<?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 Semilogo Oketola on Medium]]></title>
        <description><![CDATA[Stories by Semilogo Oketola on Medium]]></description>
        <link>https://medium.com/@semilogooketola?source=rss-8a02bfe1359e------2</link>
        <image>
            <url>https://cdn-images-1.medium.com/fit/c/150/150/1*vUYXKNYZMird5ThW13calw@2x.jpeg</url>
            <title>Stories by Semilogo Oketola on Medium</title>
            <link>https://medium.com/@semilogooketola?source=rss-8a02bfe1359e------2</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Wed, 18 Nov 2026 18:05:00 GMT</lastBuildDate>
        <atom:link href="https://medium.com/@semilogooketola/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[PancrionDX: What Happened When I Tried to Make a Multimodal Model for Pancreatic Ductal…]]></title>
            <link>https://semilogooketola.medium.com/pancriondx-what-happened-when-i-tried-to-make-a-multimodal-model-for-pancreatic-ductal-b94096fe1a89?source=rss-8a02bfe1359e------2</link>
            <guid isPermaLink="false">https://medium.com/p/b94096fe1a89</guid>
            <category><![CDATA[ai-in-healthcare]]></category>
            <category><![CDATA[multi-omic]]></category>
            <category><![CDATA[pancreatic-cancer]]></category>
            <dc:creator><![CDATA[Semilogo Oketola]]></dc:creator>
            <pubDate>Sun, 24 May 2026 00:23:00 GMT</pubDate>
            <atom:updated>2026-06-13T19:43:35.039Z</atom:updated>
            <content:encoded><![CDATA[<h3>PancrionDX: What Happened When I Tried to Make a Multimodal Model for Pancreatic Ductal Adenocarcinoma</h3><iframe src="https://cdn.embedly.com/widgets/media.html?src=https%3A%2F%2Fwww.youtube.com%2Fembed%2FXvpNtwmLu4g%3Ffeature%3Doembed&amp;display_name=YouTube&amp;url=https%3A%2F%2Fwww.youtube.com%2Fwatch%3Fv%3DXvpNtwmLu4g&amp;image=https%3A%2F%2Fi.ytimg.com%2Fvi%2FXvpNtwmLu4g%2Fhqdefault.jpg&amp;type=text%2Fhtml&amp;schema=youtube" width="854" height="480" frameborder="0" scrolling="no"><a href="https://medium.com/media/35ad531fa6529ca4023c27e69f93c7f8/href">https://medium.com/media/35ad531fa6529ca4023c27e69f93c7f8/href</a></iframe><p>I’ve written about Syngenix. I’ve written about PanEcho. Both of those were about individual tools tackling individual data types, one genomic, one imaging.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*6G8gwd3NQ1Yh8zJ2t0BpDA.png" /><figcaption>PancrionDX</figcaption></figure><p>PancrionDX was always the bigger vision behind both of those projects: a multimodal classification system for pancreatic ductal adenocarcinoma that pulls from CT imaging, transcriptomic expression profiles, and genomic mutation data simultaneously. The hypothesis: A single data modality can only tell you so much. Tumors are multidimensional, so the model should be too.</p><p>What I actually found when I built it was more interesting than I expected, and more humbling than I wanted.</p><h3>The Question I Was Actually Asking</h3><p>Most early PDAC detection efforts fall into one of two camps: imaging-based or genomics-based. Imaging tells you what the tumor looks like right now. Genomics tells you what’s happening at the molecular level. But stage classification, knowing whether a tumor is early or advanced, requires understanding <em>progression</em>. And progression is a dynamic process. No single snapshot fully captures it.</p><p>So the core question driving PancrionDX wasn’t just “can AI detect PDAC?” It was something more specific: does combining modalities actually improve classification, and if so, which ones are carrying the real signal? That distinction matters. Adding data sources to a model doesn’t automatically make it smarter. It can make it noisier. I wanted to know whether multimodal fusion was genuinely useful or just a more complicated way to get similar results.</p><h3>Building the Dataset, and Why That Was Half the Problem</h3><p>Before training anything, I had to figure out what I was training on.</p><p>I grounded the dataset in TCGA-PAAD literature, generating synthetic patient files that reflected biologically realistic variability across three modalities: radiomic features extracted from CT imaging, transcriptomic expression profiles, and genomic mutation presence data. The word <em>realistic</em> is doing a lot of work in that sentence. I didn’t just randomly sample features across a plausible range. I built in mutation prevalence rates consistent with what’s actually documented in PDAC research, pathway co-dependencies between genes, and transcriptomic heterogeneity that approximates real tumor behavior across patients.</p><p>That decision turned out to be the most consequential one I made in the entire project. And it was the first thing that broke my initial assumptions.</p><h3>What Biologically Realistic Data Did to the Genomic Model</h3><p>Here’s what I expected going in: all three modalities would contribute meaningfully to classification, and combining them would push performance higher than any individual model.</p><p>Here’s what actually happened: the moment I introduced biologically accurate mutation frequencies into the training data, the genomic-only model fell apart.</p><p>The reason makes sense in retrospect, even if it caught me off guard at the time. Driver mutations like KRAS and TP53 are nearly universal in PDAC. KRAS is mutated in over 90% of cases. TP53 isn’t far behind. If almost every tumor carries the same mutations regardless of what stage it’s at, then mutation presence can’t tell you anything about stage. It’s like trying to distinguish runners in a race by checking whether they’re wearing shoes. Everyone’s wearing shoes. That doesn’t tell you who’s ahead.</p><p>Mutation presence reflects whether cancer exists. It doesn’t reflect how far it’s progressed. Stage is a function of <em>dynamics</em>, of how the tumor has grown, invaded surrounding tissue, and remodeled its local environment over time. That information doesn’t live in a mutation flag. It lives somewhere else.</p><p>Before I introduced biological realism into the dataset, the genomic model looked fine. Artificially inflated or randomly distributed mutation signals gave it something to work with. Once I forced it to reflect actual PDAC mutation frequencies, that false signal collapsed and the model had nothing left to stand on. That’s actually the most important result in this whole project, not because it’s a good outcome for the genomic model, but because it’s an honest one. A model that only works on unrealistic data would have fallen apart the first time anyone tested it on real patients.</p><h3>Where the Real Signal Lives</h3><p>Once the genomic model degraded, transcriptomic and radiomic models retained their predictive power. That told me something important: disease stage is better captured by dynamic, state-dependent features than by static mutation flags.</p><p>In the full multimodal setting, radiomic features dominated. They accounted for over 80% of model importance. Transcriptomics contributed a meaningful secondary signal. Genomic mutation presence was nearly irrelevant.</p><p>That result deserves some unpacking, because it’s easy to misread.</p><h4>The Radiomic Model: Tumor Phenotype as a Stage Marker</h4><p>The core claim behind the radiomic component was this: a tumor’s physical appearance in imaging encodes real information about where it is in its progression. Not just whether it’s present, but how advanced it is and what it’s doing to surrounding tissue.</p><p>Here’s what each feature in the radiomic model was actually measuring:</p><p><strong>Tumor Size</strong> tells you about growth over time. Larger tumors tend to correspond to later-stage disease, not because size defines stage, but because it reflects how long and how aggressively the tumor has been growing.</p><p><strong>Tumor Compactness</strong> captures shape regularity. Irregular, non-compact tumors suggest invasive growth patterns. A well-defined, compact mass behaves very differently from one with spreading, poorly-defined edges.</p><p><strong>Texture Heterogeneity</strong> measures internal imaging disorder. Aggressive tumors tend to have more chaotic internal structure, reflecting necrosis, vascular disorganization, and rapid uncoordinated cell proliferation happening inside the mass.</p><p><strong>Edge Sharpness</strong> is about containment. Blurry tumor borders in imaging suggest the cancer is actively invading surrounding tissue rather than remaining localized. Sharp, well-defined edges indicate a more contained lesion.</p><p><strong>Necrosis Score</strong> tracks dead tissue regions within the tumor. Necrosis tends to develop when tumor growth outpaces its own blood supply, which is more common in larger, more advanced tumors.</p><p><strong>Intensity Variance</strong> measures variation in the imaging signal across the tumor mass. High variance reflects biological heterogeneity within the tumor itself, which tends to increase with disease progression.</p><p>None of these features are arbitrary choices. Each one is a proxy for a biological process that’s associated with stage advancement. The radiomic model wasn’t just finding patterns in pixel values. It was detecting physical evidence of how far the disease had progressed.</p><p>That’s why it dominated the multimodal model. Radiomic features are carrying phenotypic information that mutation data simply doesn’t contain.</p><h4>What Transcriptomics Adds</h4><p>Transcriptomic features contributed secondary signal in the multimodal setting, and that also makes biological sense. Gene expression reflects what the tumor is actively doing at a functional level. Expression patterns shift as tumors progress: invasion and angiogenesis pathways get upregulated, cell adhesion and apoptosis pathways get suppressed. These molecular changes accompany and often precede the morphological shifts that radiomics picks up on the CT scan.</p><p>So transcriptomics and radiomics aren’t redundant. They’re measuring related but distinct aspects of the same underlying progression process. Transcriptomics captures the molecular activity. Radiomics captures the physical consequences. Used together, they cover more of the story than either does alone.</p><h3>What Fusion Actually Bought</h3><p>This was the part I was most uncertain about going in.</p><p>Combining modalities doesn’t automatically improve a model. If one modality is mostly noise, including it can hurt performance. If two modalities are highly correlated, you’re not gaining new information by fusing them. Fusion only helps when the individual signals are genuinely capturing independent aspects of the same phenomenon.</p><p>In the case of PancrionDX, fusion worked. Not just in terms of accuracy but in terms of robustness.</p><p>When I ran genomic-only experiments with realistic mutation frequencies, performance degraded badly. But in the multimodal fusion model, the radiomic and transcriptomic signals compensated. The combined model didn’t just outperform the genomic model. It was more <em>stable</em> than any single-modality model I ran. That stability is the real finding here, more than any specific accuracy number.</p><p>Single-modality models are brittle because they depend entirely on one signal being reliable. The fusion model absorbed the fact that genomic features weren’t contributing useful information and still produced accurate predictions because other channels were. That fault-tolerance is what multimodal learning is actually for in a clinical context. You want a system that doesn’t collapse when one data type is noisy or uninformative for a given patient.</p><h3>The Numbers, and What They Actually Mean</h3><p>The performance metrics are encouraging: around 91% accuracy, an ROC-AUC of roughly 0.92, precision near 94%, recall around 89%. For a proof-of-concept system, that’s a genuinely strong signal.</p><p>But I want to be honest about what those numbers mean and where they fall short.</p><p>The training data is synthetic. It’s grounded in TCGA-PAAD literature and designed to approximate real tumor biology, but no real patient records were used. That means the model has learned patterns from a carefully constructed simulation of reality. How well those patterns transfer to actual clinical data is still an open question. More than that, it’s the central open question that determines whether any of this is actually useful.</p><p>There’s also an interpretability caveat worth naming explicitly. I observed gene-level feature contributions from markers including BRCA1 and CHEK2. That’s interesting, and worth following up on. But gene-level importance scores from a model trained on synthetic data could reflect dataset structure as much as true biology. I’m not ready to claim those findings are clinically meaningful without validation on independent patient cohorts. The signals are worth investigating. They’re not yet worth acting on.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*G5aNHUyD7lohadHIDvpL_g.png" /><figcaption>Analysis Results From Testing PancrionDX</figcaption></figure><h3>What I Learned That I Didn’t Expect to Learn</h3><p>The biggest takeaway from PancrionDX wasn’t about machine learning architecture or feature engineering. It was about what questions different data types are actually capable of answering.