Loading…
Structure-based drug design that schedules noise separately for 3D coordinates and 2D topology, reaching a 95.9% PoseBusters valid rate on CrossDock.
No providers recorded yet. Browse all providers
A generated ligand can score well against a pocket and still be a molecule no chemist would draw: bond lengths off the reference distribution, a strained ring, an internal energy no conformer search would tolerate. PoseBusters-style physical checks catch these, and generators that look strong on the CrossDock test set lose much of their geometric validity when the targets change. The cause traced here is the noise schedule, not model capacity. A pocket-conditioned generator builds two things at once — continuous 3D atom positions and a discrete 2D graph of atom types, bonds, charges and aromaticity — and the schedule decides which becomes clean first. When both advance on a shared timestep the path is 3D-driven: the model learns to denoise topology given clean coordinates, but gains nothing from clean topology when placing atoms.
MolPilot repairs the path rather than the network. For a single modality the variational lower bound depends only on the endpoints of the noise schedule and not on its shape, which is why schedule design is normally a free hyperparameter. That invariance breaks once two modalities are coupled: the bound becomes a path integral over the plane of continuous and discrete noise levels, so the trajectory across that plane changes it. VLB-Optimal Scheduling (VOS) turns this into a search. Training samples the two timesteps independently, giving a generalized objective defined over the whole plane instead of one diagonal; the trained model's loss surface is then gridded and dynamic programming returns the monotone path of least cumulative cost, which becomes the sampling schedule. That path is two-stage: shape-driven sketching fixes 3D geometry first, then topology-driven docking adjusts the conformation to the settled graph.
The model comes from the Institute for AI Industry Research at Tsinghua University, with the Shanghai Institute of Materia Medica and Sichuan University, and was presented at ICML 2025. It is the third project in the MolCRAFT series, after MolJO, and like its siblings is built on a Bayesian flow network, which carries a distribution over the clean molecule rather than a noisy sample — the property that makes two modalities commensurable enough to schedule against each other.
The network is an SE(3)-equivariant graph transformer with a bond-prediction head over four bond classes, trained on the standard CrossDocked2020 split of 100,000 training poses and 100 validation poses, re-preprocessed with kekulized bond orders. Training runs 30 epochs with independent timestep sampling; the released checkpoint is epoch 26, and generation uses 100 sampling steps. On the in-distribution CrossDock test set MolPilot reports a 95.9% PoseBusters passing rate against 84.6% for MolCRAFT, with 98.5% passing the internal energy check, 97.9% fully connected, and an average Vina Score of -6.88 kcal/mol. On 180 out-of-distribution PoseBusters proteins filtered to under 30% sequence identity with the training sequences it holds 79.1%, where baselines fall to between 32% and 63%. Used as a docking model on 298 PoseBusters V2 complexes, 44.0% of its poses land within 2 Å RMSD of the crystal ligand while staying physically valid, and its ProLIF interaction profiles match the co-crystallized references at a Tanimoto similarity of 0.551.
The intended use is hit generation against a solved or modeled structure: a script takes a protein PDB file plus a reference ligand, clips a 10 Å pocket around it, and samples molecules from the fixed released checkpoint with no retraining per target. Conditioned instead on a given 2D graph, the same checkpoint runs as a docking model. The emphasis on geometric validity matters where compounds feed straight into synthesis or free-energy triage.
The lasting contribution is that a multi-modal generative model's schedule design space has a real optimum that can be solved for rather than guessed — an argument that reaches any generator mixing continuous and discrete variables. The evaluation is entirely computational, using docking scores and physical validity checks as stand-ins for binding. The paper notes that its CrossDock scores exceed those of the test set's own reference ligands and discusses possible information leakage in the standard split, which is why the out-of-distribution results carry more weight. Code, weights and the preprocessed training data are released under CC BY-NC-SA, so commercial use is not permitted.
Much of this page is generated or calculated automatically. Flag anything that looks off and we will re-run it.