Loop-Diffusion

Two single-residue changes to the same nine-mer peptide antigen can move a T-cell receptor's binding affinity in opposite directions, and only a few hundred TCR-pMHC complexes have ever been experimentally resolved to learn that from. Supervised models fitted to measured binding data inherit the scarcity and batch effects of those assays; physics-based functions such as PyRosetta's cartesian-ddG protocol avoid the data problem but simplify the energetics. Loop-Diffusion takes a third route: it never sees a binding measurement, and instead learns what a plausible loop conformation looks like anywhere in the protein universe.

The premise is that the flexible loops carrying most protein function — the CDR3 loops of immune receptors, the peptide antigen seated in an MHC groove — are, structurally, ordinary loops. Crystallographic conformations sit near minima of a Boltzmann energy landscape, so a model trained to denoise loops inside their fixed local atomic environment is implicitly estimating that energy. Loop-Diffusion uses the denoising diffusion (DDPM) objective but parameterizes its output as the negative gradient of a scalar energy rather than as a noise prediction. At the smallest diffusion time step that energy approximates the physical energy up to a constant, so a mutation's predicted effect is one energy evaluation minus another — no fitting, no regression head, no re-training.

Kevin Borisiak, Gian Marco Visani and Armita Nourmohammad built it at the University of Washington, with a Fred Hutchinson Cancer Center affiliation on the senior author, and posted it in September 2024 alongside a NeurIPS 2024 Machine Learning in Structural Biology workshop poster. It comes from the same group as HERMES but is a distinct model: HERMES scores amino-acid identity from a rotationally invariant projection of a residue's neighborhood, while Loop-Diffusion scores loop coordinates.

#Key Features

• Energy, not just a denoiser: The energy gradient plays the role of the DDPM noise prediction, so a single forward pass returns a scalar usable directly as a score.
• All-atom local context: Each loop is a point cloud of every atom within 10 Å of its alpha carbons, with the environment held fixed and only loop coordinates noised.
• Zero-shot mutational scoring: Binding effects come from wild-type and mutant structures alone, so the model transfers to interfaces it never saw and to targets with no affinity data.
• Scoring only in this release: Training samples diffusion times from the lower half of the noise schedule; extending it to higher noise levels, the step that would grant generative ability, is listed as future work.

#Technical Details

Training data is 433k loop neighborhoods from 20k non-redundant structures in the ProteinNet split of CASP12 at a 30% similarity cutoff. Loops are identified with DSSP inside PyRosetta at lengths of 4 to 20 residues; each atom carries a one-hot element type, a PyRosetta partial charge and a binary loop-versus-environment flag, hydrogens omitted. The energy network is an e3nn graph convolutional network of depth 3, irreducible representations up to l = 4 at multiplicity 8, 3 radial basis functions and 100 radial neurons, plus a 10-dimensional sinusoidal time embedding, trained under a linear beta schedule from 1e-4 to 2e-3 with T = 2000.

Evaluation is on ATLAS, restricted to mutations on the peptide antigen or on one of a receptor's two CDR3 loops. On human MHC class I systems it reaches Pearson r = 0.48 and AUROC 0.79 for peptide mutations (n = 52) against 0.25/0.76 for TCRdock, 0.24/0.60 for DSMBind and 0.35/0.72 for PyRosetta, plus the best CDR3 AUROC at 0.81 (n = 37). Across all ATLAS systems (n = 55 each) it again leads, at AUROC 0.76 and 0.79; TCRdock, built on AlphaFold-Multimer, runs only on class I. DSMBind adds ESM-2 embeddings to its energy model, where Loop-Diffusion uses no sequence model.

#Applications

The immediate use is ranking candidate substitutions when engineering a TCR or a peptide vaccine epitope: given a wild-type complex and an in-silico mutant, the model returns a signed energy difference separating affinity-enhancing from affinity-reducing changes. Because it needs only structures and no target-specific labels, it fits behind any structure source and applies to systems with no measured binding data. The same machinery extends to any functional loop at a protein interface, with antibody CDR3s the context the authors name next.

#Impact

Loop-Diffusion is a compact argument that curating pretraining data to match the geometry of the downstream target matters more than model scale: a modest equivariant network trained on generic loops beats structure-prediction and score-matching baselines on immune-receptor energetics it was never shown. The caveats are equally concrete: correlations are modest in absolute terms, spanning r = 0.20 to 0.48 over a few dozen mutations per set, and the work is a single arXiv version and workshop poster, without peer review and with no code or weights released. Its value sits in the recipe.