Pocket-based molecule generation with an SE(3)-equivariant language model leads to a potent and selective HPK1 inhibitor with in vivo efficacy

Deep learning shows promise in structure-based drug discovery, yet challenges persist in generating pharmacologically plausible molecules with valid 3D conformation and decent binding mode in the pocket. We introduce SE3-BiLingoMol, an SE(3)-equivariant Transformer for pocket-based 3D molecule generation, addressing two key limitations of existing language-model approaches. First, it uses Geometric Algebra Transformers for SE(3)-equivariant handling of continuous 3D coordinates. Second, a bidirectional attention mechanism mitigates conformational errors accumulated during autoregressive sampling. These innovations enable SE3-BiLingoMol to generate 2D drug-like, 3D geometrically valid molecules with superior binding modes. Validated on DUD-E dataset containing over 100 targets, the model achieved state-of-the-art performance in de novo design and optimization. We applied SE3-BiLingoMol to design potent and selective inhibitors for HPK1, a promising immunotherapy target. Through an iterative human-AI workflow, integrating AI generation with experimental validations (X-ray crystallography, bioassays), we identified Cmpd. 6. This novel tetracyclic compound demonstrates potent HPK1 inhibition, excellent cellular activity, favorable pharmacokinetics, and robust anti-tumor in vivo efficacy as monotherapy and with PD-1 blockade. Our work establishes a sophisticated generative AI framework for 3D molecule design and demonstrates its application in developing cancer immunotherapy.