</p><p>I went in treating this primarily as a modeling problem. Get the architecture right, tune the hyperparameters, run the experiments. That part was manageable. The harder problem was biological: what does each data type actually encode, and is that encoding informative for the specific task of stage classification?</p><p>Mutation presence encodes cancer status. It doesn’t encode cancer stage. That distinction eliminated one-third of my initial model assumptions. Once I accepted it, the direction became clearer: build around features that encode dynamics and progression, not just the presence or absence of genetic alterations.</p><p>The other thing I didn’t fully anticipate was how much biological realism functions as a form of stress testing. By designing the training data to reflect true PDAC mutation frequencies rather than artificially inflated ones, I surfaced a failure mode before it could cause problems downstream. An unrealistic dataset would have produced a genomic model that looked functional in development and failed on real patients. Forcing the data to be realistic broke the model early. That’s the right order to discover things.</p><p>It also reinforced something I already suspected from building Syngenix: the quality and structure of your training data shapes everything. You can’t rescue a model from bad data with a better algorithm. The preprocessing and data design decisions are where the real work happens.</p><h3>What Comes Next</h3><p>PancrionDX has moved past the question of whether multimodal learning can work for PDAC stage classification. That question has a provisional answer: it can, and the dominant signal is radiomic.</p><p>The open questions now are about generalization. Does the model hold up beyond TCGA-PAAD derived patterns? How does it behave when tested against actual clinical imaging acquired with varying CT protocols? Can the radiomic features be extracted reliably from real scans rather than synthetic ones? How do the transcriptomic contributions hold up in independent patient cohorts with different expression distributions?</p><p>None of those questions can be answered without external validation data, which means the next phase involves getting the system in front of real clinical datasets and testing it against ground truth it wasn’t trained on. That requires collaboration with oncologists and access to patient data that I don’t currently have. That’s the gap between where this project is now and where it needs to get to before it means anything outside of a development context.</p><p>I also need to be more disciplined about the pipeline architecture. Right now, parts of the inference layer are not as cleanly separated from the experimental scaffolding as they should be. That creates fragility. For the model to be stable enough to validate properly, the inference pipeline needs to produce identical outputs for identical inputs every single time, without any variability introduced by the surrounding code. That’s the next engineering priority before the research priorities.</p><h3>Closing Thoughts</h3><p>The original hypothesis was that combining radiomic, transcriptomic, and genomic signals would outperform single-modality approaches for PDAC stage classification. What the experiment actually showed is that it does, but not for the reason I initially assumed.</p><p>The gain from fusion isn’t that all three modalities are contributing equally useful information. The gain is that fusion produces a more robust system that can still function when one signal fails. The genomic signal failed when tested realistically. The multimodal model didn’t.</p><p>That result changes how I think about the project going forward. PancrionDX isn’t trying to find the best modality for PDAC classification. It’s trying to build a system that’s fault-tolerant enough to be clinically useful. Fault-tolerant means it degrades gracefully when data is missing or uninformative, that it doesn’t overclaim certainty, and that its performance in testing reflects how it would actually behave with real patients.</p><p>The model is past proof-of-concept. The engineering and validation work is where the real challenge starts.</p><h3>Try It / Learn More</h3><p><strong>Live demo:</strong> <a href="https://pancriondx.vercel.app/">/https://pancriondx.vercel.app/</a><br><strong>Contact:</strong> <a href="https://www.linkedin.com/in/semilogo-oketola/">LinkedIn</a> or <a href="http://semilogooketola@outlook.com/">Email</a> if you want to chat about this project</p><p><strong>For medical professionals:</strong> If you’re a radiologist or oncologist interested in validating this approach, I’d love to talk. Seriously, please reach out.</p><p><strong>For patients/concerned individuals:</strong> Please don’t use this for actual medical decisions yet. It’s not validated. Talk to your doctor about pancreatic cancer screening options (endoscopic ultrasound, MRI/MRCP, etc.).</p><p>If you found this interesting, take a look at my <a href="https://www.linkedin.com/in/semilogo-oketola/">LinkedIn</a>, and <a href="https://semilogooketola.lovable.app/">personal website</a></p><p><em>PancrionDX is part of a larger early detection platform for pancreatic cancer. If you’re a researcher, clinician, or someone working on oncology AI and want to talk through any of this, my </em><a href="https://www.linkedin.com/in/semilogo-oketola/"><em>LinkedIn</em></a><em> is open.</em></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=b94096fe1a89" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Building PanEcho: What I Learned Creating an AI Tool to Help Catch Pancreatic Cancer Early]]></title>
            <link>https://semilogooketola.medium.com/building-panecho-what-i-learned-creating-an-ai-tool-to-help-catch-pancreatic-cancer-early-396e43a415e8?source=rss-8a02bfe1359e------2</link>
            <guid isPermaLink="false">https://medium.com/p/396e43a415e8</guid>
            <category><![CDATA[pancreatic-cancer]]></category>
            <category><![CDATA[radiomics]]></category>
            <category><![CDATA[earlydetectionsaveslives]]></category>
            <category><![CDATA[machine-learning]]></category>
            <dc:creator><![CDATA[Semilogo Oketola]]></dc:creator>
            <pubDate>Sat, 31 Jan 2026 00:10:51 GMT</pubDate>
            <atom:updated>2026-02-03T12:12:03.390Z</atom:updated>
            <content:encoded><![CDATA[<iframe src="https://cdn.embedly.com/widgets/media.html?src=https%3A%2F%2Fwww.loom.com%2Fembed%2Ff8ccf118b29d4fabb522f19c8f0f80d3&amp;display_name=Loom&amp;url=https%3A%2F%2Fwww.loom.com%2Fshare%2Ff8ccf118b29d4fabb522f19c8f0f80d3&amp;image=https%3A%2F%2Fcdn.loom.com%2Fsessions%2Fthumbnails%2Ff8ccf118b29d4fabb522f19c8f0f80d3-3a2a9077752db016.gif&amp;type=text%2Fhtml&amp;schema=loom" width="1110" height="832" frameborder="0" scrolling="no"><a href="https://medium.com/media/fcab27736601f07f4edafd12c1cd7c1d/href">https://medium.com/media/fcab27736601f07f4edafd12c1cd7c1d/href</a></iframe><p>So, here’s the thing about pancreatic cancer: it’s terrifying. Not in a movie-villain kind of way, but in a quiet, sneaky way that makes it one of the deadliest cancers out there. Most people don’t even know they have it until it’s way too late to do anything about it. The 5-year survival rate? 12%. That’s it.</p><p>This isn’t my first time tackling this problem. A few months back, I built Syngenix, a tool that uses RNA-seq data and machine learning to predict pancreatic cancer risk from gene expression patterns. That project taught me a lot about the molecular side of PDAC (Pancreatic Ductal Adenocarcinoma), but it had a fundamental limitation: it required expensive genomic testing that most people don’t have access to.</p><p>That got me thinking: what if we could use something people already get: CT scans? Imaging is way more accessible than RNA sequencing. Almost everyone at risk for pancreatic cancer gets an abdominal CT scan at some point. The challenge is that subtle early-stage tumors can be missed by human radiologists.</p><p>Enter PanEcho.</p><p><strong>The bigger picture:</strong> Both Syngenix and PanEcho are actually pieces of a larger project I’m working on called PancrionDX, a comprehensive pancreatic cancer screening platform that combines multiple data sources (imaging, genomics, clinical risk factors) to catch this disease as early as possible.</p><p>Full transparency: I’m not a doctor. I’m not a radiologist. I’m not a geneticist. I’m just a student that genuinely believes that AI can help save lives if we use it responsibly.</p><h3>Project Overview</h3><p><strong>What I Built:</strong> An AI-powered web app for pancreatic cancer screening from CT scans</p><p><strong>How it fits into PancrionDX:</strong> Imaging analysis component (complements Syngenix’s genomic analysis)</p><h3>The Core Features</h3><ul><li>Upload CT scans (supports DICOM medical imaging format + regular images)</li><li>AI analysis that looks for concerning patterns in the pancreas</li><li>Risk scoring (0–100%) with clear Normal vs. Suspicious classification</li><li>Longitudinal tracking: upload follow-up scans and see if things are changing over time</li><li>PDF reports you can actually share with your doctor</li><li>Accessibility: designed to work with screen readers and keyboard navigation</li></ul><h3>How PanEcho Complements Syngenix</h3><p>→ Syngenix (Genomic):</p><ul><li>Analyzes RNA-seq gene expression data</li><li>Requires expensive genomic testing</li><li>Catches molecular changes before tumors form</li><li>High accuracy but limited accessibility</li></ul><p>→ PanEcho (Imaging)</p><ul><li>Analyzes CT scan images</li><li>Uses commonly available imaging</li><li>Catches structural/anatomical changes</li><li>Lower barrier to entry, widely accessible</li></ul><p><strong>The PancrionDX vision:</strong> Combine both approaches. If you have genomic data, Syngenix can flag molecular risk. If you have imaging, PanEcho can flag anatomical concerns. Together, they provide a more complete picture than either alone.</p><h3>Tech Stack</h3><p>I went with the Next.js + Supabase + DeepMind stack because honestly, it just works:</p><ul><li><strong>Frontend:</strong> Next.js 14, React, TypeScript, Tailwind CSS</li><li><strong>Backend:</strong> Next.js Server Actions</li><li><strong>Database:</strong> Supabase (PostgreSQL with built-in auth)</li><li><strong>Storage:</strong> Supabase Storage</li><li><strong>AI:</strong> Gemini 3</li><li><strong>DICOM Processing:</strong> dcmjs + sharp</li><li><strong>Hosting:</strong> Vercel</li></ul><p><strong>Why different from Syngenix’s stack?</strong> Syngenix used Flask + scikit-learn because it was focused on traditional ML classifiers with genomic data. PanEcho uses Next.js + GPT-4o because vision models are better suited for image analysis, and GPT-4o’s prompt-based approach is more flexible than training custom models.</p><h3>The Problem I Was Trying to Solve</h3><p>Here’s what makes pancreatic cancer especially scary:</p><p>It’s silent with most people have zero symptoms until the tumour is already large, it grows fast that by the time you notice something’s wrong, it’s often spread. Early signs are subtle. A radiologist might glance at a CT scan and think “probably nothing” when there’s actually a tiny tumour forming. And sadly, access is limited. Not everyone can get their scans reviewed by a pancreatic specialist.</p><p>The math is brutal: If you catch it at stage 1, survival rate jumps to 44%. If you catch it at stage 4, it drops to 3%. Early detection literally means the difference between life and death.</p><p>When I built Syngenix, I learned that RNA-seq-based screening has a critical adoption problem: cost ($200–1,000+ per test), accessibility (most hospitals don’t routinely do gene expression profiling), and insurance rarely covers it for screening.</p><p>Meanwhile, CT scans are already routine, insurance covered, widely available, and provide immediate results. <strong>The opportunity:</strong> If we can make CT scan analysis more accurate using AI, we can screen way more people at a fraction of the cost of genomic testing.</p><h3>How I Built It (The Technical Stuff)</h3><p>The flow is pretty straightforward:</p><p>User uploads scan → DICOM processor extracts key images → GPT-4o analyzes those images → Results displayed with risk score → User can download PDF report</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*vxObcrXvfvYgjTCBQPkTJQ.png" /></figure><h3>Key Technical Decisions</h3><p><strong>1. Why Next.js 14 with Server Actions?</strong></p><p>Server Actions keep everything in one place. Your API logic lives right next to your UI components. Plus, it keeps API keys on the server where they belong meaning that there’s no chance of accidentally exposing my API key to the client.</p><p><strong>2. Why Supabase?</strong></p><p>Because I value my time and sanity. Supabase provides me with PostgreSQL database, built-in authentication, Row Level Security, and file storage — all on a free tier that’s actually usable. Could I have built all this myself? Sure. Would it have taken 10x longer? Absolutely.</p><p><strong>3. Why Gemini 3 for medical imaging?