LeashBio enables 10,000 drug-discovery campaigns in a weekend

Boltz: Boltz refers to the Boltz series of AI models developed for biomolecular structure prediction, co-folding, and binding affinity estimation, originating from MIT research and now part of Boltz PBC. These models integrate structure and affinity prediction to support small-molecule and protein design tasks. In the news, Boltz serves as a key comparator to Hermes, highlighting differences in training data sources and inference efficiency for drug discovery applications. Hermes: Hermes is LeashBio's lightweight transformer model trained solely on the company's proprietary DNA-encoded library binding datasets for protein-ligand interaction prediction. It emphasizes data quality and experimental rigor over architectural complexity to achieve strong generalization. In the news, Hermes is positioned in direct comparison to Boltz as a benchmark for efficient, scalable AI in drug discovery workflows. LeashBio: LeashBio is a biotechnology company focused on generating large-scale experimental datasets of protein-small molecule interactions using DNA-encoded libraries and automated protein production to advance AI-driven drug discovery. It develops proprietary models trained exclusively on its high-quality binding data rather than public sources. In the news, LeashBio is highlighted for its work enabling AI agents to run extensive medicinal chemistry campaigns in simulated environments and for insights from its Kaggle competition on generalization challenges. allmeasures: allmeasures is the X handle of Ian Quigley, co-founder of LeashBio, who specializes in experimental approaches to medicinal chemistry and machine learning applications in drug discovery. He leads efforts in building scalable wet-lab data generation pipelines. In the news, he explains LeashBio's methodologies, including DNA-encoded libraries, model memorization issues, and the shift toward replayable experimental worlds for AI agents. Data-Centric AI: High-quality, purpose-built experimental datasets from controlled lab measurements enable simpler AI models to achieve competitive performance in binding prediction without relying on public or contaminated data sources. Model Comparison: Lightweight architectures trained on consistent proprietary assay data can match or surpass larger, more complex models in specific drug discovery benchmarks while offering substantial gains in speed and practicality. Experimental Simulation: Replayable lab-like environments allow AI agents to iterate through thousands of medicinal chemistry campaigns, bridging wet-lab data generation with reinforcement learning for accelerated discovery cycles.