Leash Bio is building AI models trained on reproducible molecular-binding data from DNA-encoded libraries—a high-throughput screening method that uses DNA barcodes to test millions of molecules at once. In a podcast interview, co-founder and CEO Ian Quigley explains how the company collects clean experimental data and uses it to train models that avoid common pitfalls: hidden shortcuts in public datasets, memorization of protein-specific patterns, and overfitting to unreliable measurements.
Quigley discusses a large Kaggle competition (BELKA) with four billion measurements that showed even state-of-the-art models struggle on unseen chemical libraries, suggesting current approaches may learn chemist identity rather than true chemistry. Leash's solution prioritizes reproducibility and careful data splits. The company is also exploring how measured chemical landscapes can become a sandbox for reasoning agents—allowing AI systems to propose and evaluate thousands of medicinal chemistry campaigns in a weekend, with rapid experimental feedback.
Key challenges remain: moving hits from lab screens to drugs that work in animals, the cost and complexity of drug discovery, and the limits of DNA-encoded libraries themselves.
This is an AI-written summary. The original reporting belongs to @LatentSpacePod. How we write these · Spot a mistake?