the reference genome or assembly;</li><li>whether coordinates are one-based or zero-based;</li><li>allele orientation and strand;</li><li>variant identifiers and allele harmonization;</li><li>gene annotation version;</li><li>transcript or isoform definition;</li><li>whether LD and expression reference panels match the study population.</li></ul><p>A sophisticated model cannot rescue mismatched coordinates or reversed alleles.</p><h3>Step 1: GWAS finds a locus, not a mechanism</h3><p>GWAS tests variants across the genome for association with a phenotype. For a quantitative trait, the basic idea can be represented as:</p><p>phenotype ~ genotype dosage + covariates</p><p>Covariates may include age, sex, principal components, environmental variables, experimental batches, or other study-specific factors. Mixed models can additionally account for relatedness and population structure.</p><p>The immediate GWAS output is a set of variant-level statistics: effect estimates, standard errors, P values, allele frequencies, and sometimes conditional signals. A significant SNP therefore supports a statistical statement:</p><blockquote><em>At this locus, genotype is associated with the phenotype under the fitted model.</em></blockquote><p>It does not automatically identify the functional variant or the responsible gene.</p><h3>Why one causal signal produces many associated SNPs</h3><p>Nearby variants are often correlated through linkage disequilibrium, or LD. If SNP A affects the trait, SNPs B, C, and D may also show strong association simply because their genotypes track SNP A in the sampled population.</p><p>The most significant variant is commonly called the lead SNP. “Lead” means that it is a useful statistical representative of the locus. It does not mean that it is the causal nucleotide.</p><p>LD also explains why a Manhattan plot often shows a peak rather than one isolated point. The peak defines a region that requires additional analysis.</p><h3>Why the nearest gene is only a positional candidate</h3><p>A lead SNP may lie inside one gene, near another promoter, and within a regulatory element that contacts a third gene. It may affect transcription, splicing, chromatin accessibility, RNA stability, protein sequence, or none of these directly because it is only tagging another variant.</p><p>Distance is useful for generating candidates, but “nearest gene,” “associated gene,” and “causal gene” are not synonyms.</p><h3>Step 2: Fine-mapping narrows the variant set</h3><p>Statistical fine-mapping uses association strength and LD structure to distribute probability across variants in a locus. Instead of declaring the smallest-P-value SNP causal, it may produce a credible set of variants that together capture a chosen posterior probability.</p><p>Fine-mapping is valuable because it replaces an arbitrary window around the lead SNP with a probabilistic candidate set. Yet its resolution depends on sample size, ancestry, LD accuracy, model assumptions, imputation quality, and the number of causal signals allowed in the region.</p><p>A credible set is therefore a prioritized list, not an experimentally confirmed answer.</p><p>Functional annotations can help rank variants within this set: coding consequence, splice prediction, chromatin accessibility, enhancer marks, transcription-factor motifs, and allele-specific activity. These annotations add biological plausibility but should not be treated as independent proof when they were used to construct the prior.</p><h3>Step 3: eQTL analysis links genotype to expression</h3><p>An expression quantitative trait locus, or eQTL, is a variant associated with the abundance of a gene’s RNA. A simple model is:</p><p>gene expression ~ genotype dosage + covariates</p><p>A cis-eQTL usually refers to a variant tested near the gene, whereas a trans-eQTL influences a more distant gene, potentially on another chromosome. The exact cis window is an analysis choice rather than a biological law.</p><p>If the GWAS locus is also an eQTL for Gene A, we gain an important molecular clue: variation in the region is related both to the trait and to Gene A expression.</p><p>But two cautions are essential.</p><h3>eQTLs are context dependent</h3><p>Regulatory effects can be tissue-, cell-type-, developmental-stage-, ancestry-, sex-, and environment-specific. An eQTL detected in blood may be absent in liver. A signal in bulk tissue may reflect a change in cell composition rather than regulation within a cell type. A response eQTL may appear only after stimulation.