</strong></p><p>This was probably my biggest gamble. Gemini 3 isn’t specifically trained on medical images, but it’s gotten incredibly good at vision tasks. I considered fine-tuning a specialized model (too expensive/time-consuming) or using a dedicated medical imaging AI (most are proprietary and $$). Turns out, with careful prompt engineering, Gemini 3 can actually identify radiomic features pretty well.</p><p><strong>Learning from Syngenix:</strong> With Syngenix, I trained custom Random Forest and SVM models on TCGA gene expression data. That required extensive feature engineering and model tuning. With PanEcho, Gemini’s zero-shot learning meant I could skip all that. Trade-off: less control over the model, but way faster to build.</p><h4>The Data Model</h4><p>I kept it simple with three main tables:</p><ul><li><strong>Scans Table:</strong> User info, file location, clinical context (age, symptoms, family history), processing status, 90-day retention</li><li><strong>Analyses Table:</strong> Which scan, classification (normal vs. suspicious), risk score (0–100%), AI’s reasoning, previous scans for comparison</li><li><strong>Usage Tracking:</strong> Scans per user, rate limiting (5 scans per month to control costs)</li></ul><p><strong>PancrionDX integration note:</strong> When I eventually merge this with Syngenix, I’ll add genomic_analyses and combined_risk_assessments tables that integrate imaging + genomic + clinical data.</p><h3>How I Got Gemini 3 to Analyze Medical Images</h3><p>This was the hardest part. You can’t just throw CT scans at Gemini 3 and hope for the best.</p><p><strong>The Challenge:</strong> CT scans have dozens or hundreds of slices. I needed to extract the RIGHT slices (where the pancreas actually is), send only a few to Gemini 3 (to keep costs down), and get consistent, structured responses.</p><p><strong>My Solution:</strong></p><p><strong>Step 1: Extract key slices</strong></p><ul><li>Pancreas is usually in the middle 30-40% of an abdominal CT scan</li><li>Extract 5-8 evenly spaced slices from that region</li><li>Convert them to PNG (Gemini 3 understands images better than raw DICOM)</li></ul><p><strong>Step 2: Structured prompting</strong></p><pre>const systemPrompt = `<br>You are an expert radiologist AI specializing in pancreatic imaging.<br>Your role is to screen (not diagnose) for suspicious findings.<br>Always use plain language suitable for patients AND clinicians.<br>Never make definitive diagnoses-recommend physician consultation.<br>`;<br><br>const userPrompt = `<br>Analyze this pancreatic CT scan for PDAC (pancreatic cancer).<br>Patient info:<br>- Age: ${age}<br>- Symptoms: ${symptoms}<br>- Family history: ${familyHistory}<br><br>Look for these specific features:<br>- Mass lesions (especially &gt;2cm)<br>- Pancreatic duct dilation (&gt;3mm is concerning)<br>- Tissue atrophy or calcifications<br>- Vascular involvement<br>- Enlarged lymph nodes<br><br>Return JSON with:<br>{<br>  &quot;classification&quot;: &quot;normal&quot; or &quot;suspicious&quot;,<br>  &quot;risk_score&quot;: 0-100,<br>  &quot;summary&quot;: &quot;2-3 sentences in plain language&quot;,<br>  &quot;detailed_findings&quot;: {...}<br>}<br>`;</pre><p><strong>Comparison to Syngenix’s approach:</strong> With Syngenix, I had to manually engineer features and train models on labeled data. With PanEcho, Gemini 3 handles feature extraction automatically. Way simpler, but also less transparent. I can’t inspect model weights or feature importance like I could with Syngenix’s Random Forest.</p><p><strong>Step 3: Validation</strong></p><p>I don’t trust AI responses blindly. I validate that the risk score is 0–100, classification matches risk score (❤0% must be “normal”), all required fields are present, and response is valid JSON. If validation fails, I return a partial result with a big disclaimer.</p><h3>Building the UI for Anxious Users</h3><p>When people are worried about cancer, the LAST thing they need is a confusing interface.</p><p><strong>Design principles:</strong></p><ul><li><strong>Simplicity:</strong> Clean, lots of white space, one clear action per page</li><li><strong>Progressive disclosure:</strong> Show the most important info first (Normal vs. Suspicious, risk score), hide technical details in expandable sections</li><li><strong>Empathetic language:</strong> “We’re analyzing your scan…” not “PROCESSING… 47%”</li><li><strong>Prominent disclaimers:</strong> Every results page says “This is a screening tool, not a diagnosis”</li></ul><h3>The Hard Problems I Had to Solve</h3><h4>Problem 1: Keeping API Costs Under Control</h4><p><strong>The Issue:</strong> Gemini 3 vision is expensive. Sending 50 CT scan slices could cost $5+ per analysis.</p><p><strong>My Solution:</strong> Only send 5–8 key slices, compress images to 2048px, for longitudinal analysis send only NEW scan images, store previous results as TEXT.</p><p><strong>Result:</strong> Cost dropped from ~$5 per analysis to ~$0.35. That’s the difference between sustainable and bankruptcy.</p><p><strong>Syngenix comparison:</strong> With Syngenix, after the initial model training (one-time cost), inference was basically free. PanEcho has ongoing API costs for every analysis. This is a key difference between traditional ML and LLM-based approaches.</p><h4>Problem 2: Making Longitudinal Analysis Work</h4><p><strong>The Issue:</strong> For tracking tumor growth, I need to compare multiple scans over time with more complex AI reasoning and higher costs.</p><p><strong>My Solution:</strong> A 4-step chain:</p><ol><li>Analyze the new scan independently</li><li>Compare with previous scans (look for size changes)</li><li>Calculate risk trajectory (increasing/decreasing/stable)</li><li>Synthesize everything into a comprehensive report</li></ol><p>I only send images in step 1. Steps 2–4 work with text summaries.</p><p><strong>Result:</strong> Longitudinal analysis costs ~$0.90 (vs $3+ if I sent all images multiple times).</p><p><strong>PancrionDX integration:</strong> This same multi-step approach will work for combining imaging + genomic data.</p><h4>Problem 3: Rate Limiting Without Annoying Users</h4><p><strong>My Solution:</strong> 5 scans per user per month, clear messaging (“You have 3 scans remaining this month”), shows reset date, option to contact support for exceptions. It’s transparent and fair.</p><h4>Problem 4: Data Retention and Privacy</h4><p><strong>My Solution:</strong> 90-day automatic deletion, users notified 10 days before, can extend up to 3 times (total 360 days), can delete anytime manually, Row Level Security ensures users only see their own data.</p><p><strong>PancrionDX consideration:</strong> When integrating genomic data (way more sensitive), I’ll need encrypted at rest storage, audit logs, and option for on-premise deployment.</p><h4>Problem 5: Making It Accessible</h4><p><strong>What I did:</strong> Keyboard navigation, screen reader support, color contrast (4.5:1 minimum), never rely on color alone, focus indicators, respect reduced motion preferences. I ran automated accessibility audits (axe-core) and manual testing with screen readers.Development Process</p><h3>Breaking It Into Chunks</h3><p>I broke the project into 18 testable chunks. Each chunk had clear acceptance criteria, automated tests, and a manual testing checklist. I didn’t move to the next chunk until the current one was fully working.</p><p><strong>Why this worked:</strong> Small wins kept me motivated, testing caught bugs early, I could show progress, and easy to debug (only one new thing could be broken at a time).</p><p><strong>Lesson from Syngenix:</strong> I didn’t chunk Syngenix as carefully, which led to weeks stuck debugging feature engineering and model training simultaneously. The incremental approach works way better.</p><h3>Testing: Automated + Manual</h3><p><strong>Unit tests (Vitest):</strong></p><pre>it(&#39;rejects files over 100MB&#39;, () =&gt; {<br>  const largeFile = new File([&#39;x&#39;.repeat(150 * 1024 * 1024)], &#39;huge.dcm&#39;);<br>  expect(validateFileSize(largeFile, 100)).toBe(false);<br>});</pre><p><strong>Integration tests (Playwright):</strong> Full upload flow tested on actual TCIA scans (public pancreatic CT dataset)</p><p><strong>Manual testing:</strong> Accessibility (keyboard-only navigation, screen reader), mobile responsiveness, real-world edge cases</p><h3>What I Learned</h3><h4>Things That Worked Better Than Expected</h4><p><strong>1. Server Actions are actually great: </strong>No CORS headaches, API keys stay server-side automatically, less boilerplate, co-located with UI code.</p><p><strong>2. Supabase is a cheat code: </strong>Setting up auth, database, storage, and RLS would’ve taken me a week. With Supabase? An afternoon.</p><p><strong>3. Gemini 3 is better at medical imaging than I expected: </strong>With careful prompting, Gemini 3 can identify mass lesions, measure duct dilation, spot vascular involvement, and explain findings in plain language.</p><p><strong>Syngenix lesson applied:</strong> When I built Syngenix, I spent weeks training custom models from scratch. PanEcho taught me that sometimes off-the-shelf foundation models are good enough, you just need to learn how to use them effectively.</p><p><strong>4. Tailwind + shadcn/ui = speed: </strong>Let me build a professional-looking UI in days, not weeks.</p><h4>Things That Were Harder Than Expected</h4><p><strong>1. DICOM files are a nightmare: </strong>Medical imaging standards are complex. DICOM files can be non-standard, corrupted, huge (500MB+), or multi-series. I spent days debugging weird edge cases.</p><p><strong>2. Prompt engineering is an art, not a science: </strong>Getting Gemini 3 to return consistent results took 15+ iterations. It’s not hard, but it requires patience.</p><p><strong>Syngenix comparison:</strong> With Syngenix, I could debug model failures by inspecting feature importance. With Gemini 3, it’s a black box. I can only tweak the prompt and hope for better results. Less control, but faster to iterate.</p><p><strong>3. Token costs add up FAST: </strong>My first implementation cost $4.80 per analysis. After optimization: $0.35. 14x cheaper.</p><p><strong>4. Medical accuracy is HARD to validate:</strong> How do I know if Gemini 3’s analysis is correct? I’m not a radiologist. I tested with TCIA scans, but that’s not enough for real-world use.</p><p><strong>Syngenix had the same problem:</strong> Even though I validated against TCGA ground truth labels, I couldn’t be 100% sure my gene expression classifier was clinically accurate without real-world testing.</p><p><strong>PancrionDX solution:</strong> By combining imaging (PanEcho) + genomics (Syngenix) + clinical factors, we can cross-validate. If all three agree on high risk, that’s more reliable than any single modality alone.</p><h4>Mistakes I Made</h4><p><strong>1. Underestimating token costs: </strong>My initial design would’ve cost $5+ per analysis. Should’ve calculated costs upfront.</p><p><strong>2. Not planning for edge cases sooner:</strong> “What if the DICOM file is corrupted?” I encountered all of these and more. Building error handling after the fact is way harder.</p><p><strong>3. Overcomplicating the initial design: </strong>My first architecture diagram had microservices, message queues, and a separate Python backend. Way overkill. Keep it simple, especially for v1.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*a8mgFq25N0vlMyg7lEAf9g.png" /><figcaption>Landing Page for PanEcho</figcaption></figure><h3>The Results (So Far)</h3><h4>Performance Metrics</h4><p>After testing with 10 TCIA scans:</p><ul><li>Processing time: 35–45 seconds on average (target was &lt;60s) ✅</li><li>API cost: $0.35 per initial scan, $0.90 per longitudinal (within budget) ✅</li><li>Upload success rate: 97% (3% failed due to corrupted DICOM files)</li><li>Classification accuracy: Hard to measure without radiologist review, but results seem reasonable</li></ul><h4>What’s Working</h4><p><strong>The good:</strong></p><ul><li>File upload + DICOM processing is solid</li><li>Gemini 3 analysis returns structured, useful results</li><li>UI is clean and accessible</li><li>Longitudinal tracking shows meaningful trends</li><li>Deployed on Vercel free tier without issues</li></ul><p><strong>The needs-work:</strong></p><ul><li>Need medical validation (radiologist review)</li><li>False positive rate unknown</li><li>Should add explanations (why did the AI flag this?)</li><li>Mobile experience could be smoother</li></ul><h3>Cost Analysis</h3><p>For a typical user (uploads initial scan + 2 follow-ups per month):</p><ul><li>Google AI API: ~$1.50/month</li><li>Supabase: $0 (free tier)</li><li>Vercel: $0 (free tier)</li></ul><p><strong>Total cost per user:</strong> ~$1.50/month</p><p>At 100 users: $150/month (manageable)<br>At 1,000 users: $1,500/month (need to monetize or get funding)</p><p><strong>PancrionDX cost projection:</strong> When combining both tools:- imaging analysis (~$1.50/month), genomic analysis (free after model training), clinical risk assessment (minimal compute cost). <strong>Total:</strong> ~$2–3/month per user for comprehensive screening. Way cheaper than traditional genomic testing ($200–1,000).