FlowSBDD

Structure-based drug design asks a generative model to place a 3D ligand inside a protein pocket: given the pocket atoms and their coordinates, produce atom types and positions for a molecule that binds. Diffusion models took over this task because they generate chemically sensible geometry, but the formulation carries two costs. The training objective is fixed to denoising, which leaves no natural place to add a chemistry-specific term, and the reverse process is a long stochastic chain that makes sampling slow. The initial distribution is likewise fixed to Gaussian noise, discarding any prior knowledge about what a plausible ligand looks like.

FlowSBDD, from the University of Texas at Austin group behind rectified flow with a collaborator at Rutgers University, rebuilds the task on rectified flow instead. Rather than learning to reverse a noising process, the model learns a velocity field that transports an initial state to a ligand along as straight a path as possible, solved at inference by an ordinary differential equation. Because the training objective is a transport cost rather than a denoising likelihood, extra loss terms and alternative starting distributions slot in directly. The work was accepted to the ELLIS 2024 ML4Molecules workshop, and joins pocket-conditioned generators such as TransDiffSBDD in the catalog.

#Key Features

• Rectified flow transport: A neural velocity field is trained to follow the linear interpolation between the initial state and the true ligand, for atom coordinates and one-hot atom types jointly, then integrated with an Euler solver at sampling time.
• Bond distance loss: An auxiliary term penalises the difference in Euclidean distance between bonded atom pairs, summed over every bond in the ligand, which supplies explicit geometric supervision the coordinate loss alone does not.
• Flexible prior distribution: The Gaussian starting point can be rescaled or replaced entirely — including by the output of another generator — so sampling becomes a refinement of an existing distribution rather than a walk from pure noise.
• Fast sampling: The straight-path formulation needs far fewer integration steps than a diffusion reverse chain, which is where the speed advantage comes from.

#Technical Details

The interaction between ligand atoms and pocket atoms is modelled with an SE(3)-equivariant graph neural network in which hidden embeddings and coordinates are updated alternately in each layer. Training uses CrossDocked2020, filtered to complexes with docking RMSD below 1 and sequence identity below 30%, giving 100,000 protein-ligand complexes for training and 100 held-out complexes for testing. Generated molecules are scored with AutoDock Vina.

Under its best setting the model reports an average Vina Dock score of -8.50 and a median of -8.36, ahead of DecompDiff by 1.3% on the average, with high-affinity rates of 63.4% and 70.9%, QED of 0.47, synthetic accessibility of 0.51 and diversity of 0.75 — a diversity margin of 4.2% over TargetDiff. Ranked by median Vina energy per pocket, FlowSBDD produces the best-binding molecules for 50% of the 100 test targets, against 24% for TargetDiff, 17% for Pocket2Mol and 9% for an autoregressive baseline. Generating 100 molecules for one pocket takes 144 seconds with 100 Euler steps, compared with 3,428 seconds for TargetDiff and 7,785 for the autoregressive model. Ablating the bond loss costs 17.6% of median QED and 29.4% of average Vina Dock.

#Applications

The setting is hit generation against a known protein structure: supply a pocket and the model proposes drug-like 3D molecules positioned inside it, with no per-target retraining and no manual specification of the binding site. The sampling speed makes large candidate libraries per target practical, and the replaceable prior means the model can be used as a refinement stage over ligands produced by molecular dynamics or by another generator.

#Impact

FlowSBDD's contribution is to show that the flexibility rectified flow brings — arbitrary auxiliary losses, arbitrary source distributions — translates into pocket-conditioned molecule generation, at diffusion-level quality with an order of magnitude less sampling time. The evidence base is narrow: results come from a short workshop paper confined to the CrossDocked2020 benchmark with docking scores as the only affinity proxy, with no wet-lab validation, and the authors note the model underperforms on the Vina Score metric. No code, weights or license have been released, so independent reproduction is not currently possible.