Sagace
Models for living systems
Sagace is a research lab exploring how models can represent, understand, and reason about living systems.
Biology is difficult to understand and experiments are expensive. Palaestra tests whether models can interpret evidence, form hypotheses, work through failure, and choose the next useful experiment.
Verifiable environments for biological reasoning.
A curated suite of verifiable tasks that train and evaluate models on real biological, chemical, and pharmaceutical reasoning. Each task is usable interoperably — as a reinforcement-learning environment or as a benchmark — with a hidden oracle and a runnable verifier.
Themed collections of related tasks — one box is one benchmark. Public boxes may be used for training; private boxes are held out for evaluation.
Progress in biology should be something you can prove.
Models will reason about living systems only as well as we can measure them. Palaestra builds the verifiable environments to train, evaluate, and trust that reasoning — one task at a time, with a hidden oracle behind every one.
Training frontier models for biology, or evaluating them? Palaestra is opening to research partners.
A field guide to the public benchmarks the community already uses to measure models on biological reasoning — the terrain Palaestra is built to extend.
Mutation-effect and fitness prediction across 250+ deep mutational scanning assays and curated clinical variants.
Notin et al. · 2023Five biologically grounded tasks measuring how well learned protein representations transfer.
Rao et al. · 2019The biennial blind experiment assessing prediction of protein 3D structure from sequence.
Prediction Center · 1994–Palaestra is built for the labs training the models. These are built for the scientists at the bench: each takes a compound and returns more than a number — the pathways that moved, the evidence behind them, and an honest account of where the model stops knowing. All three run in your browser.
Per-pathway predictors — CYP inhibition and metabolism, transporters, nuclear receptors — composed into a liver-injury call you can point a mechanism at, with a live screen against 193 marketed drugs.
Predicted HepG2 single-cell response read as pathway-level stress, combined with exposure to give a triage decision — and an applicability domain that says when to abstain.
Both demos run the shipped models against a fixed compound library. Tell us what you are screening and we will open them up to your series — and to the full technical reports behind each call.
