We built Manas AI because wet labs keep synthesizing what computers should have eliminated.

The standard virtual screening workflow hands a pharma team 2,000-5,000 candidates ranked by shallow docking scores. Those scores correlate poorly with actual binding — Pearson r around 0.4-0.5 for most commercial tools. The team synthesizes the top 50, runs binding assays, and gets back a hit rate of 2-5%. Millions in synthesis cost, months of time, and most of the information could have been predicted computationally with better models.

Siddhartha Mukherjee spent several years watching that happen at a Boston biotech before deciding to rebuild the stack from scratch. The core insight: binding affinity prediction only works at scale if you train on heterogeneous structural data spanning multiple target classes, not just the kinase-heavy datasets most published models use. The ADMET gap is even larger — the 8-12 endpoints most tools cover leave out the assays that actually kill Phase I candidates.

Manas AI was founded in Cambridge in 2023. We built the binding affinity GNN, the 48-endpoint ADMET ensemble, and the pan-proteome off-target screen as one integrated pipeline. We are not a general-purpose virtual screening tool, and we do not sell software licenses. We run campaigns — computationally intensive, scientifically accountable, ending at an IND-enabling candidate with a documented rationale. Our measure of success is that number, not hit-list length.