DTMol

Most deep-learning docking models attack the blind docking problem: given a whole protein and a ligand, find both the binding site and the pose. In practical drug discovery the site is usually already known — from a crystal structure, a homolog, or a pocket-detection run — and the remaining question is narrower and more valuable: how does this molecule sit in this pocket? DTMol targets that pocket-based setting directly, and the restriction pays off twice. The search space collapses to the pocket, and the freed compute budget buys a heavier architecture that can model all-atom interactions between ligand and side chains.

DTMol comes from Carl Kingsford's group at Carnegie Mellon University with collaborators at Shanghai Jiao Tong University and the Institute of Process Engineering of the Chinese Academy of Sciences, and was posted as a preprint in April 2025. It is the first application of the diffusion transformer architecture to pocket-based docking, and it composes two pieces that are usually kept apart: molecular representation encoders pretrained on large small-molecule and protein-pocket corpora, and a generative diffusion process over ligand coordinates. Related catalog entries such as PocketXMol unify docking with design across many tasks; DTMol instead specializes hard on pose prediction inside a known pocket.

#Key Features

• Diffusion transformer for docking: A denoising diffusion process generates ligand poses conditioned on the pocket, with a transformer decoder that attends over every ligand and pocket atom rather than over a coarse graph.
• Parallel invariant and equivariant flows: The decoder runs SE(3)-invariant and SE(3)-equivariant transformer layers side by side, giving the model rotation- and translation-consistent coordinate updates while keeping the implementation and training straightforward.
• Pretrained molecular encoders: Uni-Mol encoders pretrained separately on small molecules and on protein pockets supply the input representations, transferring structural priors from datasets far larger than the set of solved ligand-pocket complexes.
• Wet-lab validated screening: Rather than stopping at retrospective benchmarks, the authors ran a blinded TR-FRET binding assay on the model's top DrugBank candidates against JAK2.

#Technical Details

Noise is applied in the forms best matched to each degree of freedom: a variance-preserving DDPM process perturbs atom coordinates, while a variance-exploding score-matching process perturbs the rotation Euler vector and the translation vector, which start at zero. Generation solves the reverse process, with 20 reverse diffusion steps used at inference. Training ran on the deduplicated PDBBind v2020 collection of 19,120 protein-ligand complexes, with the pretrained encoders frozen for the first 200 epochs and then fine-tuned jointly with the decoder; the full run took 500 epochs across two 80 GB A100 GPUs, roughly seven days, and inference needs more than 20 GB of GPU memory.

On the time-split PDBBind test set of 363 complexes, DTMol places 45.05% of top-1 poses within 2 Å RMSD, against 36.51% for Gnina, 33.22% for DiffBindFR, 28.93% for DiffDock, and 26.64% for Smina. On the 428-complex PoseBusters benchmark, which contains only structures released from 2021 onward and therefore no training overlap, it reaches 77.65% top-1 under 2 Å versus 65.65% for Gnina, with a median top-1 RMSD of 1.253 Å. On the 85-structure PoseBusters Astex Diverse set it reaches 80.05% top-1 under 2 Å.

#Applications

The model fits hit-finding campaigns against targets with a characterized pocket, where thousands of candidates need ranking before any of them reach a bench. The authors demonstrated exactly that workflow: 9,137 DrugBank v5.1.8 compounds were docked against the JAK2 pocket defined by ligand D7D in PDB entry 6BBV, poses were rescored with smina, and the top 20 were combined with the top 20 from AutoDock Vina for testing. Twenty-seven of the 40 were assayed by TR-FRET at 100 nM and 15 were retested at 50 nM, all under blinded compound identities. DTMol's predicted ranking was the only one with a positive Spearman correlation to measured inhibition (0.192, against -0.146 for AutoDock), and two compounds from its candidate list were identified as new JAK2 inhibitors.

#Impact

DTMol's contribution is twofold: it shows that a diffusion transformer with pretrained encoders sets a higher bar on pocket-based docking benchmarks than both physics-based search and earlier generative models, and it backs the claim with prospective wet-lab results, which remains rare for docking papers. The practical caveats are real. The work is a preprint and has not been peer reviewed. The GitHub repository publishes training and inference code but no license file, so no reuse rights are granted, and the trained checkpoints and processed data are distributed only by email request rather than as a download. The model also assumes the pocket is known, which excludes targets where the binding site itself is the unknown.