Loading…
Discrete diffusion model for protein sequence and MSA generation, enabling controllable de novo design directly in sequence space without structure.
Structure-first protein design generates a backbone geometry and then searches for a sequence that folds into it, which confines it to folds at or near the Protein Data Bank — a small, biased sample of what proteins actually do. Intrinsically disordered proteins are the clearest casualty: they never adopt a stable structure, they carry out essential regulatory and signaling roles, and a backbone-first pipeline has nothing to hand them.
EvoDiff answers this by running the diffusion process over amino-acid tokens rather than coordinates. A forward process corrupts a natural sequence — either by masking one residue at a time in a random order (order-agnostic autoregressive diffusion, OADM) or by mutating residues through a BLOSUM or uniform substitution matrix (discrete denoising diffusion probabilistic models, D3PM) — and a network learns to reverse it; OADM is the stronger of the two for unconditional generation. Because corruption is defined position-by-position, any subset of positions can be held fixed at sampling time. One trained checkpoint therefore generates unconditionally, inpaints a functional domain, scaffolds a motif from its residue identities alone, or writes a new query sequence into an existing multiple sequence alignment. Two families implement this: EvoDiff-Seq over single sequences and EvoDiff-MSA over alignments.
EvoDiff was developed at Microsoft Research in Cambridge, Massachusetts, with collaborators at Microsoft Research AI for Science, the University of Toronto, and Stanford University, and released under an MIT license. The authors are Sarah Alamdari, Nitya Thakkar, Rianne van den Berg, Neil Tenenholtz, Robert Strome, Alan M. Moses, Alex X. Lu, Nicolò Fusi, Ava P. Amini, and Kevin K. Yang. Designs across three systems — disordered mitochondrial targeting signals, metal-binding proteins, and protein binders — were synthesized and shown to express, fold, and carry the expected secondary structure.
evodiff.pretrained, the package installs from PyPI, and the repository ships a Docker image and an Azure AI Foundry notebook.EvoDiff-Seq trains on 42 million UniRef50 sequences using the dilated convolutional backbone introduced with CARP, whose widening receptive fields capture long-range dependencies without attention's quadratic cost. Each corruption scheme is trained at 38M and 640M parameters, with left-to-right autoregressive baselines at both sizes; the 640M model carries the headline results. EvoDiff-MSA uses the MSA Transformer, whose row and column attention model within-sequence and between-sequence dependencies jointly, trained on the OpenFold dataset of 401,381 alignments over roughly 140,000 PDB chains, subsampled to 512 residue positions and 64 sequences either at random or by greedily maximizing diversity. Training consumed 4,128 GPU-hours on 32GB V100s.
Evaluation combines distributional and structural checks: amino-acid and secondary-structure KL divergence, perplexity under protein language models, Fréchet distance, and minimum Hamming distance to the training set to confirm generations are genuinely novel. Foldability comes from a self-consistency pipeline that folds each generated sequence with OmegaFold and re-designs it with ESM-IF1 and ProteinMPNN. Disorder is scored with DR-BERT, and motif scaffolding is benchmarked on the 17 problems compiled for RFdiffusion, judged by predicted local distance difference test and motif RMSD.
EvoDiff suits design problems where a target backbone either does not exist or is not the right specification. Groups working on intrinsically disordered regions — abundant in transcription factors, signaling hubs, and aggregation-prone disease proteins — can generate candidates in a regime structure-based tools cannot address. MSA conditioning supports family-scale engineering: new members that inherit a family's evolutionary constraints while diversifying at tolerant positions, useful as directed-evolution starting libraries. Motif scaffolding applies to enzyme and binder engineering, where a functional site must sit in a new sequence context; because the only input is sequence, it reaches any protein with an alignment.
EvoDiff's durable contribution is the demonstration that discrete diffusion over sequence, conditioned on evolutionary data, yields proteins that survive wet-lab characterization across functionally distinct classes — not just favorable in-silico metrics. Generating disordered proteins fills a real gap in the design toolkit, and the order-agnostic formulation has been picked up in later sequence-generation work. A permissive license and pip-installable checkpoints have made it a common baseline for sequence-space generative methods. The limitations are equally clear: there is no direct structural control, so a user cannot request a target fold, and three-dimensional plausibility must be checked with a separate structure predictor. The 640M flagship is small next to the largest protein language models, and on tightly folded, highly conserved families dedicated structure-based methods remain stronger.
Much of this page is generated or calculated automatically. Flag anything that looks off and we will re-run it.