We just published a new benchmark measuring virtual cell models on the same 5 biological tasks. Until now, most virtual cell benchmarks have measured models against individual tasks (perturbation prediction, cell-type embedding, etc). VCBench is among the first to run the models through the same five tasks, against the same simple baselines, scored the same way, with everything published. Our analysis revealed some interesting findings that are invisible to single-task benchmarks. For example, only one model beat the best simple baseline on any task - TranscriptFormer, on predicting protein levels. But the design choice that won it that task destroys its score on another: temporal ordering, i.e. putting cells in developmental order. A single-task benchmark can't see this tradeoff; a multi-task one can. @AppliedSciAI isn’t building a virtual cell model at this stage. Instead we’re building small, deeply-conditioned models in narrow domains - starting with cardiotoxicity. Had we raised a splashy $200m seed, we’d probably be swinging for the fences trying to scale a general-purpose model. But something feels right about constraining ourselves to a very narrow problem faced by biotechs + pharma (efficiently uncovering hidden cardiac safety risks in small molecules), and hopefully over time generalizing the resulting intelligence and real-world data to other organs and biological domains. While it’s unclear what an eventual “virtual cell” might look like - my guess is it’ll be something like a federation of domain-specific models trained on conditional data, sharing what generalizes and keeping what doesn’t. Our goal with VCBench (and the open tools released with the benchmark) is to show where the current generation of virtual cell models plateaus, and what the next generation likely needs. If you're building with AI in life sciences - or want a best-in-class, AI-native screening model for cardiac toxicity in your small molecule programs - we'd love to talk. Blog post: xitter.cf/AppliedSciAI/status/20…

In Silico Predictions 🤝 Agentic Reasoning We're connecting CardioSafe, our new cardiac toxicity prediction model, to Alexandria, our SOTA scientific literature agent. So when a model flags a compound as risky, scientists get the full context around the prediction, not just a number. Traditional toxicity screening tells a chemist which compounds passed or failed. It doesn't tell them the structural features driving liability, which clinical precedents are most relevant, and what to make next. That reasoning requires synthesizing evidence across published literature, patents, experimental assay results, chemical databases & clinical trials. CardioSafe gives you the number; Alexandria gives you the why: what's known about the target and chemical series, what's shown up in trials, what safety signals already exist. A hybrid approach combining predictive toxicology + agentic AI, so discovery teams can move faster while designing better, safer drugs. 🌐 Alexandria: • Ranked #1 on Lab-Bench2 across literature QA, figure reading & data table reasoning. • Multi-source evidence across literature, trials, patents, structure databases & FDA data. • One question, one grounded answer across sources. ❤️‍🩹 CardioSafe: • Predicts blocking across hERG, Nav1.5, Cav1.2, and IKs from one structure. • Trained on the largest curated multi-channel dataset and benchmarked on out-of-distribution compounds. • 76% more blockers caught at equal spend vs. the next best published baseline; 39% lower cost per confirmed-safe lead. We built in this order on purpose. Get the hardest evidence source - the publications - to state of the art first, widen the inputs, then layer prediction models on top. CardioSafe is the first of those models. It won't be the last. The broader question we’re interested in: What should the interface between predictive models + scientific reasoning actually look like? Especially in settings where a wrong answer can waste months, kill a promising series, or miss a safety liability. Curious to hear from medicinal chemists, toxicologists, biologists and AI researchers working in life sciences: When a model flags a compound as risky, what context would you want before trusting or rejecting the prediction? Demo access to our platform is now open - link below!

What would a Figma for science look like? Drug discovery reasoning is inherently spatial and non-linear. A canvas lets the work breathe. Humans will remain in the AI x bio loop for the foreseeable future, and some of the most exciting work is happening at the UX layer.