Research Publications

We present MolGen-2, a transformer-based generative model trained on 1.8 billion chemical structures that designs novel small molecules optimized simultaneously for target binding affinity, ADMET properties, and synthetic accessibility. In prospective validation across 14 oncology targets, 31% of generated candidates showed sub-nanomolar IC50 values.

We describe a two-stage pipeline combining backbone diffusion with sequence design for the de novo engineering of enzymes with novel catalytic activities. Across 47 designed enzymes, 38 folded as predicted and 19 demonstrated measurable catalytic activity, with six achieving kcat/KM values within 10-fold of natural homologs.

Analysis of whole-genome sequencing data from 87,000 individuals across 12 ancestral populations identified 344 novel genetic loci associated with rare metabolic disorders, with AI-derived polygenic risk scores demonstrating 3.2-fold improvement in predictive accuracy over existing models for non-European ancestry groups.

We describe AutoLab-1, a fully autonomous drug discovery platform that integrates AI experimental design, robotic liquid handling, in-line mass spectrometry, and reinforcement learning to iteratively optimize lead compounds. In a blinded comparison against parallel human teams, AutoLab-1 achieved equivalent lead quality 40-fold faster with 60% lower reagent consumption.

RepurposeAI, a heterogeneous knowledge graph combining FDA drug databases, proteome-scale interaction data, and clinical outcome records, identified 18 FDA-approved drugs with computational evidence for repositioning against neglected tropical diseases. Six candidates were confirmed active in cell-based assays at clinically achievable concentrations.

A prospective multicenter study of 1,240 patients with advanced non-small cell lung cancer demonstrated that AI-guided tumor genomic analysis reduced median time to matched therapy selection from 21 days to 5.7 days, with the AI-recommended regimen concordant with multidisciplinary tumor board consensus in 89% of cases.