Jake Smith

We know a molecular ML model will struggle when it deals with unfamiliar input molecules. This is essentially the out-of-distribution problem in statistics. But here’s the intricacy of molecular prediction: the chemical space is orders of magnitude larger than any available training set, and we often aim to explore structurally distinct compounds in search of new opportunities. That means we can’t afford to be overly restrictive by discarding unfamiliar inputs. Instead, we need to quantify how unfamiliar a molecule is to the model, so we can better assess the reliability of the model prediction. That’s the motivation behind a new study by Derek van Tilborg et al.. They introduced a Joint Molecular Model that simultaneously learns: 🔹Chemical property prediction 🔹Molecular reconstruction from latent space Here’s the clever part: the model’s reconstruction error serves as a proxy for unfamiliarity. If a molecule can’t be accurately reconstructed from its latent representation, it likely lies outside the model’s learned chemical space, indicating lower confidence in the prediction. This strategy of joint modeling offers a practical, interpretable way to map the model’s limits. Additionally, the unfamiliarity landscape may be leveraged to guide workflows like virtual screening and reinforced learning applications. 📄 Molecular deep learning at the edge of chemical space, ChemRxiv, Aug 11, 2025 🔗 https://lnkd.in/eV56xDU5

174

4 Comments

People often ask me for ideas of projects they can work on to solve real-life problems in computational biology. Well, St. Jude Children's Research Hospital is hosting a biohackathon 9/29 - 10/1! All levels welcome. I think these are amazing learning opportunities. Link to hackathon in comments. Many challenges are available to participants. One stood out to me as particularly useful. ~~~~~~~~~~~~~~~~~~~~~~~~~ Title: Omics Dataset Validation Prior to Download Challenge Summary: I want to make it much easier to evaluate publicly available omics datasets—like single-cell RNA-seq, spatial transcriptomics, and DNA methylation arrays—before we actually download anything. The main source for these datasets will be the Gene Expression Omnibus (GEO). The goal is to avoid spending time and resources on downloading large amounts of data, only to find out that most of it isn’t useful for our research. For example, we want to quickly spot datasets that are missing gene names, lack spatial coordinates, or have their data split across multiple files instead of being neatly organized in a single h5ad object. To do this, we’re planning to build a tool that can cross-check datasets by looking at their metadata and file formats. This will involve some web scraping to automatically collect and organize all the relevant metadata from GEO. The main output will be comprehensive metadata tables that summarize the key features and file details of each dataset. We’ll also include some visualization options—like bar plots, scatter plots, and filtered tables or SQL queries—to help us explore and filter the datasets more effectively. Once we’ve curated a solid list of promising datasets, we can then use existing tools like GEOquery to download and assess only those that are most likely to meet our needs. I think the biggest challenge will be building a good web scraping module for GEO. I already have a pretty clear idea of several key variables we can extract and where to find them on GEO DataSets web site, but I have no experience with web scraping. That part will probably be the most difficult, and it’s something I’ll need to learn as we go. There are some existing solutions out there, like Polly (Elucidata) and GEO2R, which are designed to offer a complete end-to-end analysis pipeline. However, they don’t really address this initial step of getting a clear overview of the data format availability for every queried dataset. Our focus is specifically on making this first stage of dataset evaluation much more transparent and efficient. Number of Team Members Needed: This team (3 to 6 members) would like to recruit additional members… Useful Tools/Packages/Software: Web scraping (e.g. Selenium), data visualization (R, Python), familiarity with GEO datasets, API development, and experience with OMICs data formats (e.g., RNA-seq, spatial OMICs, etc) Submitter: Maycon Marção, Scientist, St. Jude Children's Research Hospital ~~~~~~~~~~~~~~~~~~~~~~~~~