</p><p>The most relevant eQTL dataset is not necessarily the largest one; it is the one that best matches the biology of the phenotype.</p><h3>An overlapping eQTL is not yet a shared causal signal</h3><p>The same genomic region can contain multiple causal variants. One could drive the disease association while another, correlated variant drives expression. Their association peaks may look similar because of LD.</p><p>Visual overlap or matching lead SNPs is suggestive, but formal colocalization is stronger.</p><h3>Do not forget splicing and other molecular traits</h3><p>Gene-level expression can miss regulatory effects that change transcript structure without substantially changing total RNA abundance. A splicing QTL, or sQTL, associates genotype with exon usage, intron excision, or transcript proportions.</p><p>Depending on the scientific question, useful molecular QTL layers may also include protein abundance, chromatin accessibility, methylation, metabolites, or cell-state phenotypes. The appropriate mediator is not always total gene expression.</p><h3>Step 4: TWAS converts SNP information into a gene-level association</h3><p>A transcriptome-wide association study starts with a reference dataset containing both genotypes and measured expression. For each gene with sufficient heritable expression, a prediction model estimates weights for nearby SNPs:</p><p>predicted expression = w1*SNP1 + w2*SNP2 + ... + wk*SNPk</p><p>The model may use elastic net, sparse regression, Bayesian approaches, or related strategies. The trained weights are then integrated with a target GWAS, using either individual-level data or GWAS summary statistics plus an LD reference.</p><p>The result tests whether the genetically predicted component of expression is associated with the trait.</p><p>This distinction is crucial. A standard TWAS does not test whether measured expression in the target cohort correlates with phenotype. It tests a genetically predicted component constructed from a reference model.</p><h3>What a significant TWAS result means</h3><p>A defensible interpretation is:</p><blockquote><em>Genetically predicted expression of Gene A is associated with the phenotype under this prediction model and reference context.</em></blockquote><p>It does not, by itself, prove that raising Gene A expression will change the phenotype.</p><p>Prediction weights can tag a nearby trait-associated variant through LD. Multiple neighboring genes may share eQTL architecture and all become significant. A model trained in the wrong tissue or ancestry may poorly represent the relevant regulatory biology. Some genes cannot be tested because their expression is not sufficiently predictable in the available reference panel.</p><p>TWAS is therefore best viewed as gene prioritization, not automatic causal-gene identification.</p><p>Method Statistical unit Primary question What it does not prove GWAS Variant Is genotype associated with the phenotype? The lead SNP is functional, or the nearest gene is causal eQTL Variant-gene pair Is genotype associated with expression? Expression mediates the trait association sQTL Variant-splicing event Is genotype associated with RNA splicing? The affected isoform causes the phenotype TWAS Predicted gene expression Is genetically predicted expression associated with the phenotype? Changing expression will change the phenotype Colocalization Two regional association patterns Are both signals consistent with a shared causal variant? Expression is the causal mediator Functional experiment Perturbation Does changing the variant or gene alter the predicted mechanism? Universal relevance across every biological context</p><h3>Step 5: Colocalization separates two very different stories</h3><p>Colocalization compares the complete regional association patterns for two traits, often GWAS and eQTL signals. Its central question is:</p><blockquote><em>Are these signals consistent with the same causal variant, or with distinct causal variants in LD?</em></blockquote><p>This is more informative than asking whether both datasets contain a significant SNP in the same window.