</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*U1aaoN1BTXw80CPu7lQ9Ag.png" /><figcaption>Scan results page from PanEcho</figcaption></figure><h3>What’s Next</h3><h4>Immediate Priorities</h4><p><strong>1. Medical validation: </strong>Get a radiologist to review a sample of analyses. Questions: Are the identified features actually concerning? Are we missing obvious findings? What’s the false positive/negative rate? (Same priority for Syngenix: Need oncologist validation)</p><p><strong>2. Explainability: </strong>Show WHICH parts of the image influenced the AI’s decision. Options: attention maps, saliency visualization, side-by-side comparisons with annotations.</p><p><strong>3. Better error handling: </strong>Current error messages are okay, but could be more helpful with troubleshooting steps.</p><h4>Building PancrionDX</h4><p><strong>The integration plan:</strong></p><p><strong>Phase 1: Data Pipeline</strong></p><ul><li>Build unified upload system (CT scans, genomic files, clinical forms)</li><li>Standardize data formats across modalities</li><li>Create combined database schema</li></ul><p><strong>Phase 2: Multi-Modal Risk Assessment</strong></p><ul><li>Train fusion model combining PanEcho imaging risk score, Syngenix genomic risk score, and clinical risk factors</li><li>Weighted ensemble approach (imaging 40%, genomic 30%, clinical 30%)</li><li>Output: single comprehensive risk score with confidence intervals</li></ul><p><strong>Phase 3: Clinical Validation</strong></p><ul><li>Partner with hospital for prospective study</li><li>500+ patients with multiple data modalities</li><li>Compare PancrionDX vs. standard of care</li><li>Measure: sensitivity, specificity, early detection rate</li></ul><p><strong>Technical challenges:</strong> Handling missing data, calibrating risk scores across modalities, explaining combined risk, HIPAA compliance.</p><h4>Potential Vision</h4><p><strong>1. FDA clearance pathway:</strong> If validation results are good, pursue FDA 510(k) clearance as a Computer-Aided Detection device. This is expensive ($100k+), time-consuming (1–2 years), and requires clinical data, but necessary for serious medical use.</p><p><strong>2. Real deployment:</strong> Get this into actual clinics where it can help real people. That’s the whole point.</p><h3>Reflections: What This Project Taught Me</h3><h4>Technical Skills</h4><p>This project leveled up my skills in Next.js 14, prompt engineering, medical imaging (DICOM format, CT scan anatomy), database design, accessibility, and cost optimization.</p><h4>Non-Technical Lessons</h4><p><strong>1. Build for real users, not imaginary ones:</strong> I initially designed complex features that “users might want.” Then I talked to people actually at risk for pancreatic cancer. They wanted simple, clear results, fast processing, and ability to share with their doctor. Always validate with real users.</p><p><strong>2. Constraints breed creativity: </strong>The $0.35 cost target forced me to get creative. If I had unlimited budget, I probably would’ve built something worse.</p><p><strong>3. Incremental progress &gt; perfectionism:</strong> I used to spend weeks planning before writing code. The 18-chunk approach taught me to ship small, working pieces, get feedback early, and iterate based on reality, not theory.</p><p><strong>4. AI is a tool, not magic: </strong>GPT-4o is impressive, but it’s not infallible. It hallucinates findings sometimes, misses subtle abnormalities, and can’t replace human judgment. The right mental model: AI is an assistant, not a replacement.</p><p><strong>5. Medical software is different:</strong> Building health tech comes with responsibilities. People’s lives could depend on this. False negatives could delay life-saving treatment. False positives cause unnecessary anxiety. Legal liability is real. I added disclaimers, but I also lose sleep thinking: “What if someone trusts this too much and doesn’t see a doctor?” This isn’t just a coding exercise. It matters.</p><h3>Final Thoughts</h3><p>Building PanEcho was one of the most challenging and rewarding projects I’ve done. It’s not just about code, it’s about potentially helping people catch cancer early enough to actually treat it.</p><p>The honest truth:</p><ul><li>This tool isn’t ready for serious medical use (needs validation)</li><li>Gemini 3 is good but not perfect</li><li>I’m not a doctor, and this isn’t medical advice</li><li>There are serious ethical and legal questions to figure out</li></ul><p>But also:</p><ul><li>AI CAN analyze medical images reasonably well</li><li>The tech stack (Next.js + Supabase + Gemini 3) works great</li><li>There’s real potential here to save lives</li><li>Someone with more resources should pursue this seriously</li></ul><p>If I had to start over, I’d:</p><ul><li>Start with a smaller scope (maybe just initial analysis, no longitudinal)</li><li>Focus on one cancer type and do it really well</li><li>Partner with a hospital for validation data</li></ul><p>But I’m proud of what I built. It works. It’s fast. It’s accessible. And maybe, just maybe, it’ll inspire someone to take this idea and run with it properly.</p><h3>Try It / Learn More</h3><p><strong>Live demo:</strong> <a href="https://panecho-pancrion-dx.vercel.app/">https://panecho-pancrion-dx.vercel.app/</a><br><strong>Code:</strong> <a href="https://github.com/itswkosi/panecho">https://github.com/itswkosi/panecho</a><br><strong>Contact:</strong> <a href="https://www.linkedin.com/in/semilogo-oketola/">LinkedIn</a> or <a href="http://semilogooketola@outlook.com/">Email</a> if you want to chat about this project</p><p><strong>For medical professionals:</strong> If you’re a radiologist or oncologist interested in validating this approach, I’d love to talk. Seriously, please reach out.</p><p><strong>For developers:</strong> All the code, prompts, and architecture decisions are documented in the repo. Feel free to fork it, improve it, or use it as a starting point for your own medical AI projects.</p><p><strong>For patients/concerned individuals:</strong> Please don’t use this for actual medical decisions yet. It’s not validated. Talk to your doctor about pancreatic cancer screening options (endoscopic ultrasound, MRI/MRCP, etc.).</p><h3>Acknowledgments</h3><p>Big thanks to:</p><ul><li>The Cancer Imaging Archive (TCIA) for public CT scan datasets</li><li>Google for Gemini 3 API access</li><li>Vercel and Supabase for generous free tiers</li><li>Everyone who gave feedback on early versions</li></ul><h3>A final note on responsibility:</h3><p>I built this because I think AI can help save lives. But I also know that releasing half-baked medical software can cause harm. If you’re thinking of building something similar:</p><ul><li>Talk to medical experts FIRST</li><li>Plan for validation from day one</li><li>Be transparent about limitations</li><li>Never oversell the accuracy</li><li>Think hard about the ethics</li></ul><p>We’re playing with people’s health here. Take it seriously. Alright, rant over. Go build cool (and responsible) stuff.</p><p>If you found this interesting, take a look at my <a href="https://www.linkedin.com/in/semilogo-oketola/">LinkedIn</a>, and <a href="https://semilogooketola.lovable.app/">personal website</a></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=396e43a415e8" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[SynGenix: Bringing RNA-Seq Classifiers to Pancreatic Cancer Risk Prediction]]></title>
            <link>https://semilogooketola.medium.com/syngenix-bringing-rna-seq-classifiers-to-pancreatic-cancer-risk-prediction-0f4edf1a8d17?source=rss-8a02bfe1359e------2</link>
            <guid isPermaLink="false">https://medium.com/p/0f4edf1a8d17</guid>
            <category><![CDATA[pancreatic-cancer]]></category>
            <category><![CDATA[rna-seq]]></category>
            <category><![CDATA[ai]]></category>
            <dc:creator><![CDATA[Semilogo Oketola]]></dc:creator>
            <pubDate>Sat, 22 Nov 2025 14:50:47 GMT</pubDate>
            <atom:updated>2025-12-01T13:21:41.493Z</atom:updated>
            <content:encoded><![CDATA[<p>Pancreatic cancer models usually lean on imaging, biomarkers, or vague risk calculators. I wanted to answer a different question: <em>Can raw RNA-seq data alone reveal early signals of high-risk biology?</em><br> That question became SynGenix.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*T_RO5PE3f1O0-iyLHU0U1A.png" /></figure><p>Instead of asking clinicians to run multi-step pipelines, SynGenix takes the raw counts directly and turns them into a genomic risk score, one that compares your expression profile to patterns seen in known high-risk and low-risk cohorts. The idea started small: rebuild a miniature version of the PanCanAtlas RNA-seq workflow to understand the logic behind each preprocessing decision. But the deeper I went, the less it felt like a toy and the more it became a way to test how far careful QC, normalization, and feature selection can take a classifier built on expression data alone.</p><p>I built SynGenix as the foundation for PancrionDx, my long-term vision for a comprehensive early detection model for pancreatic cancer. But what seemed like a modest prototype turned into something deeper: a full exploration of RNA sequencing technology, preprocessing pipelines inspired by PanCanAtlas, and the challenge of translating complex genomic data into interpretable risk scores.</p><h3>Inspiration</h3><p>My fascination with pancreatic cancer genomics began when I explored the PanCanAtlas studies. Their RNA-seq pipelines are clinical-grade: strict QC, normalization, batch-correction, robust feature selection. I thought, <em>If I replicate their approach in miniature, I can start to understand what really matters.</em></p><p>At first, I treated their published methods like a recipe: download counts, normalize, run a classifier. But replication without comprehension only gets you so far. I started asking: <em>Why did they filter genes this way? Why this normalization algorithm instead of another? What’s the impact of including or excluding low-count genes?</em> These weren’t academic questions; they were design decisions that shape every downstream model. That curiosity changed everything. SynGenix wasn’t just a toy: it became a sandbox to test, fail, and rebuild.</p><h3>Learning RNA-Seq</h3><h4>Choosing the Right Data</h4><p>My first error was picking a microarray dataset because it was easy to access. It felt safe. But microarray data measures expression differently and comes with its own set of biases. When I compared microarray to RNA-seq, I realized they’re not interchangeable, at least not for what I wanted to build. The disparity in technology could introduce misleading signals. So I pivoted to RNA-seq. That shift forced me to answer deeper questions:</p><ul><li>What do raw read counts actually represent?</li><li>How does sequencing depth bias affect interpretation?</li><li>Why should I filter out genes with low expression or low variance?</li></ul><p>Diving into these questions was humbling. I spent nights reading about how sequencers convert RNA molecules into digital reads, why library preparation matters, and how GC content can skew results. StatQuest videos helped me build intuition, but I quickly supplemented them with research papers and documentation from tools like DESeq2 and edgeR.</p><h4>Building a Preprocessing Pipeline</h4><p>Once I had raw counts, I built a pipeline with multiple stages:</p><ol><li><strong>Quality Control (QC):</strong> I wrote scripts to detect outlier samples, mismatched metadata, and samples with abnormal library sizes.</li><li><strong>Filtering:</strong> I dropped genes with consistently low counts and no real variation across samples.</li><li><strong>Normalization:</strong> I experimented with different strategies (e.g., TPM, CPM, and more sophisticated methods) until I found the balance between signal preservation and noise control.</li><li><strong>Feature Engineering:</strong> I focused on genes with the most variance, plus known biomarkers drawn from literature.</li></ol><p>This process taught me one of the most important lessons: <em>most of the model’s power comes from getting the input right</em>. A great classifier can’t fix a poorly normalized dataset.</p><h3>Modeling</h3><p>With clean, normalized data in hand, I turned to model training.</p><ul><li>I tried logistic regression first because I heard it was a simple, interpretable, a good baseline.</li><li>Then I explored Random Forests<strong> </strong>and XGBoost, which gave better performance but required more careful tuning.</li><li>For validation, I used stratified train/validation splits, ensuring the class balance held even when I shuffled seeds.</li></ul><p>The goal wasn’t just “get a high AUC.” It was to create a stable model,<strong> </strong>one that wouldn’t fall apart when I tweaked normalization or feature selection. I realized early on that swings in performance often came from preprocessing changes, not the algorithm itself.