DrugFlow

DrugFlow is a generative model for structure-based drug design that produces a complete 3D ligand for a given protein binding pocket: atom coordinates, atom types, and the bonds between them, generated jointly rather than assembled in stages. It was developed at the Laboratory of Protein Design and Immunoengineering at EPFL by Arne Schneuing, Ilia Igashov, and colleagues, with co-authors at VantAI, the University of Oxford, and Aithyra, and was published at ICLR 2025.

Its central design choice is to use the right generative process for each data type. Ligand coordinates are continuous, so DrugFlow transports them with conditional flow matching. Atom and bond types are categorical, so they are transported with discrete Markov bridges, which interpolate between a categorical prior and the data distribution without embedding discrete labels into a continuous relaxation. The paper also reframes how such models should be judged: rather than reporting the best absolute docking score a sampler can reach, the authors measure how closely the distribution of generated molecules matches real protein-ligand data across chemistry, geometry, and physics, on the grounds that a model never trained to optimize a score should not be ranked by it.

Three extensions build on the base model. An uncertainty head, trained end-to-end, assigns a per-atom confidence that flags out-of-distribution samples. A preference alignment scheme adapted to both flow matching and Markov bridges shifts a pretrained model toward desirable property regions. FlexFlow additionally samples protein side-chain torsion angles alongside the ligand, addressing the same receptor-flexibility problem that Apo2Mol approaches through apo-to-holo pocket generation.

#Key Features

• Multi-domain generative process: Continuous flow matching handles atom coordinates while discrete Markov bridges handle atom and bond types, so each modality is modeled in its natural space.
• Learned molecule size: A virtual-node type lets the model add or delete atoms during sampling, so ligand size is drawn from a learned conditional distribution given the pocket rather than fixed in advance.
• Per-atom uncertainty: An end-to-end uncertainty head assigns confidence scores that correlate with steric clashes and ligand efficiency, and separate distribution tails from modes, giving a ranking signal at inference time.
• Preference alignment: A direct-preference-style objective, formulated for both flow matching and Markov bridges, tunes a pretrained checkpoint toward higher drug-likeness, synthesizability, or docking efficiency.
• Flexible pockets with FlexFlow: The FlexFlow variant jointly samples side chain torsion angles on a hypertorus with the ligand, enabling design against unbound receptor conformations.

#Technical Details

The backbone is an E(3)-equivariant heterogeneous graph neural network built from geometric vector perceptrons, operating on a graph whose nodes are ligand atoms and pocket residues represented by their C-alpha positions plus vector features to each residue atom. Separate message and update functions are learned for ligand-ligand, ligand-pocket, pocket-ligand, and pocket-pocket edges, with a 10 Angstrom cutoff. DrugFlow has 12.1 million weights and samples in 500 steps; FlexFlow has 12.6 million. Training uses 100,000 CrossDocked protein-ligand pairs with 100 held-out test proteins, split at 30% sequence identity, with ligands failing PoseBusters filters removed.

Evaluation measures distances between the generated and training distributions. DrugFlow reaches a Frechet ChemNet Distance of 4.28 against 11.64 for DiffSBDD, 12.70 for Pocket2Mol, and 13.77 for TargetDiff, and leads by a wide margin on the Jensen-Shannon divergence between the joint distributions of drug-likeness, synthetic accessibility, lipophilicity, and Vina efficiency. It ranks first or a close second on every marginal distribution the authors measure, including bond lengths, bond angles, ring systems, and normalized counts of non-covalent interactions. FlexFlow recovers bound side-chain conformations at a median RMSD of 1.75 Angstroms, compared with 1.98 Angstroms for Rosetta repacking in the absence of the ligand.

#Applications

DrugFlow suits hit-generation workflows where the goal is a diverse, physically plausible set of pocket-complementary starting points rather than a single score-maximized compound. Because it learns a distribution rather than an objective, teams can retrain it on a curated in-house corpus or apply preference alignment to move it toward their own criteria, and the per-atom uncertainty gives a cheap filter for discarding implausible samples before docking. FlexFlow extends this to targets available only in apo or predicted form.

#Impact

DrugFlow's most portable contribution is methodological: it argues that generative drug design models should be assessed as distribution learners, and supplies a concrete battery of divergence-based metrics over chemistry, geometry, and interaction profiles to do so. Code is released under an MIT license, and four checkpoints (base, uncertainty-head, FlexFlow, and preference-aligned) along with processed CrossDocked data are archived on Zenodo under CC BY 4.0, which makes both the model and its evaluation directly reusable. The model is compact by current standards at roughly 12 million parameters, and the authors note that its atom-removal capacity is bounded by the maximum number of virtual nodes seen during training, so it cannot yet select molecule size in a fully open-ended way.