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*gULctXKbQAdvSZSTUUvGxg.png" /></figure><p><em>Figure 2. Two loci can produce a significant GWAS, an eQTL, and even a TWAS</em></p><p><em>signal. Colocalization helps distinguish a shared genetic signal from two different causal variants correlated through LD.</em></p><p>In a classic Bayesian coloc framework, the hypotheses include no association, association with only one trait, two traits with different causal variants, and two traits sharing a causal variant. Strong support for the shared-variant hypothesis increases confidence that the trait and expression signals have the same local genetic origin.</p><p>However, colocalization still does not prove mediation. A shared variant might alter Gene A expression and affect the phenotype through a separate mechanism. Results also depend on priors, variant coverage, allele harmonization, LD, and assumptions about the number of causal variants. Multi-signal loci require conditioning, fine-mapping, or methods designed to accommodate multiple signals.</p><h3>Step 6: Build a convergent evidence package</h3><p>A compelling candidate gene is supported by several lines of evidence that fail in different ways. A practical hierarchy is:</p><ol><li><strong>Robust GWAS locus:</strong> quality-controlled association that survives appropriate structure and multiple-testing correction.</li><li><strong>Fine-mapped variants:</strong> a credible set supported by matching LD and adequate imputation or sequencing.</li><li><strong>Relevant molecular QTL:</strong> eQTL, sQTL, or another QTL in the appropriate tissue, cell type, stage, or condition.</li><li><strong>TWAS association:</strong> a well-performing prediction model with aligned alleles and ancestry-matched LD.</li><li><strong>Colocalization:</strong> support for a shared local genetic signal rather than simple regional overlap.</li><li><strong>Functional coherence:</strong> the variant maps to a plausible regulatory or coding mechanism and the gene fits known biology.</li><li><strong>Perturbation evidence:</strong> allele editing, reporter assays, CRISPR perturbation, knockdown, overexpression, model-organism genetics, or other experiments test the proposed direction.</li></ol><p>No single item is always decisive. Convergence is powerful because statistical, genomic, and experimental evidence have different sources of error.</p><h3>A worked interpretation</h3><p>Imagine a GWAS for seed oil content identifies a locus containing Genes A, B, and C.</p><ul><li>The lead SNP sits in an intron of Gene B.</li><li>Fine-mapping yields six plausible variants, including one near the promoter of Gene A.</li><li>In developing seed tissue, the locus is a cis-eQTL for Gene A but not for Gene B.</li><li>A seed-specific expression model makes Gene A significant in TWAS.</li><li>GWAS and Gene A eQTL signals colocalize after conditioning on a second regional signal.</li><li>The prioritized variant alters a transcription-factor motif in accessible chromatin.</li><li>Editing the motif changes Gene A expression and oil accumulation in the predicted direction.</li></ul><p>The conclusion is not strong because Gene A was the nearest gene. It is strong because multiple independent observations connect the locus, regulatory effect, gene, and phenotype.</p><p>If the experiment instead showed that the GWAS and eQTL signals were driven by distinct variants in LD, Gene A could remain biologically interesting, but the genetic evidence for mediation would be substantially weaker.</p><h3>Where pathway enrichment fits</h3><p>After TWAS, researchers often perform GO, KEGG, Reactome, or ranked gene-set enrichment. These analyses answer whether the prioritized genes concentrate in particular functional categories.</p><p>Use the correct background. If only 5,000 genes had valid expression-prediction models, those 5,000 genes — not every protein-coding gene — are the natural background for an over-representation test.</p><p>Enrichment can reveal a coherent biological theme, but it does not repair an invalid gene list or prove that a pathway causes the phenotype. Regional LD, correlated prediction models, and clustered genes can also violate naive independence assumptions.</p><p>Treat pathway results as a map for follow-up, not the end of the causal argument.