</p><p>Once trained, the model outputs a probability. But I didn’t want to stop there. I built a risk scoring layer that translates that probability into something more interpretable:</p><ul><li>I overlaid the score on the distribution of my training cohort</li><li>I computed confidence intervals</li><li>I made it clear where the classifier was more or less certain</li></ul><p>This risk score isn’t a diagnosis. It’s a genomic fingerprint. It’s a way of saying, “your expression pattern is more similar to high-risk samples than to low-risk ones,” with caveats attached.</p><h3>Designing for Real-World Use</h3><p>Models don’t live in isolation, people use them. I sketched out a web interface in Figma Make<strong> </strong>so I could imagine how a researcher or clinician might upload data and get results. Through design, I uncovered some interesting questions:</p><ul><li><strong>Input format: </strong>What file types should the user upload? Raw counts? Normalized matrices?</li><li><strong>Transparency:</strong> Do I show feature importance per gene, or just a high-level risk score?</li><li><strong>Interpretation:</strong> Do I present confidence intervals? Cohort distribution? What if a user misinterprets a probability?</li><li><strong>Trust:</strong> How do I prevent someone from over-relying on a risk number that wasn’t validated prospectively?</li></ul><p>Design forced clarity. The interface is about how much interpretation I can and should hand to someone else, not just aesthetic quality. Because if a researcher can’t interpret results correctly, the classifier is useless, or worse, dangerous.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/932/1*_GfGsKzAXpCcng6kvTdd4Q.png" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*TI5Rklpymc0nOxQCEAWUbQ.png" /><figcaption>UI design for SynGenix featuring manual or file upload</figcaption></figure><iframe src="https://cdn.embedly.com/widgets/media.html?src=https%3A%2F%2Fwww.loom.com%2Fembed%2Fd16bbab0d5de4f6e986f201b95c612c6&amp;display_name=Loom&amp;url=https%3A%2F%2Fwww.loom.com%2Fshare%2Fd16bbab0d5de4f6e986f201b95c612c6&amp;image=https%3A%2F%2Fcdn.loom.com%2Fsessions%2Fthumbnails%2Fd16bbab0d5de4f6e986f201b95c612c6-0776c72f68cfe750.gif&amp;type=text%2Fhtml&amp;schema=loom" width="1920" height="1440" frameborder="0" scrolling="no"><a href="https://medium.com/media/f811e128675b60a78b1065c7e0d69c15/href">https://medium.com/media/f811e128675b60a78b1065c7e0d69c15/href</a></iframe><h3>What I Learned (And What Comes Next)</h3><p>Working on SynGenix taught me far more than I expected:</p><ol><li><strong>Sequencing knowledge matters:</strong> I now understand how RNA-seq works, where biases come from, and why not all datasets are created equal.</li><li><strong>Preprocessing is king:</strong> The model’s success depended more on normalization, filtering, and QC than on choosing a fancy algorithm.</li><li><strong>Validation discipline is non-negotiable:</strong> Stability across preprocessing is more valuable than a single “good” test accuracy.</li><li><strong>Interpretability is essential:</strong> Without clear communication of uncertainty and context, risk scores become black boxes.</li><li><strong>Designing for users reveals blind spots:</strong> Sketching the UI in Figma showed me weaknesses in my pipeline that I wouldn’t have seen otherwise.</li></ol><h3>What I’d Do Differently</h3><p>Now that SynGenix actually works end-to-end, it’s obvious which parts of my process were strong and which parts were… let’s say unnecessarily experimental. If I were starting this project again, these are the changes that would have saved me weeks of circling and rebuilding.</p><h4>1. Validate My Data Choice Earlier</h4><p>Switching from microarray to RNA-seq midway was very inconvenient. I had to redo preprocessing, feature engineering, and entire chunks of the model.<br> Looking back, I should have questioned my dataset <em>before</em> I touched a notebook. The technology behind the data should drive the pipeline, not the other way around.</p><ul><li>What biological question am I answering?</li><li>Which technology captures that signal best?</li><li>What biases does each platform introduce?</li></ul><p>If I had forced myself to justify the sequencing method at the very beginning, I would have avoided half the detours I ended up taking.</p><h4>2. Spend More Time on Metadata QC</h4><p>My earliest runs had inconsistencies because I didn’t scrutinize the metadata early enough. RNA-seq samples with mismatched labels, missing tumor stage fields, or unclear sample origins can break your entire classifier silently. I learned the hard way that messy metadata is a liability.</p><p>If I redid SynGenix, the <em>first thing</em> I’d build is a metadata auditor, something that checks class balance, missing fields, sample duplication, outliers in sequencing depth and inconsistent annotations</p><p>before I ever touch normalization.</p><h4>3. Prototype Multiple Preprocessing Pipelines Instead of Committing Too Early</h4><p>I underestimated how much normalization and filtering shape the downstream model. I picked one pipeline too early and only compared other approaches later, when re-running everything felt like a chore. A more disciplined approach would have been to:</p><ul><li>test multiple normalization strategies (CPM, TPM, variance stabilizing, etc.)</li><li>compare variance thresholds for feature filtering</li><li>evaluate the impact of removing highly correlated features</li><li>run a grid of preprocessing combinations</li></ul><p>before picking the final pipeline. This would have given me much more confidence in the stability of my results.</p><h4>4. Document Every Choice As I Made It</h4><p>Half of my confusion came from not documenting why I made certain calls. When you revisit code weeks later, you forget the reasoning behind certain filters or thresholds. Good scientific modelling requires a log of decisions and the justification behind them.</p><p>Next time, I’ll maintain an internal “lab notebook” not for code, but for reasoning, pipeline alternatives I rejected, and why I rejected them.</p><h4>5. Build the UI Earlier</h4><p>I thought UI came at the end. Turns out, designing the interface earlier would have clarified important questions:</p><ul><li>Which inputs can users realistically upload?</li><li>How much interpretation support do I need to provide?</li><li>What metadata should I require at upload time?</li></ul><p>By the time I finally drafted the UI, I realized my original output structure made no sense for a user. Next time, I’d sketch the interface at the start before writing a line of code, because UI forces you to articulate what the model <em>actually</em> does.</p><h4>6. Add More Stress Tests to Avoid Overfitting Subtly</h4><p>I validated SynGenix properly, but I was still too trusting of decent metrics without pushing harder on edge cases. If I were to restart, I would:</p><ul><li>test smaller subsets of data</li><li>run permutations to ensure the signal isn’t random</li><li>test different train/validation splits</li><li>compare performance after shuffling metadata</li><li>to check how robust the classifier really is.</li></ul><p>A good model should survive being poked hard; mine survived, but I should have poked harder.</p><h4>7. Start With a Scoping Document</h4><p>SynGenix grew organically, which is a nice way of saying it grew without a solid plan. A simple one-page scoping document would have helped define my goals for this project, list constraints, choose proper data source, outline preprocessing requirements and define “done”</p><p>Without that, I occasionally drifted into unnecessary side quests (like trying five feature selection strategies back-to-back). Next time, I’ll scope first, build second, and refine third.</p><h3>Vision: SynGenix → PancrionDx</h3><p>SynGenix might seem like a small side project, but to me, it was foundational. It was a sandbox that allowed me to test every design, statistical, and engineering assumption that will go into PancrionDx, my ultimate goal: a multimodal early-detection system for pancreatic cancer.</p><p>Over time, I hope to:</p><ul><li>Validate SynGenix on independent cohorts</li><li>Incorporate more complex data (for example, clinical metadata, mutational profiles)</li><li>Create a robust, user-facing tool that researchers can actually use for hypothesis generation or risk stratification</li><li>Collaborate with labs to test it prospectively</li></ul><h3>Final Thoughts</h3><p>Building SynGenix was humbling. It forced me to confront how much I didn’t know; not just about modelling, but about the raw biological processes behind my data. It showed me that ‘simple’ tools demand serious discipline, and that interpretability isn’t a feature you tack on at the end. It’s integral from day one.</p><p>SynGenix taught me more about modelling than any textbook ever could, and it exposed every weak point in my assumptions. It showed me that good performance is meaningless without a defensible pipeline behind it, that uncertainty needs to be surfaced instead of hidden, and that user-facing tools demand more discipline than academic notebooks.<br> The next phase is simple: validate it on outside cohorts, challenge it with more diverse data, and push the interface toward something researchers can actually rely on. SynGenix isn’t the final version of anything; it’s the baseline I’ll measure every future iteration against as I build toward the full vision of PancrionDx.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=0f4edf1a8d17" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Investigating Pancreatic Ductal Adenocarcinoma]]></title>
            <link>https://semilogooketola.medium.com/investigating-pancreatic-ductal-adenocarcinoma-50a6a3b94a0f?source=rss-8a02bfe1359e------2</link>
            <guid isPermaLink="false">https://medium.com/p/50a6a3b94a0f</guid>
            <category><![CDATA[pancreatic-cancer]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[oncology]]></category>
            <dc:creator><![CDATA[Semilogo Oketola]]></dc:creator>
            <pubDate>Mon, 27 Oct 2025 04:31:14 GMT</pubDate>
            <atom:updated>2025-10-27T04:31:14.612Z</atom:updated>
            <content:encoded><![CDATA[<h4>Table of Contents:</h4><p>∘ <a href="#4bd4">Introduction</a><br> ∘ <a href="#350f">How Pancreatic Cancer Develops Biologically:</a><br> ∘ <a href="#642e">Detection Methods: Traditional vs. Modern</a><br> ∘ <a href="#7e8f">Innovators in the Field</a><br> ∘ <a href="#f7fb">Challenges and Future Directions</a><br> ∘ <a href="#e935">Conclusion</a><br> ∘ <a href="#60c4">References:</a></p><h4><strong>Introduction</strong></h4><p>The pancreas is a very important gastrointestinal organ with two significant roles. One, it produces digestive enzymes, such as amylase, lipase, trypsin, and chymotrypsin, to allow the breakdown of carbohydrates, fats, and proteins in the small intestine. Two, it produces hormones regulating blood sugar levels, such as insulin, which reduces blood sugar, and glucagon, which elevates blood sugar. It also produces other hormones, such as somatostatin, which regulates digestive enzyme release and maintains homeostasis in the system by inhibiting hormone and enzyme secretion.</p><p>When pancreatic cancer occurs, both digestive and hormonal operations of the pancreas are disrupted, impairing nutrient absorption and blood sugar management. What makes this cancer so dangerous is its tendency to remain asymptomatic in its early stages; meaning it is often detected only after it has spread to nearby organs. Pancreatic cancer remains one of the deadliest cancers, with a mere 9–13% chance of survival over five years. To understand why pancreatic cancer is so difficult to treat, one has to look at how it develops at the cellular and genetic level.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*yYkV75PGNsxG2MvYDh3Rag.png" /><figcaption>Fig 1: Comparison of Regular Pancreas and Pancreas affected with PDAC</figcaption></figure><h4><strong>How Pancreatic Cancer Develops:</strong></h4><p>There are two general categories of pancreatic cancer: exocrine and neuroendocrine, with the exocrine cancers representing about 95% of the pancreatic cancers. Exocrine pancreatic cancer begins in the pancreatic cells that secrete digestive enzymes. Pancreatic ductal adenocarcinoma (PDAC) is the most frequent form of exocrine pancreatic cancer and occurs in the epithelial cells lining the ducts through which these enzymes flow. PDAC develops when mutated glandular or ductal cells acquire alterations that cause them to grow in an uncontrollable way and proliferate.</p><p>Less common are pancreatic neuroendocrine cancers, which begin at the neuroendocrine cells of the pancreas. These cells are specialized cells that combine the features of both neurons and hormone producing cells called endocrine cells. They receive signals from the nervous system and release hormones that regulate blood sugar levels, help control metabolism and aid in digestion by influencing the release of other hormones and digestive enzymes. Sometimes, these cells mutate and begin to multiply uncontrollably, forming neuroendocrine tumors (NETs). These tumors can be classified as functional, where they produce hormones and cause hormone related symptoms; and non-functional, where no hormones are produced. They begin at the islet cells that produce hormones and are often slower progressing than exocrine tumors. Diagnosis and treatment of NETs are dependent on a range of factors including tumor type, site, hormone production, aggressiveness, and spread. Neuroendocrine tumors generally have a better prognosis than exocrine cancers.