140

15 Comments

🧫💻🧬🤓 A challenge in analyzing scRNA-seq data is that critical genes like transcription factors are lowly expressed and hard to detect, while signal from highly expressed, less relevant genes like ribosomal genes dominate the analysis.Our latest work addresses just this! 📄 Paper: https://lnkd.in/g8Uz4dNv We quantify gene importance by measuring how reliant the model is on each gene for embedding each cell by probing the model weights for encoding that cell. Turns out that those lowly expressed transcription factors are really important and those ribosomal genes aren’t very important. And it quantifies this for all genes! This framework scales beautifully, allowing users to search 20 million cells for combinations of important genes (i.e. gene signatures) in just two minutes. Using this tool, we searched for the MS1 gene program across 400 studies. This revealed the MS1 signature in myeloid cells from patients with septic shock, severe COVID-19, and unexpectedly, Kawasaki disease. With collaborators at Boston Children's, we validated these findings, confirming that MS1-like cells are present in Kawasaki disease and even identified a treatment that may reduce MS1 cell prevalence! 💻 Code: https://lnkd.in/gyKb4m4g From Maxwell Gold, Tommaso Biancalani, Hector Corrada Bravo, Miguel Reyes and many other great colleagues! #SingleCellRNAseq #Genomics #FoundationModels #AIinMedicine #Biotech #DrugDiscovery #Immunology #Genentech

413

7 Comments

Stanford and UCSD researchers just made it possible to train medical segmentation AI models with as few as 9 examples. Semantic segmentation is critical for clinical AI, but progress has been constrained by the need for thousands of expert-annotated masks. 𝗚𝗲𝗻𝗦𝗲𝗴 𝗶𝘀 𝗮 𝗴𝗲𝗻𝗲𝗿𝗮𝘁𝗶𝘃𝗲 𝗔𝗜 𝗳𝗿𝗮𝗺𝗲𝘄𝗼𝗿𝗸 𝘁𝗵𝗮𝘁 𝗽𝗿𝗼𝗱𝘂𝗰𝗲𝘀 𝗮𝘂𝘅𝗶𝗹𝗶𝗮𝗿𝘆 𝗱𝗮𝘁𝗮 𝗲𝗻𝗱-𝘁𝗼-𝗲𝗻𝗱, 𝘂𝘀𝗶𝗻𝗴 𝘀𝗲𝗴𝗺𝗲𝗻𝘁𝗮𝘁𝗶𝗼𝗻 𝗽𝗲𝗿𝗳𝗼𝗿𝗺𝗮𝗻𝗰𝗲 𝗶𝘁𝘀𝗲𝗹𝗳 𝘁𝗼 𝗴𝘂𝗶𝗱𝗲 𝘁𝗵𝗲 𝗴𝗲𝗻𝗲𝗿𝗮𝘁𝗶𝗼𝗻 𝗽𝗿𝗼𝗰𝗲𝘀𝘀.  1. Matched or exceeded baseline model performance using 8–20× less labeled training data across 11 segmentation tasks and 19 datasets.  2. Improved Dice scores by 10–20% in low-data settings like placental vessel, polyp, and foot ulcer segmentation with as few as 40–50 examples.  3. Outperformed UNet and DeepLab in out-of-domain generalization, e.g., achieving a 16.1% Dice gain over nnUNet when segmenting lungs across datasets.  4. Eliminated need for unlabeled images by learning from real mask-to-image generation and using a three-level optimization strategy that aligns generation with validation performance.  5. Extended to 3D tasks using Pix2PixNIfTI, improving volumetric segmentation of hippocampus and liver even with only 40 volumes. I thought this was pretty cool: GenSeg’s three‑tier multi‑level optimisation (MLO) nests GAN training, synthetic‑data creation and segmentation validation so the generator’s architecture is tuned by segmentation loss From what I've seen in AutoML pipelines, this is conceptually close to differentiable Neural Architecture Search (e.g., DARTS) and reinforcement‑learning coils in AutoML pipelines But here transplanted to the data dimension rather than weights, showing cross‑pollination between NAS research and data‑centric ML. It'd be interesting to see GenSeg on sub‑populations (e.g., skin‑tone stratified ISIC splits) to check for any synthetic‑data bias propagation It's cool to see GenSeg converting scarce annotations into a leverage point rather than a bottleneck Here's the awesome paper: https://lnkd.in/gDPc6ye5 Congrats to Li Zhang, Basu Jindal, Ahmed M. Alaa, Robert Weinreb , James Zou, Pengtao Xie, and co! Free access to the best LLMs and ask any medical questions here: labela.ai/healtholymp I post my takes on the latest developments in health AI – 𝗰𝗼𝗻𝗻𝗲𝗰𝘁 𝘄𝗶𝘁𝗵 𝗺𝗲 𝘁𝗼 𝘀𝘁𝗮𝘆 𝘂𝗽𝗱𝗮𝘁𝗲𝗱! Also, check out my health AI blog here: https://lnkd.in/g3nrQFxW