</p><h3>A compact reporting checklist</h3><p>Before claiming that a GWAS locus implicates a gene, report:</p><ul><li>genome build, alleles, and harmonization procedure;</li><li>population structure and relatedness controls;</li><li>locus definition and conditional signals;</li><li>LD reference population;</li><li>fine-mapping assumptions and credible sets;</li><li>molecular-QTL tissue, cell type, stage, and sample size;</li><li>expression-model performance and genes that were testable;</li><li>colocalization model, priors, and multi-signal handling;</li><li>whether evidence supports association, a shared variant, mediation, or experimental causality;</li><li>functional experiments and alternative explanations.</li></ul><p>Precision in language is part of reproducibility. “Positional candidate,” “associated gene,” “colocalized gene,” and “functionally validated gene” describe different evidence levels.</p><h3>The central lesson</h3><p>The route from SNP to candidate gene is not a straight arrow. It is an evidence ladder.</p><p>GWAS locates trait-associated regions. Fine-mapping prioritizes variants. eQTL and sQTL analyses connect variants to molecular phenotypes. TWAS reorganizes compatible genetic signals into gene-level associations. Colocalization tests whether molecular and trait signals may share a causal variant. Functional experiments finally challenge the proposed mechanism.</p><p>A significant SNP is not automatically a causal variant. A significant TWAS gene is not automatically a causal gene. A colocalized eQTL is not automatically proof of mediation.</p><p>The goal is not to run the largest number of analyses. It is to ask a distinct question at each step, respect the assumptions behind the answer, and build a biological claim that is no stronger — and no weaker — than the evidence allows.</p><h3>Further reading</h3><ul><li><a href="https://doi.org/10.1016/j.ajhg.2017.06.005">Visscher et al. — 10 Years of GWAS Discovery</a></li><li><a href="https://doi.org/10.1126/science.aaz1776">GTEx Consortium — atlas of genetic regulatory effects across human tissues</a></li><li><a href="https://doi.org/10.1038/ng.3506">Gusev et al. — integrative approaches for large-scale TWAS</a></li><li><a href="https://doi.org/10.1038/s41588-019-0385-z">Wainberg et al. — opportunities and challenges for TWAS</a></li><li><a href="https://doi.org/10.1371/journal.pgen.1004383">Giambartolomei et al. — Bayesian colocalization using summary statistics</a></li><li><a href="https://doi.org/10.1038/s41576-018-0016-z">Schaid et al. — statistical fine-mapping from associations to candidate causal variants</a></li></ul><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=8cfc0fdaa4e7" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[10 Signaling Pathways Every Bioinformatician Should Know]]></title>
            <description><![CDATA[<div class="medium-feed-item"><p class="medium-feed-image"><a href="https://medium.com/@zongmei08/10-signaling-pathways-every-bioinformatician-should-know-dee9104fb8a2?source=rss-c0d277c14888------2"><img src="https://cdn-images-1.medium.com/max/1600/1*OFNAvnNdDeNtPz_vYygpMw.png" width="1600"></a></p><p class="medium-feed-snippet">A practical guide to PI3K, MAPK, NF-kB, Wnt, TGF-beta, JAK/STAT, AMPK, p53, Notch, and Hippo &#x2014; plus how to analyze them without&#x2026;</p><p class="medium-feed-link"><a href="https://medium.com/@zongmei08/10-signaling-pathways-every-bioinformatician-should-know-dee9104fb8a2?source=rss-c0d277c14888------2">Continue reading on Medium »</a></p></div>]]></description>
            <link>https://medium.com/@zongmei08/10-signaling-pathways-every-bioinformatician-should-know-dee9104fb8a2?source=rss-c0d277c14888------2</link>
            <guid isPermaLink="false">https://medium.com/p/dee9104fb8a2</guid>
            <dc:creator><![CDATA[Zongmei Gao]]></dc:creator>
            <pubDate>Tue, 27 Oct 2026 16:26:57 GMT</pubDate>
            <atom:updated>2026-10-27T16:26:57.371Z</atom:updated>
        </item>
        <item>
            <title><![CDATA[ENCODE, GENCODE, ClinVar, GTEx: A Beginner’s Map of Genomics Terms]]></title>
            <description><![CDATA[<div class="medium-feed-item"><p class="medium-feed-image"><a href="https://medium.com/@zongmei08/encode-gencode-clinvar-gtex-a-beginners-map-of-genomics-terms-28650f05baba?source=rss-c0d277c14888------2"><img src="https://cdn-images-1.medium.com/max/640/1*hS2bVqJJaOhwBqKNylOwEw.png" width="640"></a></p><p class="medium-feed-snippet">How to tell databases, assays, atlases, annotations, and QTLs apart &#x2014; and choose the right one for your biological question</p><p class="medium-feed-link"><a href="https://medium.com/@zongmei08/encode-gencode-clinvar-gtex-a-beginners-map-of-genomics-terms-28650f05baba?source=rss-c0d277c14888------2">Continue reading on Medium »</a></p></div>]]></description>