</p><p>This review focuses on exocrine pancreatic cancers, more specifically the pancreatic ductal adenocarcinoma. Like neuroendocrine cancer, adenocarcinomas can also be found elsewhere, as they develop from glandular epithelial cells. PDAC is the outcome of a combination of mutations in key driver genes, most commonly in the oncogene KRAS and the tumor suppressor genes CDKN2A, TP53 and SMAD4.</p><p>KRAS is mutated early in PDAC development and is the main driving force behind the disease. Hyperactivated KRAS initiates critical signaling pathways such as MAPK and PI3K that promote overgrowth, survival, and metabolic reprogramming of cells to make the cancer aggressive and resistant to treatment. CDKN2A is a tumor suppressor gene that encodes two crucial proteins: p16INK4A, which binds to and blocks the activity of CDK4 and CDK6 proteins, which are needed for the cell cycle to progress, preventing abnormal cell divisions; and p14ARF, which protects the p53 tumor suppressor protein responsible for regulating cell division from being broken down. Mutations in this gene can cause these proteins to stop working, allowing cells to grow and divide too quickly and out of control. TP53 and SMAD4 are additional tumor suppressor genes. Their loss tends to occur later in disease evolution, inducing progression from localized precancerous lesions to invasive carcinoma.</p><p>There are also some inherited gene mutations linked to pancreatic cancer like mutations in the BRCA1 and BRCA2 genes, which are essential in repairing damaged DNA, preventing the accumulation of errors in the genetic code. When the BRCA1/2 gene is mutated, the protein cannot perform this function correctly, leading to an accumulation of genetic damage that triggers uncontrolled cell growth and tumor formation.</p><p>PDAC is highly metastatic and has the tendency to spread to nearby or even distant organs, including the liver, lymph nodes, peritoneum (the lining of the abdominal cavity), lungs, bones, and gastrointestinal tract. The liver is the most common site of metastasis, especially for carcinomas occurring in the body or tail of the pancreas. Metastatic cells can obstruct bile ducts of the liver, producing jaundice, hyperbilirubinemia and elevation of liver enzyme levels. Lymph node spread, usually peripancreatic nodes, occurs either by direct invasion along contiguous growth (spread by touch) or by lymphatic dissemination. The peritoneum is also involved as a frequent site, with ascites (fluid in the abdomen), abdominal or back pain, bowel or urinary obstruction, and cancer-related weight loss being symptoms expressed due to metastasis. PDAC also invades major blood vessels such as the celiac axis and superior mesenteric artery, leading to an interruption of blood supply to the liver, stomach, intestines, and other organs supplied by its branches. With its complex biology and virulent spread, effective detection methods for PDAC are critical, though diagnosing PDAC has been mostly challenging.</p><figure><img alt="Fig 1: Progression of Pacreatic Ductal Adenocarcinoma Tumor Cells" src="https://cdn-images-1.medium.com/max/1024/1*H6YlOCKYqzw-IZBJoNHnAg.png" /><figcaption>Fig 2: Progression of Pancreatic Ductal Adenocarcinoma Tumour Cells</figcaption></figure><h4><strong>Detection Methods: Traditional vs. Modern</strong></h4><p>Traditional methods of diagnosing pancreatic ductal adenocarcinoma (PDAC) are relatively based on imaging, laboratory investigations, and tissue biopsy. Imaging is often the first technique that is used for diagnosis, where CT scanning is most commonly used to gauge tumor size, location, and vascular invasion. While highly sensitive for tumors larger than 2 cm in diameter, CT scans are not as likely to detect smaller tumors. Additional imaging technologies, such as endoscopic ultrasound (EUS) and positron emission tomography (PET) scans, also are frequently employed to enhance diagnostic accuracy.</p><p>Serum biomarkers, such as cancer antigen 19–9 (CA19–9), also play a significant role in diagnosis. CA19–9 testing is employed to monitor the level of this protein in the blood as a way of tracking pancreatic cancer growth and response to treatment. But it is not helpful for the earliest detection because CA19–9 levels rise only when cancer has progressed to stage II or stage III. In addition, there may be elevated levels in benign conditions such as pancreatitis or bile duct obstruction, and there are individuals who do not secrete CA19–9 altogether. In addition, the biological features of PDAC include the production of a dense desmoplastic stroma, a fibrous tissue barrier that acts as a physical and biochemical shield that conceals tumors during CT scans and prevents successful therapy. During biopsies, the needle collects mostly fibrous tissue rather than the tumor cells, leading to false negatives or inconclusive results during diagnosis.</p><p>Fortunately, there are several encouraging next-generation technologies transforming the PDAC detection and diagnosis landscape:</p><ul><li>Liquid Biopsies: These non-invasive tests quantify circulating tumor DNA (ctDNA), circulating tumor cells (CTCs), and other molecular biomarkers in blood, urine, or bodily fluids. Liquid biopsies may be able to detect cancer months or even years before symptoms develop.</li><li>Advanced Imaging Modalities: New molecular imaging and high-resolution modalities facilitate early visualization of the tumor. Molecular imaging is leveraging the application of radiotracers or targeted contrast agents to identify tumor-specific proteins and metabolite activity, such as PET tracers targeting PDAC-specific antigens. High-resolution MRI and newer EUS modalities, including diffusion-weighted MRI and contrast-enhanced EUS, increase sensitivity for occult or small lesions.</li><li>Artificial Intelligence (AI) and Machine Learning: Deep learning models are being trained to read medical images and histopathology slides for subtle abnormalities that may precede tumor growth. When combined with patient data such as age, genetics, or diabetes onset, AI systems can stratify patients by risk and potentially detect PDAC earlier than current methods.</li><li>Genetic and Risk Screening: Germline testing for BRCA1/2, PALB2, or CDKN2A mutations enables the early identification of high-risk individuals who can be safeguarded by close monitoring and preventive measures.</li><li>Biomarker Panels: Instead of relying solely on CA19–9, researchers are developing multi-analyte diagnostic panels that detect multiple biomarkers (proteins, DNA, RNA, and metabolites) simultaneously in a single sample. This approach should increase sensitivity and specificity, particularly for asymptomatic patients.</li><li>Nanotechnology and Biosensors: Wearable biosensors and new nanotechnology are being developed to detect molecular or metabolic changes in real time. Currently experimental, such devices have the potential to make continuous, non-invasive surveillance of cancer possible one day.</li></ul><p>Together, these technologies are great innovations in how pancreatic cancers are being diagnosed and treated. While these emerging technologies have much to offer for changing PDAC detection and diagnosis, it is the manner in which they are being used in real-world research and clinical settings that offers the greatest potential.</p><h4><strong>Innovators in the Field</strong></h4><p>Building upon these emerging diagnostic principles, several biotechnology companies are now translating these innovations into real-world clinical tools for the detection and diagnosis of PDAC. GRAIL’s Galleri test, Guardant Health, Freenome, and PAC-MANN are worthy of note.</p><p>GRAIL’s Galleri test is a multi-cancer early detection assay that is blood-based and can identify over 50 types of cancer, including PDAC. It translates DNA methylation signatures, fixed DNA methyl group patterns that regulate gene expression, to find signals of cancer in cell-free DNA (cfDNA). Epigenetic markers are such fixed signs of tissue- and cell-type-specific activity. Galleri, in a new clinical trial, identified cancers seven times as frequently as traditional screening tests currently identify them, with over half diagnosed at early, treatable stages. The test was discovered to have approximately 62% sensitivity for early-stage cancers, such as PDAC, and is currently underway at FDA review. Despite the fact that Galleri is for screening purposes rather than for diagnosis, it offers data on both the existence and likely tissue of origin of detected cancers.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/407/1*UxLgkkPnIb1_ZDiUdkIl2g.jpeg" /><figcaption>Fig 3: Systemic review and meta analysis diagnostic performance of DNA alterations in pancreatic juice for the detection of pancreatic cancer</figcaption></figure><p>Guardant Health developed several liquid biopsy platforms, Guardant360 and Guardant Reveal, that analyze cfDNA for genomic alterations and minimal residual disease (MRD), tiny leftover cancer after treatment. In 2024, Guardant launched Shield, an FDA-approved blood methylation test in patients 45 years and above. Shield detects several cancers, including PDAC, pre-symptomatically by analyzing cfDNA for methylation signature change in cancer-related genes. These technologies allow doctors to diagnose tumors at an earlier stage, monitor recurrence, and formulate more specific treatment regimens.</p><p>Freenome is developing a multiomics platform to detect PDAC in its earliest and most treatable state. It has designed a platform to merge genomic, transcriptomic, methylomic, and proteomic information, including DNA mutations, gene expression levels, methylation patterns, and protein biomarkers, through high-performance machine models. Combined as a whole, the integrated platform can detect weak molecular cues linked to the development of early PDAC, providing meaningful information about disease presence and possibly, disease stage, to clinicians.</p><p>Finally, PAC-MANN, or Protease-ACtivated MAgnetic NaNosensor, was developed by researchers at the OHSU Knight Cancer Institute and Cancer Early Detection Advanced Research Center (CEDAR). PAC-MANN detects elevated protease activity in the blood, a biomarker indicative of cancer development. It makes use of a fluorescent peptide attached to magnetic nanoparticles. When these peptides are broken down or cleaved by active proteases, they emit a fluorescent signal, indicating the presence of cancer related protease activity. This technique offers a sensitive, non-invasive way to discern molecular alterations in the blood related to PDAC.</p><p>These technologies combined are the future of early cancer detection, merging molecular, computational, and nanoscale approaches to catch pancreatic cancer much earlier than is currently possible. While these companies have laid the groundwork for redefining early detection, real-world application has technical and biological constraints.</p><h4><strong>Challenges and Future Directions</strong></h4><p>As with any new technology, there are various problems in the effectiveness and scalability of these strategies. For instance, liquid biopsies are not able to effectively detect early pancreatic tumors since such tumors release low levels of circulating tumor DNA (ctDNA). Additionally, tumor heterogeneity, which is the genetic variation between patients, makes universal detection models difficult to achieve; and benign diseases or other cancers can trigger spurious signals. High sequencing costs and analytical complexity also limit widespread clinical adoption.</p><p>The most difficult challenge is the absence of large, high-quality datasets of early PDAC scans to train strong models with. The “black box” nature of much of the AI is also an important issue: clinicians can see the inputs and outputs but not the decision-making process within. This uninterpretability has a tendency to raise skepticism and slow clinical adoption. In addition, AI-based diagnostic instruments must undergo strict clinical validation and approval for regulatory use prior to deployment.</p><p>Likewise, biomarker panels, nanotechnology, biosensors, multiomics liquid biopsies, and other AI-derived approaches all encounter similar obstacles. Most are still in the experimental or research phase and are being developed further to increase their sensitivity, specificity, and practical usefulness.</p><p>Nonetheless, these approaches have already made great strides in early cancer detection. Liquid biopsies are becoming increasingly sensitive, now capable of identifying trace amounts of tumor DNA and presenting multiple molecular analytes, like ctDNA, cfRNA, proteins, and exosomes, to improve detection rates early on. AI imaging is pioneering the identification of faint patterns out of CT, MRI, and PET scans that the human eye cannot see, and it can predict risk even before lesions appear. Meanwhile, new imaging technologies are being created to detect sub-centimeter cancers at lower doses of radiation and be paired with liquid biopsy data for precise image-guided sampling. Panels of biomarkers, supplemented with machine learning tools, are revolutionizing early-stage cancer screening. These tools integrate analysis of proteins, miRNA, and metabolites to recognize molecular patterns linked to PDAC formation. Finally, nanotechnology is the next frontier, with sensitivity in diagnosing and potentially curing cancers at the molecular level.