1,176

21 Comments

Excited to announce Biomni X Tamarind Bio — towards an AI protein designer. We show: 🤖 Agentic protein optimization: Iteratively improve thermostability with LLM reasoning and calling AlphaFold-2, ThermoMPNN, and more. 🧠  Natural language access to cutting-edge tools like Boltz-2, Chai-1, ImmuneBuilder, for anyone. Learn more: https://lnkd.in/d4Semkjd

113

3 Comments

BindCraft is now published in Nature Portfolio! The open-source pipeline for de novo protein binder design achieves significantly improved 10–100% success rates with nanomolar affinity without the need for high-throughput screening or experimental optimization. Designs demonstrated functional and therapeutic potential, from modulating gene editing to controlling allergens. BindCraft also showed generalization beyond known proteins, designing strong binders against a de novo beta-barrel fold with no natural sequence homologues. Since its release last Fall, BindCraft has seen wide uptake from the community under a highly permissive license. Read my latest at GEN (Genetic Engineering & Biotechnology News): https://lnkd.in/gUTDXwDG Martin Pacesa Sergey Ovchinnikov Bruno Correia Pranam Chatterjee

194

2 Comments

Version 2.1 of the QC:MXP repeat-injection derived batch correction app is now available to download at qcmxp.org. Just for fun I thought it would be interesting to revisit an older data set from Craig Wheelock's lab, which Stacey Reinke & Craig published in 2022 https://lnkd.in/dhQjy5as . This 16 batch data set of 880 subjects (+224 pooled QC injections) across 100+ curated metabolites is exceptionally high quality - the best non-quantified data set I have ever seen. Even so, across 16 batches, there was a significant amount of within and between batch systematic bias ("drift"). Fortunately, my spline correction algorithm did a good job reversing this effect.

118

1 Comment

Academic careers follow predictable paths: postdoc, assistant professor, tenure track. Industry careers are more flexible and faster-moving. You can advance through performance, not just time served. Promotions happen based on impact and readiness, not seniority. Be proactive about your development and don’t wait for others to manage your career trajectory. Take ownership of your advancement.

73

15 Comments

Target-aware molecule generation with graph-transformer GANs Drug discovery starts with finding small molecules that stick to the right protein—but the search space is mind-boggling (10³⁰–10⁶⁰ possible molecules). Most AI generators make “drug-like” molecules, yet aren’t steered to a specific target. Atabey Ünlü and coauthors introduce DrugGEN: a generative adversarial network that represents molecules as graphs and uses graph transformers to create target-specific candidates. Instead of optimizing abstract scores, the discriminator learns directly from known binders of a chosen protein and nudges the generator toward that protein’s chemical “neighborhood.” They put DrugGEN to the test on AKT1, a cancer-relevant kinase. The pipeline goes beyond generation—candidates are triaged with docking, deep learning activity prediction, and molecular dynamics. Crucially, several de novo molecules were synthesized and showed low-micromolar AKT1 inhibition in vitro, validating end-to-end practicality. Attention maps over the molecular graph highlight atoms the model deems important, offering interpretability. Why this matters: it’s a clean, target-centric route to novel chemistry at realistic drug sizes, complementary to diffusion/RL approaches. The team released open code and models, so you can retrain DrugGEN on your protein of interest given known actives. Paper: https://lnkd.in/dDwNHaBJ #DrugDiscovery #MachineLearning #DeepLearning #GraphNeuralNetworks #Transformers #GenerativeAI #GANs #ComputationalBiology #AIforScience #Cheminformatics #AIForDrugDiscovery #MolecularDesign #Bioinformatics #MedicinalChemistry #DrugDevelopment #AIResearch

110

2 Comments

Interested in cutting-edge ML x Bio but have more of a computational, data science, or machine learning background? Christopher Fleetwood from Hugging Face compiled an excellent tutorial that I recommend: https://lnkd.in/gfnaZCWG This details the inputs and outputs to perturbation prediction models, the learning task, the loss, and the evaluation metrics. Check out the bottom of the blog for a sneak peek of what's to come with Arc Institute x Hugging Face!

