DSMBind

The Structural Antibody Database holds thousands of antibody-antigen crystal structures but binding affinities for only 566 of them, and the SKEMPI mutation benchmark has roughly 6,000 ΔΔG measurements in total. PDBbind is larger, but mixes IC50, Ki, Kd and EC50 labels that are not comparable to each other. Supervised affinity predictors are therefore trained on either too little data or noisy data, and a bigger network fixes neither.

DSMBind removes the labels from the training loop entirely. It is an energy-based model over a protein complex: an SE(3)-invariant encoder embeds every backbone and side-chain atom, and a feed-forward layer scores each atom pair within a distance cutoff, summing to the binding energy. Training maximizes the likelihood of crystal structures under that energy, using denoising score matching so the intractable partition function never has to be estimated. Because the gradient of the energy with respect to atom coordinates is the force on each atom, the torque between the two binding partners follows directly, and Euler's rotation equations convert that torque into an angular velocity and a rotation matrix. This construction — Neural Euler's Rotation Equations, or NERE — makes SE(3) denoising score matching equivariant without constraining the encoder, and it lets the perturbation be a rigid rotation and a side-chain torsion change rather than Gaussian noise, so every decoy stays chemically plausible.

The work comes from the Broad Institute, presented as NERE at NeurIPS 2023 for protein-ligand and antibody-antigen binding, then extended in a December 2023 preprint under the name DSMBind with protein-protein mutation effects and a wet-lab nanobody design campaign. Two catalog entries build on it directly: EBM-DDG swaps its frame-averaging encoder for invariant point attention, and AdFlip uses DSMBind scores to steer sequence sampling.

#Key Features

• No affinity labels anywhere in training: The training sets are unlabelled crystal structures — 4,806 PDBbind protein-ligand complexes, 3,416 SAbDab antibody-antigen complexes, and roughly 27,000 non-redundant protein-protein complexes — with any measured affinities discarded.
• One energy function across three ligand classes: The same objective covers small molecules (atoms encoded by a message-passing network over the molecular graph), antibodies and nanobodies (residues encoded by frozen 2,560-dimensional ESM-2 embeddings), and general protein-protein interfaces.
• Physics-derived rotation prediction: NERE reads the score of the energy model as an atomic force, accumulates it into a torque, and inverts the inertia matrix to recover the rotation — no keypoint alignment, no spherical harmonics.
• Zero-shot ΔΔG: A mutational effect is the energy difference between mutant and wild-type complex, with no regression head fitted to SKEMPI.
• Virtual-screening throughput: About 0.1 second per complex, roughly 8 million compounds per day, against 4-20 per day for MM/GBSA and FEP+.

#Technical Details

On the 285-complex PDBbind core set, NERE reaches Pearson 0.656 on crystal structures and 0.651 on AutoDock Vina-docked poses, ahead of MM/GBSA (0.647) and its own Gaussian-noise ablation (0.638). On the Merck free energy perturbation benchmark it averages Spearman 0.388 across five seeds, close to the supervised KDeep (0.390). On SKEMPI it reaches Spearman 0.403, matching FoldX (0.407) while far outscoring ESM-1v and inverse-folding baselines. On the 566-complex SAbDab test set it reaches 0.374, above a supervised network on the same encoder (0.350) and an AlphaFold 2 predicted-aligned-error baseline (0.291), and it scores 0.71 AUROC separating trastuzumab CDR3 variants that improve on a 0.2 nM parent. The outputs are comparative: the DSM objective is invariant to an additive constant, so the scores rank complexes but are not calibrated ΔG values.

#Applications

The PD-L1 campaign shows the workflow end to end. All three nanobody CDRs were randomized, ESMFold folded each candidate, template-based docking placed it against a crystallized PD-L1-nanobody complex, and DSMBind scored the result; a student model distilled from 200,000 scored sequences ranked a billion candidates, and the top 24 from each of two libraries went to ELISA. Two bound PD-L1 at 200 nM and 500 nM, one of them specifically, with under 30% CDR identity to anything in the training set. The same ranking step slots into small-molecule virtual screening, antibody affinity maturation, and interface mutation scanning.

#Impact

DSMBind established that the likelihood of a complex structure predicts binding better than the likelihood of a protein sequence, a result protein language models cannot reach and that downstream energy models and guided-design samplers have since taken up. The MIT-licensed code ships the three inference checkpoints the tutorial notebook loads, with training and evaluation data archived on Zenodo. NVIDIA distributes a separately trained DSMBind checkpoint through BioNeMo under Apache 2.0; it was trained on 25,561 PDB complexes with the Gaussian-noise objective and reports Pearson 0.64 on CASF-16, so it is a different artifact from the authors' weights. The repository pins PyTorch 1.13, sidechainnet below 0.7.6, and a development branch of SRU++, which makes environment setup the main obstacle to reuse.