</p><p>Through continued study and multidisciplinary collaboration, these technologies will likely progress far beyond their current capabilities. The future holds vast potential for earlier, more accurate, and less invasive detection of pancreatic and other cancers.</p><h4><strong>Conclusion</strong></h4><p>Pancreatic cancer has long been among the toughest cancers to diagnose and treat, based on its aggressive biology and late-presenting symptoms. Yet, with continued investment in research, cross-disciplinary collaboration, and the introduction of next-generation technologies, the future of pancreatic cancer diagnosis is more promising than ever before. While challenges like tumor heterogeneity, low concentrations of ctDNA in early disease, limited datasets, and regulatory hurdles remain, the pace of innovation offers real reason for optimism.</p><p>Every new platform, algorithm, and biomarker panel brings us closer to a revolutionary goal: screening for pancreatic cancer early enough to treat it successfully and save lives. What was once nearly unthinkable is now possible. We are moving toward a future where pancreatic cancer will no longer be a silent killer, but a disease that we can screen for early, treat with precision, and one day conquer.</p><p>The path ahead is clear: tireless research, technological innovation, and cross-disciplinary collaboration to rewrite the landscape of pancreatic cancer detection. With these advancements, we can envision a future where early detection is commonplace, interventions are timely and personalized, and survival rates soar exponentially. The horizon is now marked not by limitations but by possibility for a future where pancreatic cancer can be anticipated, intercepted, and conquered.</p><h4><strong>References:</strong></h4><p>Mayo Clinic. (n.d.-a). <em>Neuroendocrine tumors — Symptoms &amp; causes</em>. Retrieved October 23, 2025, from<a href="https://www.mayoclinic.org/diseases-conditions/neuroendocrine-tumors/symptoms-causes/syc-20354132?utm_source=chatgpt.com"> https://www.mayoclinic.org/diseases-conditions/neuroendocrine-tumors/symptoms-causes/syc-20354132</a></p><p>Mayo Clinic. (n.d.-b). <em>Pancreatic cancer — Symptoms &amp; causes</em>. Retrieved October 23, 2025, from<a href="https://www.mayoclinic.org/diseases-conditions/pancreatic-cancer/symptoms-causes/syc-20355421#dialogId24828029"> https://www.mayoclinic.org/diseases-conditions/pancreatic-cancer/symptoms-causes/syc-20355421#dialogId24828029</a></p><p>Ottenhof, N. A., et al. (2011). <em>Molecular characteristics of pancreatic ductal adenocarcinoma</em>. PMC. Retrieved from<a href="https://pmc.ncbi.nlm.nih.gov/articles/PMC3068308/?utm_source=chatgpt.com"> https://pmc.ncbi.nlm.nih.gov/articles/PMC3068308/</a></p><p>Yonezawa, S., et al. (2008). <em>Precursor lesions of pancreatic cancer</em>. PMC. Retrieved from<a href="https://pmc.ncbi.nlm.nih.gov/articles/PMC11854847/"> https://pmc.ncbi.nlm.nih.gov/articles/PMC11854847/</a></p><p>GRAIL, Inc. (2025, June 18). <em>GRAIL announces positive top-line results from the Galleri® PATHFINDER 2 registrational study</em> [Press release]. Retrieved from<a href="https://grail.com/press-releases/grail-announces-positive-top-line-results-from-the-galleri%E2%93%A1-pathfinder-2-registrational-study/?utm_source=chatgpt.com"> https://grail.com/press-releases/grail-announces-positive-top-line-results-from-the-galleri%E2%93%A1-pathfinder-2-registrational-study/</a></p><p>Guardant Health, Inc. (2025, April 22). <em>Guardant Health to present data highlighting advances in multi-cancer detection and liquid and tissue molecular profiling at AACR</em> [Press release]. Retrieved from <a href="https://www.biospace.com/press-releases/guardant-health-to-present-data-highlighting-advances-in-multi-cancer-detection-and-liquid-and-tissue-molecular-profiling-at-aacr">https://www.biospace.com/press-releases/guardant-health-to-present-data-highlighting-advances-in-multi-cancer-detection-and-liquid-and-tissue-molecular-profiling-at-aacr</a></p><p>BioSpace. (2021, September 29). <em>New study finds Freenome’s multiomics approach successful in detecting pancreatic cancer</em> [Press release]. Retrieved from<a href="https://www.biospace.com/new-study-finds-freenome-s-multiomics-approach-successful-in-detecting-pancreatic-cancer?utm_source=chatgpt.com"> https://www.biospace.com/new-study-finds-freenome-s-multiomics-approach-successful-in-detecting-pancreatic-cancer</a></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=50a6a3b94a0f" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[The Emerging Role of AI Facial Analysis in Early Symptom Detection and Diagnosis of Rare Genetic…]]></title>
            <link>https://semilogooketola.medium.com/the-emerging-role-of-ai-facial-analysis-in-early-symptom-detection-and-diagnosis-of-rare-genetic-919a02d71e86?source=rss-8a02bfe1359e------2</link>
            <guid isPermaLink="false">https://medium.com/p/919a02d71e86</guid>
            <category><![CDATA[genomics]]></category>
            <category><![CDATA[facial-analysis]]></category>
            <category><![CDATA[symptom-detection]]></category>
            <category><![CDATA[rare-disease]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <dc:creator><![CDATA[Semilogo Oketola]]></dc:creator>
            <pubDate>Sat, 11 Oct 2025 18:39:48 GMT</pubDate>
            <atom:updated>2025-10-29T01:08:06.116Z</atom:updated>
            <content:encoded><![CDATA[<h3><strong>The Emerging Role of AI Facial Analysis in Early Symptom Detection and Diagnosis of Rare Genetic Diseases</strong></h3><p>What if a simple picture could reveal hidden genetic markers linked to disease?</p><p>Our faces contain more information than we realize; subtle signs shaped by our deoxyribonucleic acid (DNA). Genetic information housed within the DNA influences not only inner biology but also exterior features; and in some cases, these differences are visible in facial features. Behind every outwardly visible human trait, such as our colored eyes, facial features, even the way bodies respond to disease, are complex patterns of genetic instructions. These instructions, held in our DNA, shape not only who we are, but also how our bodies function and occasionally fall apart. Mapping genetic information to their physical resultant biological features has never been easy. This is likely because the majority of diseases and characteristics are the product of the interaction between many genes, not just a singular change or mutation, and determining their causes genetically is a complex puzzle.</p><p>For geneticists, genotype to phenotype mapping has long been the key to unlocking disease. But traditional mapping requires years of patient history, expensive sequencing, and expert analysis. Even with these tools, some genetic diseases are elusive, going undetected for years or even mislabeled altogether. This is especially relevant in the cases of rare or complex diseases. While each occurs infrequently, rare genetic diseases affect over 300 million individuals worldwide, majority of whom experience delayed or misdiagnoses due to similar or unfamiliar presentations of their illnesses.</p><p>Advanced AI developments have made it possible to identify patterns in complex or rare diseases. With programs like DeepGestalt and GestaltMatcher, AI has demonstrated promise in identifying rare genetic conditions from facial features with increased accuracy. The science of AI-driven phenotype–genotype mapping could bring new opportunities into the medical field.</p><p>While most current studies are focused on inherited conditions/diseases, in the future, genetic mapping with AI-driven phenotype-genotype could be useful for more complex diseases like cancer. For example, research has shown that pancreatic cancer has established connections with mutations such as KRAS and TP53. Thus, this brings up the important question: could AI spot the early subtle visual cues linked to the mutations that result in conditions like pancreatic cancer?</p><p>In this article, I will discuss how AI links genetic markers with visual features, highlight existing breakthroughs, and consider what this might mean for the future of research and diagnosis of rare genetic diseases.</p><h4>Table of Contents</h4><p>∘ <a href="#feae">The Science of Genotype and Phenotype:</a><br> ∘ <a href="#03b3">AI Meets Genetics: The Rise in Facial Analysis:</a><br> ∘ <a href="#797c">Case Study: SHEPHERD</a><br> ∘ <a href="#4162">The Future of AI Facial Analysis</a><br> ∘ <a href="#2c9e">Conclusion:</a><br> ∘ <a href="#b2b3">More About Me:</a><br> ∘ <a href="#44bd">Reference List:</a></p><h4><strong>The Science of Genotype and Phenotype:</strong></h4><p>Variations in the human body are moderately to highly heritable; and resemblances between relatives show the strong relationship between inherited genetics and phenotypes. The craniofacial structure is incredibly polygenic(controlled by multiple inheritable genes). In a 2012 GWAS study, researchers found that variants near PAX3 and TP63, which include rs7559271, rs974448 and rs9288572, are associated with facial features like eyebrow shape or nose width. The effects of genetic variants on phenotype often depend on genetic background; however, many mutations have inconsistently penetrant effects even when the background is controlled. The genetic background refers to all the other genes within an organism i.e the rest of the genome that might influence how a mutation behaves. When scientists “control the background,” it means they’re keeping the other genes, the environment and any other influencing factors constant, so that only the mutation should change. Even if two organisms are genetically identical except for one mutation, that mutation could affect one and not the other, making its effect inconsistent.</p><p>This leads us to the broader concept of genotype-phenotype (geno-pheno) mapping. Many loci (the fixed physical location of a gene on a chromosome) across the genome, which is the entire set of genetic information (DNA or RNA) within an organism, each contribute small, additive effects, such as the shape, size, and width of the nose, for instance. These effects cause the variants or mutations. Geno-pheno mapping seeks to look at patterns across many variants rather than single mutations visible in genes. It tries to find connections between specific genetic loci and specific phenotypic outcomes, for example finding the link between <em>PAX1</em> and <em>GLI3</em>, which influence nostril breadth, and <em>RUNX2</em>, which controls the width of the nasal bridge. Some mutations are incompletely penetrant meaning that not everyone with the mutation shows the trait. In essence, geno-pheno mapping is a complex process and is not always straightforward. The same variant can have different phenotypic effects.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/834/1*0KPYsglgWXQO4rm8PW6TQA.png" /><figcaption><strong>Figure 1: Illustration of complete and variable genetic penetrance and expressivity in affected individuals.</strong></figcaption></figure><p>There are 2 basic approaches to mapping genetic effects to phenotypic variation.</p><ul><li>One measures genotypic values with known physiological and biochemical relationships to the phenotype of interest.</li><li>The other uses random markers to associate any potential genetic variant with variations in the trait of interest</li></ul><p>One could also utilize a one-shot model by encoding genotypes and phenotypes jointly in a shared representation space where<strong> </strong>both<strong> </strong>genotype patterns and phenotypic traits are embedded, and the relationships are learned directly, rather than running multiple independent mappings. This same logic of mapping patterns applies not only to visible traits, but also to complex diseases such as cancer.</p><p>Just like craniofacial traits, cancers are not only driven by a singular gene. In cancers, genetics are involved through a cascade of accrued mutations (genetic changes that accumulate over time in an organism’s cells) that change the functioning of crucial genes that regulate cell growth, division, and death. The majority of these mutations are acquired throughout one’s life, but a few percent are inherited. These changes ruin the normal cell cycle, and cells start growing and dividing abnormally and give rise to tumors. The genes associated with pancreatic cancer for example, include mutations like APC, KRAS, TP53, CDKN2A, SMAD4 as well as inherited risk genes like BRCA2, PALB2. Some of these are inherited germline mutations (permanent changes to a parent’s reproductive cells that are incorporated into the DNA of every cell in the offspring’s body), while others are acquired somatic mutations (DNA changes in a body’s non-reproductive cells that occur after conception). The phenotypes produced by the TP53 and KRAS mutations include extreme tumor aggressiveness, poor prognosis, and resistance to therapy.