            <link>https://medium.com/@zongmei08/encode-gencode-clinvar-gtex-a-beginners-map-of-genomics-terms-28650f05baba?source=rss-c0d277c14888------2</link>
            <guid isPermaLink="false">https://medium.com/p/28650f05baba</guid>
            <dc:creator><![CDATA[Zongmei Gao]]></dc:creator>
            <pubDate>Tue, 27 Oct 2026 15:28:09 GMT</pubDate>
            <atom:updated>2026-10-27T15:28:09.465Z</atom:updated>
        </item>
        <item>
            <title><![CDATA[Doublet Detection in Single-Cell RNA-seq: A Practical scDblFinder Workflow]]></title>
            <description><![CDATA[<div class="medium-feed-item"><p class="medium-feed-image"><a href="https://medium.com/@zongmei08/doublet-detection-in-single-cell-rna-seq-a-practical-scdblfinder-workflow-2233c1aebfbc?source=rss-c0d277c14888------2"><img src="https://cdn-images-1.medium.com/max/640/1*jrpsbGjxROTrD2kR-KTfwA.png" width="640"></a></p><p class="medium-feed-snippet">When to run it, what &#x201C;sample&#x201D; really means, and how to avoid deleting real biology</p><p class="medium-feed-link"><a href="https://medium.com/@zongmei08/doublet-detection-in-single-cell-rna-seq-a-practical-scdblfinder-workflow-2233c1aebfbc?source=rss-c0d277c14888------2">Continue reading on Medium »</a></p></div>]]></description>
            <link>https://medium.com/@zongmei08/doublet-detection-in-single-cell-rna-seq-a-practical-scdblfinder-workflow-2233c1aebfbc?source=rss-c0d277c14888------2</link>
            <guid isPermaLink="false">https://medium.com/p/2233c1aebfbc</guid>
            <dc:creator><![CDATA[Zongmei Gao]]></dc:creator>
            <pubDate>Mon, 26 Oct 2026 17:38:12 GMT</pubDate>
            <atom:updated>2026-10-26T17:38:12.062Z</atom:updated>
        </item>
        <item>
            <title><![CDATA[Ten Classic Signaling Pathways Every Biologist Should Recognize]]></title>
            <description><![CDATA[<div class="medium-feed-item"><p class="medium-feed-image"><a href="https://medium.com/@zongmei08/ten-classic-signaling-pathways-every-biologist-should-recognize-d0032eddc847?source=rss-c0d277c14888------2"><img src="https://cdn-images-1.medium.com/max/2400/1*mYGC4-EG0gmRYA9SEsCt0g.png" width="2400"></a></p><p class="medium-feed-snippet">A practical guide to GPCR, MAPK, PI3K, JAK-STAT, NF-kB, TGF-beta, Wnt, Notch, Hedgehog, and Hippo signaling</p><p class="medium-feed-link"><a href="https://medium.com/@zongmei08/ten-classic-signaling-pathways-every-biologist-should-recognize-d0032eddc847?source=rss-c0d277c14888------2">Continue reading on Medium »</a></p></div>]]></description>
            <link>https://medium.com/@zongmei08/ten-classic-signaling-pathways-every-biologist-should-recognize-d0032eddc847?source=rss-c0d277c14888------2</link>
            <guid isPermaLink="false">https://medium.com/p/d0032eddc847</guid>
            <dc:creator><![CDATA[Zongmei Gao]]></dc:creator>
            <pubDate>Fri, 23 Oct 2026 16:05:39 GMT</pubDate>
            <atom:updated>2026-10-23T16:05:39.448Z</atom:updated>
        </item>
        <item>
            <title><![CDATA[Gene Expression Regulation: From the Lac Operon to Chromatin, Enhancers, and RNAi]]></title>
            <description><![CDATA[<div class="medium-feed-item"><p class="medium-feed-image"><a href="https://medium.com/@zongmei08/gene-expression-regulation-from-the-lac-operon-to-chromatin-enhancers-and-rnai-9efdcd6e5979?source=rss-c0d277c14888------2"><img src="https://cdn-images-1.medium.com/max/2400/1*AlWg5x2r36bgDKk0DNLr9g.png" width="2400"></a></p><p class="medium-feed-snippet">A beginner-friendly map of how cells decide when, where, and how strongly a gene should be expressed</p><p class="medium-feed-link"><a href="https://medium.com/@zongmei08/gene-expression-regulation-from-the-lac-operon-to-chromatin-enhancers-and-rnai-9efdcd6e5979?source=rss-c0d277c14888------2">Continue reading on Medium »</a></p></div>]]></description>
            <link>https://medium.com/@zongmei08/gene-expression-regulation-from-the-lac-operon-to-chromatin-enhancers-and-rnai-9efdcd6e5979?source=rss-c0d277c14888------2</link>
