Oct 14, 2026 · 1h 50m
LeashBio tests whether chemistry models learn science or shortcuts
Lab-in-the-Loop at Compute Speed — Ian Quigley, Leash Bio
Reliable molecular models depend on datasets that expose causal binding behavior rather than the habits, artifacts, and selection biases surrounding experiments.
- 1DNA-encoded libraries make protein–small-molecule binding measurements scalable, but tags, batch effects, and assay artifacts complicate interpretation.
- 2LeashBio’s failed Kaggle competition showed that large datasets can still reward memorization instead of generalization across proteins and chemicals.
- 3Causal activity cliffs and lab-in-the-loop simulations offer a path toward agents that learn from verifiable molecular consequences.
Don't miss
Ian Quigley describes an experiment in which models identify individual chemists from molecular structures, revealing how easily chemistry systems can learn creator fingerprints instead of chemical physics.
The brief
Ian Quigley explains why LeashBio built DNA-encoded libraries: sequencing turns pooled binding experiments into scalable measurements, but controls must separate genuine hits from barcode, DNA-tag, and assay artifacts.
The company’s Kaggle competition delivered a sharp warning: despite an unusually large dataset, winning models performed no better than chance, forcing a rethink of diversity, splits, and molecular representations.
Hermes and harder activity-cliff tests expose how models can memorize chemical similarity, molecular weight, or chemist fingerprints instead of learning binding mechanisms that transfer out of distribution.
LeashBio responds with causal experiments, mutating protein pockets and competing ligands to create informative negatives that reveal which molecular changes actually alter binding.
The longer-term vision is a language-model agent trained on measured counterfactuals, exploring medicinal-chemistry decisions in a virtual lab whose feedback remains tied to physical experiments.