214

4 Comments

Part 2 of ‘Why am I not getting any response after applying to 100 postdoc positions? What is the single most important accomplishment any PI/ group leader would be looking for when hiring postdocs? PUBLICATIONS So, why would you list them at the end of your 5 page CV? Well, in case I forgot to mention, a 5 long long CV for your first postdoc position is also a big ‘NO’. If you have completed your undergrad, masters and PhD back to back with no other relevant work experience in between, then your CV should not be longer than 2-3 pages. Applying with a long 5 page CV doesn’t necessarily mean extraordinary accomplishments. It can also hint at the inability to present relevant information precisely/ concisely. If your research career was extraordinary and you managed to publish 20+ first author papers, list the most relevant 10 in the first page of your CV and add a google scholar link. You can also mention high impact publications individually in the ‘Overview’ or ‘Key Achievement’ section at the top of your CV. For example, ‘20+ first-author publications in high impact peer-reviewed journals including Nature.’ Same applies to bioinformaticians, if you have developed multiple codes, apps, packages, tools, and websites, list the most relevant ones and then link the rest to a common online platform. Reduce white space. Reduce noise. Declutter. Your CV gives you the opportunity to tell your story. Not to write a laundry list of technical skills. #CV #jobapplication #CVtips

66

In this week’s AI ∩ Bio series, we explore a paper that flips the script on traditional drug discovery. Instead of asking what happens if we perturb one target at a time, the authors ask a different question: given a diseased state and a healthy one, what interventions most directly shift a cell from here to there? Summary: Our paper this week introduces PDGrapher, a causally inspired graph neural network model that predicts combinatorial perturbagens capable of reversing disease phenotypes. Essentially we're asking the question: given a diseased state and a desired healthy one, which interventions are most likely to get us to healthy? PDGrapher is built on graph neural networks (GNNs), machine learning models built to learn from data represented as graphs — structures where nodes and edges capture relationships. In biology, those graphs can be networks of genes, proteins, or signaling pathways, making GNNs a natural fit for capturing how cellular systems behave. Using this architecture, PDGrapher directly proposes potential target sets. In tests across 19 datasets spanning 11 cancers, it ranked known drug targets higher and ran up to 25× faster than comparable AI models. Key Scientific Insight: PDGrapher works by embedding gene expression data onto biological networks and linking two modules: one that proposes targets to perturb, and another that predicts what the treated expression profile would look like. A cycle objective ties these together, keeping predictions consistent with biology. >>On chemical perturbations, it consistently outperformed other models, recovering validated oncology targets like KDR (VEGFR2) and TOP2A. >>On genetic knockouts, performance was more variable, reflecting the biological reality that cells often compensate when genes are missing. The key advance is not raw accuracy alone but the problem formulation: shifting from simulating responses to directly identifying interventions that matter. Leadership Angle: For diagnostics and translational leaders, PDGrapher is less a simulator than a decision aid. It offers three important signals for adoption: -->Scalability — direct intervention discovery scales better as the number of possible combinations explodes. -->Generalization — leave-cell-out results suggest some portability across related contexts, a must for preclinical triage. -->Caveats — current evidence is from cell lines and LINCS/CMap profiles; real-world use will require prospective testing in primary cells, tissues, and in vivo systems. Mentorship Angle: For early-career scientists, the lesson is about problem framing. PDGrapher didn’t succeed by adding more complexity but by asking a sharper question: from “what happens if I perturb everything?” to “which interventions directly solve the problem?” The discipline lies in defining the decision, making assumptions explicit, and stress-testing where models weaken. Carry that mindset forward: it’s what turns clever modeling into credible science.

149

10 Comments