            <guid isPermaLink="false">https://medium.com/p/9efdcd6e5979</guid>
            <dc:creator><![CDATA[Zongmei Gao]]></dc:creator>
            <pubDate>Thu, 22 Oct 2026 17:49:50 GMT</pubDate>
            <atom:updated>2026-10-22T17:49:50.317Z</atom:updated>
        </item>
        <item>
            <title><![CDATA[How to Read qPCR Results: Ct Values, Melt Curves, and 2^-ΔΔCt Without Guessing]]></title>
            <description><![CDATA[<div class="medium-feed-item"><p class="medium-feed-image"><a href="https://medium.com/@zongmei08/how-to-read-qpcr-results-ct-values-melt-curves-and-2-%CE%B4%CE%B4ct-without-guessing-eb5f7bd17cdf?source=rss-c0d277c14888------2"><img src="https://cdn-images-1.medium.com/max/1916/1*D1ZZ3VutoqTEi-y6OtHAMA.png" width="1916"></a></p><p class="medium-feed-snippet">A practical workflow for deciding whether the run worked, calculating relative expression, and troubleshooting curves before they become&#x2026;</p><p class="medium-feed-link"><a href="https://medium.com/@zongmei08/how-to-read-qpcr-results-ct-values-melt-curves-and-2-%CE%B4%CE%B4ct-without-guessing-eb5f7bd17cdf?source=rss-c0d277c14888------2">Continue reading on Medium »</a></p></div>]]></description>
            <link>https://medium.com/@zongmei08/how-to-read-qpcr-results-ct-values-melt-curves-and-2-%CE%B4%CE%B4ct-without-guessing-eb5f7bd17cdf?source=rss-c0d277c14888------2</link>
            <guid isPermaLink="false">https://medium.com/p/eb5f7bd17cdf</guid>
            <dc:creator><![CDATA[Zongmei Gao]]></dc:creator>
            <pubDate>Tue, 20 Oct 2026 18:50:37 GMT</pubDate>
            <atom:updated>2026-10-20T18:50:37.105Z</atom:updated>
        </item>
        <item>
            <title><![CDATA[AI Virtual Cells Are Not Digital Cells — Yet]]></title>
            <description><![CDATA[<div class="medium-feed-item"><p class="medium-feed-image"><a href="https://medium.com/@zongmei08/ai-virtual-cells-are-not-digital-cells-yet-4c2556920ad0?source=rss-c0d277c14888------2"><img src="https://cdn-images-1.medium.com/max/2000/1*17_8VHe0bdbOaA-0W2HiOw.png" width="2000"></a></p><p class="medium-feed-snippet">What these models actually predict, why perturbation data matters, and how to read the first benchmarks without the hype</p><p class="medium-feed-link"><a href="https://medium.com/@zongmei08/ai-virtual-cells-are-not-digital-cells-yet-4c2556920ad0?source=rss-c0d277c14888------2">Continue reading on Medium »</a></p></div>]]></description>
            <link>https://medium.com/@zongmei08/ai-virtual-cells-are-not-digital-cells-yet-4c2556920ad0?source=rss-c0d277c14888------2</link>
            <guid isPermaLink="false">https://medium.com/p/4c2556920ad0</guid>
            <dc:creator><![CDATA[Zongmei Gao]]></dc:creator>
            <pubDate>Mon, 19 Oct 2026 15:10:09 GMT</pubDate>
            <atom:updated>2026-10-19T15:10:09.796Z</atom:updated>
        </item>
        <item>
            <title><![CDATA[Single-Cell Metabolism Without Four Separate PipelinesA practical, no-code workflow for pathway…]]></title>
            <description><![CDATA[<div class="medium-feed-item"><p class="medium-feed-image"><a href="https://medium.com/@zongmei08/single-cell-metabolism-without-four-separate-pipelinesa-practical-no-code-workflow-for-pathway-9e727c0007c9?source=rss-c0d277c14888------2"><img src="https://cdn-images-1.medium.com/max/2000/1*N92aHN4sA2DcvpSABK2FOw.png" width="2000"></a></p><p class="medium-feed-snippet">Single-cell RNA sequencing can tell us which metabolic genes are expressed in each cell. Turning that expression matrix into a metabolic&#x2026;</p><p class="medium-feed-link"><a href="https://medium.com/@zongmei08/single-cell-metabolism-without-four-separate-pipelinesa-practical-no-code-workflow-for-pathway-9e727c0007c9?source=rss-c0d277c14888------2">Continue reading on Medium »</a></p></div>]]></description>
            <link>https://medium.com/@zongmei08/single-cell-metabolism-without-four-separate-pipelinesa-practical-no-code-workflow-for-pathway-9e727c0007c9?source=rss-c0d277c14888------2</link>
            <guid isPermaLink="false">https://medium.com/p/9e727c0007c9</guid>
            <dc:creator><![CDATA[Zongmei Gao]]></dc:creator>
            <pubDate>Sat, 17 Oct 2026 20:55:41 GMT</pubDate>
            <atom:updated>2026-10-17T20:55:41.579Z</atom:updated>
        </item>
    </channel>
</rss>