</p><p>Here, mapping focuses on linking specific mutations with clinical outcomes. For example, BRCA (BReast CAncer gene) mutation has been found to be connected with improved response to PARP inhibitors, as well as the KRAS mutation with tumor aggressiveness, and tailoring therapy based on mutations. The genotype-phenotype relationship in cancer is complex, but decoding it can guide more precise diagnosis and treatment, just as it does for traits as simple as facial morphology. Understanding how genes map to traits lays the foundation for training AI models that can link genomic markers with visual features.</p><h4><strong>AI Meets Genetics: The Rise in Facial Analysis:</strong></h4><p>While geno–pheno mapping has helped us to understand genetic traits, there are still speed and scalability limitations. In the past, clinicians relied on experience and close observation, but AI tools now make it possible to detect these traits with greater speed and accuracy. AI has also opened up new opportunities in genetics, allowing computers to catch subtle physical characteristics that until now would likely be picked on only by very experienced clinicians. By combining deep learning with big data sets of images, researchers have begun training algorithms to spot genetic patterns directly from the face, a field known as AI-based facial phenotyping.</p><p>Along with recent technological advancements, AI is being trained to read genetic information from faces using computer vision and deep learning models called convolutional neural networks(CNNs). These are specialized algorithms that can detect patterns in images by recognizing edges, curves and textures across multiple layers of abstraction. These algorithms are trained to identify complex visual patterns by correlating thousands of local features like lines, edges and facial contours and comparing them with resembling genotypes. These models have significantly improved how facial patterns are analyzed in genetics, often detecting features that are barely visible to the human eye.</p><p>One of the most effective of the AI-driven genetic diagnostic breakthroughs, DeepGestalt, developed by scientists at FDNA, a biomedical company focused on using artificial intelligence to analyze facial photos and other clinical information, in 2019, uses CNNs to compare patient images with a reference library of known genetic syndromes; sometimes more accurately than doctors in diagnosing. It was trained on over 17,000 images of patients with more than 200 genetic syndromes, allowing it to compare new patient photos inputted during trials, to its database. Following its clinical trials, DeepGestalt achieved top 10 diagnostic accuracy rates of above 90% in identifying rare syndromes, even surpassing clinicians.</p><figure><img alt="Composite photos and test set results of the Specialized Gestalt Model." src="https://cdn-images-1.medium.com/max/1024/1*tNM3RtM6I2k5TxK1uZwA7A.png" /><figcaption><strong>Figure 2: Composite photos and test set results of the DeepGestalt Model.</strong></figcaption></figure><h4><strong>Case Study: SHEPHERD</strong></h4><p>Another incredibly successful deep learning model is SHEPHERD. It is a few-shot learning model developed by a team headed by Dr Marinka Zitnik in 2022, which is more focused on rare genetic diseases. The model inputs patient phenotype terms and a list of candidate genes, and operates at multiple points throughout the rare disease diagnosis process to perform causal gene discovery, retrieve “patients-like-me” with similar genetic and phenotypic features, and provide interpretable names for atypical symptoms in its knowledge graph.</p><p>To overcome the limitations of supervised learning, SHEPHERD was trained primarily on simulated rare disease patients, created using an adaptive simulation approach that generates realistic rare disease patients with varying numbers of phenotype terms and candidate genes, and learned to incorporate medical knowledge of known phenotype, gene, and disease associations. Through this diagnostic process, patients were annotated with a set of Human Phenotype Ontology (HPO); phenotype terms describing their clinical features and a set of candidate genes that may explain the patient’s syndrome.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*4Rv-MEBsTQ1t5JKuPvxhRg.png" /><figcaption><strong>Figure 3: Overview of SHEPHERD in rare disease diagnosis.</strong></figcaption></figure><p>Using SHEPHERD, rare disease experts only needed to examine 1026 genes from the expert-curated lists or 18,005 genes from the variant-filtered lists to identify the causal gene for 465 UDN (Undiagnosed Diseases Network) patients during one of their clinical trials. In comparison, with non-guided ranking, the experts would have needed to evaluate a total of 2231 expert-curated genes or 27,727 variant-filtered genes. SHEPHERD reduced the number of genes that the experts needed to consider by up to 23.3% and continues to show the potential to improve diagnostic efficiency. It was discovered that SHEPHERD was able to identify the causal gene among its top five predictions even when the patient’s presenting phenotypic abnormalities were multiple hops away from the gene causing the disease in the knowledge graph.</p><p>Patient UDN-P1 is a remarkable example of SHEPHERD accuracy. Three years after being enrolled in the study, she was diagnosed with POLR3-related leukodystrophy. Her case was also the most complex, because she had symptoms like lack of tear production, hearing loss, and hypertension, which are not typical for leukodystrophy. Only 13 out of her 46 phenotype traits were directly related to the causal gene, POLR3A; and on average, the gene was 1.98 hops away in the knowledge graph. SHEPHERD nonetheless ranked POLR3A as the leading candidate among 17 expert-curated and 86 variant-filtered candidates.</p><p>The model’s attention mechanism allows it to “weight” the phenotypic traits in different ways, up-weighting more prominent ones and down-weighting distractor traits, so that the model itself becomes resilient to atypical presentations. This suggests that SHEPHERD does not just match patterns, it also reads context. These results show how AI programs such as SHEPHERD can significantly accelerate gene discovery and diagnosis of rare disease, and reveal patterns that classical analysis would not find. As these technologies develop further, they may become valuable aids to clinicians, complementing expert judgment with computer vision.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*oHjILeLs_DhSGEEyMvLt6w.png" /><figcaption><strong>Figure 4: Causal gene discovery case studies for Patient UDN-P1 for leukodystrophy</strong></figcaption></figure><h4>Challenges</h4><p>Despite the progress of AI-driven facial analysis, several key challenges continue to limit its clinical usefulness and global applicability. First, the lack of diverse, high-quality datasets remains a major barrier. Most existing facial image databases overrepresent individuals of European descent, leading to algorithmic bias and reduced diagnostic accuracy for underrepresented populations. Rare diseases are themselves uncommon, so there’s not a lot of labelled data to train models to be highly robust.</p><p>Additionally, the “black-box” quality of most deep learning systems cuts down on interpretability; physicians can’t see what facial features were considered in making a prediction, so it’s harder to validate or trust the results. Furthermore, variations in lighting, pose, age, and image quality can all significantly affect model performance, and there is still little evidence of the performance of such systems in real-world, clinical settings outside of controlled studies. Ethical and privacy concerns surrounding the collection and storage of sensitive facial data also hinder large-scale deployment.</p><p>Finally, deployment in clinical workflows requires standardized protocols, regulatory clearance, and physician training so that AI will be used as a reliable diagnostic aid and not an opaque, autonomous entity. Surmounting these challenges, through heterogeneous data curation, explainable AI development, cross-institutional validation, and open governance frameworks, will be the key to realizing the complete potential of facial analysis for rare disease diagnosis.</p><h4><strong>The Future of AI Facial Analysis</strong></h4><p>AI facial analysis supplements traditional genotype–phenotype mapping by providing an instantaneous, visual view of gene expression in human traits. When used with genomic sequencing, these tools narrow down the range of potential genetic mutation territory to search through, making diagnoses quicker and more precise. Over time, this union might create a feedback loop where visual data provides genetic discovery with input, while genetic data improves the precision of AI models.</p><blockquote>“What once took years of testing can now take days. AI is changing how we look at genetic data entirely.” - Dr. Kimberly Huygen</blockquote><p>This marks a milestone in the understanding of the human genome by scientists and clinicians. The future of diagnosis may shift from laboratory-based testing to the utilization of smart visual recognition.This technology has already assisted doctors in identifying rare conditions in patients who might have otherwise gone undiagnosed for months or years. In addition to providing increased speed and precision, it could cut down the cost of diagnosis for patients with limited access to expensive labs and personalize treatments based on their DNA. It could also potentially uncover new, unsuspected gene-trait relationships. It is important to note that this technology is designed to support clinicians, not replace them, extending human expertise rather than eliminating it.</p><p>With every development in the way technology inscribes itself into genetics, tools like AI-based facial analysis reframe diagnosis and knowledge of disease. Already, we have witnessed their capacity to accelerate detection and disclose genetic clues imperceptible to the human eye. With careful refinement, this AI and human intuition alliance shows promise for a new era of precision in medicine, where faces do not just tell their stories but speak of their very code of life.</p><h4><strong>More About Me:</strong></h4><p>My name is Semilogo Oketola. I’m a student with a strong interest in genomics, computational biology, and artificial intelligence. I’m passionate about exploring how data-driven models and AI can advance precision medicine and rare disease diagnostics. If you are interested in connecting with me, email me at <a href="/service/mailto:semilogooketola@outlook.com">semilogooketola@outlook.com</a></p><h4><strong>Reference List:</strong></h4><ul><li>Gurovich, Y., Hanani, Y., Bar, O., Fleischer, N., Gelbman, D., Basel-Salmon, L., … &amp; Basel-Vanagaite, L. (2019). Identifying facial phenotypes of genetic disorders using deep learning. <em>Nature Medicine, 25</em>(1), 60–64. <a href="https://doi.org/10.1038/s41591-018-0279-0">https://doi.org/10.1038/s41591-018-0279-0</a></li><li>Alsentzer, E., Bohutinska, A., Lopez, R., et al. (2025). SHEPHERD: Few-shot learning for phenotype-driven diagnosis of rare diseases. <em>npj Digital Medicine, 8</em>(1). <a href="https://doi.org/10.1038/s41746-025-01749-1">https://doi.org/10.1038/s41746-025-01749-1</a></li><li>Schmitz-Abe, K., Murdock, D. R., Mahamdallie, S. S., et al. (2021). Using simulated patients to improve rare disease diagnosis with SHEPHERD. <em>eLife, 10</em>, e68623. <a href="https://elifesciences.org/articles/68623?utm_source=chatgpt.com">https://elifesciences.org/articles/68623</a></li><li>Face2Gene: Using deep learning for genetic syndrome recognition. (2018). <em>National Center for Biotechnology Information (NCBI)</em>. <a href="https://pmc.ncbi.nlm.nih.gov/articles/PMC6008943/">https://pmc.ncbi.nlm.nih.gov/articles/PMC6008943/</a></li><li>Garvan Institute of Medical Research. (n.d.). <em>Genotype vs phenotype</em>. <a href="https://www.garvan.org.au/news-resources/science-explained/genotype-phenotype">https://www.garvan.org.au/news-resources/science-explained/genotype-phenotype</a></li><li>Elsevier. (2025). <em>Integrating AI into facial phenotyping: Ethical and diagnostic perspectives.</em> <em>ScienceDirect</em>. <a href="https://www.sciencedirect.com/science/article/pii/S2666389925000236">https://www.sciencedirect.com/science/article/pii/S2666389925000236</a></li><li>Nemours Biomedical Research. (n.d.). <em>Genotype–phenotype relationships and modern AI interpretation.</em> Elsevier Pure. <a href="https://nemours.elsevierpure.com/">https://nemours.elsevierpure.com</a></li><li>FDNA. (2020, March 1). <em>How AI helps doctors identify rare genetic conditions</em> [Video]. YouTube. <a href="https://youtu.be/OejH-QteqVU">https://youtu.be/OejH-QteqVU</a></li></ul><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=919a02d71e86" width="1" height="1" alt="">]]></content:encoded>
        </item>
    </channel>
</rss>