Pearl: A Foundation Model for Placing Every Atom in the Right Location [pdf] (genesis.ml)

GenesisMolecularAI announced Pearl, a generative “foundation” model for protein–ligand cofolding that sets a new state-of-the-art for predicting accurate, physically valid binding poses. Pearl combines large-scale synthetic training data with architectural advances and inference-time controllability to overcome the data scarcity, physics violations, and limited conditioning common in prior models. On public benchmarks (Runs N’ Poses and PoseBusters) it achieves ≈85% success (RMSD < 2 Å and PB-valid), outperforming AlphaFold 3 and open-source baselines by ~14% and showing almost no loss when physical validity checks are applied. On a proprietary set of challenging drug targets, Pearl yields up to ~3.6× improvement at the stricter RMSD < 1 Å threshold important for medicinal chemistry. Technically, Pearl’s key innovations are: (1) training recipes that mix experimental and large-scale synthetic protein–ligand complexes (performance scales with synthetic dataset size); (2) an SO(3)-equivariant diffusion module that respects 3D rotational symmetry to improve generalization, sample efficiency, and produce physically plausible geometries; and (3) a generalized multi-chain templating system enabling flexible conditioning (pocket-conditional or unconditional modes) on protein, cofactor, and non‑polymeric context. Together with curriculum training and a diffusion-based generative structure module, these advances make Pearl both more accurate and more controllable for structure-based drug discovery workflows.