PPDiff

Designing a protein that binds tightly to a chosen target is the central problem in therapeutic protein engineering, and it fails in two characteristic ways. The designed sequence may not fold into the backbone it was designed for, or the backbone may fold correctly and simply not bind. Most computational pipelines invite both failures by construction: they generate a backbone first, then thread a sequence onto it with a separate inverse-folding model, so nothing in the process ever optimizes sequence and structure jointly. PPDiff, developed at Carnegie Mellon University with NEC Laboratories America, collapses that two-stage pipeline into a single non-autoregressive diffusion model that generates a binder's amino acid sequence and its alpha-carbon backbone at the same time, conditioned on the target.

The technical core is the Sequence Structure Interleaving Network with Causal attention layers (SSINC), which handles the awkward fact that a protein is simultaneously a discrete token string and a continuous point cloud. Interleaved self-attention layers capture global correlations along the sequence, k-nearest-neighbor equivariant graph convolutions model local interactions in three-dimensional space, and causal attention layers on top simplify the sequential dependencies among residues. Diffusion runs on both modalities at once: a categorical corruption process over residue types alongside a Gaussian process over coordinates.

To pretrain this at scale the authors curated PPBench, a general protein–protein complex dataset of 706,360 complexes drawn from the Protein Data Bank — orders of magnitude larger than the specialized complex datasets that preceded it. The pretrained checkpoint is applied to arbitrary targets without per-target retraining, and is separately fine-tuned for two applied settings: target-protein mini-binder design and antigen–antibody design. PPDiff was published at ICML 2025.

#Key Features

• Joint sequence–structure diffusion: Residue types and backbone coordinates are denoised together in a single generative process, so the model never produces a sequence that is inconsistent with the structure it was paired with.
• Interleaved global and local modeling: Self-attention layers capture long-range sequence correlations while kNN equivariant graph convolutions handle 3D neighborhoods, alternating within each block rather than being stacked as separate stages.
• Causal attention over residues: A causal attention head on top of the interleaving stack lets each residue attend to its predecessors, injecting sequential structure into an otherwise non-autoregressive generator.
• PPBench pretraining corpus: A curated set of 706,360 protein–protein complexes from the PDB, released with the code, which supports general-purpose pretraining rather than target-specific fitting.
• Novelty and diversity by design: Because generation is stochastic rather than a single deterministic decode, PPDiff produces candidate sets with high novelty and diversity scores alongside its success rates, giving downstream screening more to work with.

#Technical Details

The architecture comprises three interleaving blocks — each with 11 self-attention layers and one kNN equivariant graph convolutional layer — plus one causal attention layer, totaling 692M parameters. Self-attention layers are initialized from the 650M-parameter ESM-2 checkpoint. Diffusion uses 1,000 steps with a cosine schedule for the sequence channel and a sigmoid schedule for the structure channel. On the general complex design task, designs are scored with ipTM, pTM, PAE, and pLDDT computed by AlphaFold 3; PPDiff reaches a 50.00% top-1 success rate against 47.36% for an SSINC network trained without diffusion, 45.61% for an interleaving-only variant, and 35.08% for a structure-diffusion-then-ProteinMPNN pipeline, while matching the strongest baseline on novelty at 89.46%. Fine-tuned on ten experimentally characterized mini-binder targets and scored by AlphaFold 2 pAE_interaction, it reaches 23.16%; fine-tuned on 4,261 SAbDab antigen–antibody complexes split by CDR cluster at 40% identity, it reaches 16.89% with the best ipTM and novelty among tested methods.

#Applications

PPDiff is aimed at the early, in-silico stage of binder discovery: proposing candidate mini-binders against a therapeutic target, or generating antibody variants against a given antigen, with enough diversity that a downstream filter has real choices. Because the pretrained checkpoint conditions on an arbitrary target structure, it can be run on new targets directly, with fine-tuning reserved for teams holding application-specific complex data. The 3D-coordinate output feeds straight into structure-based filtering before anything reaches a bench.

#Impact

The joint sequence–structure formulation is PPDiff's substantive contribution, and PPBench is a useful artifact independent of the model — a PDB-derived complex corpus at a scale that makes general pretraining for binder design practical. Several caveats matter for anyone evaluating it against alternatives such as RFdiffusion or sequence-space generators like EvoDiff. All reported success rates are in-silico proxies from folding-model confidence scores, with no wet-lab validation; the model represents backbones only as alpha-carbon coordinates rather than full atomic detail; and every design failed against EGFR, which the authors attribute to its shallow binding surface. Adoption is early, and the release carries a NEC Laboratories America evaluation license restricted to non-commercial research use rather than a standard open-source license — a material constraint for industrial users.