DisProtEdit

Protein foundation models have made structure prediction and de novo sequence generation routine, but editing an existing protein — changing one property while holding the others fixed — remains awkward. Most multimodal protein–text models compress a sequence into a single holistic embedding, so a prompt asking for more alpha-helical content also perturbs whatever that embedding encoded about function. DisProtEdit, developed at the University of Waterloo with the Vector Institute and the National Research Council Canada, addresses this by splitting the latent space in two: one subspace carries structural semantics, the other functional semantics, and each is supervised by its own natural-language channel.

The training signal comes from SwissProtDis, a dataset the authors built by taking Swiss-Prot protein–text pairs and using a large language model to decompose each curated annotation into two separate descriptions — one structural, one functional — covering roughly 540,000 sequences. A protein encoder and two text encoders are trained so that the protein embedding sits close to the concatenation of the two text embeddings, while a maximum mean discrepancy (MMD) term pushes the structural and functional embeddings onto disjoint, angularly separated regions of the unit hypersphere so they cannot quietly share information.

Editing then happens entirely at inference time. A user rewrites the structural description, the functional description, or both; the model re-encodes the modified text, interpolates between the original protein embedding and the text-guided one, and a decoder reads out a new sequence. There is no retraining and no per-protein optimization. Released in June 2025 and presented at the ICML 2025 GenBio and FM4LS workshops, DisProtEdit builds directly on the text-guided protein design line established by ProteinDT, which serves as its principal baseline.

#Key Features

• Dual-channel language supervision: Every protein is paired with two independent descriptions rather than one, so structural and functional semantics enter the model through separate channels instead of having to be untangled afterwards.
• Angular MMD disentanglement: Instead of a single isotropic prior over the whole latent, the structure and function subspaces are matched to separate priors on disjoint hyperspheres, enforcing independence in both direction and magnitude.
• Alignment and uniformity in place of contrastive loss: Training replaces the usual InfoNCE objective with explicit alignment and uniformity terms, avoiding the large batches and false-negative pairs that complicate contrastive learning in multimodal settings.
• Compositional multi-attribute editing: Because the two channels are independent, structural and functional edits can be issued simultaneously and evaluated jointly — a setting no prior text-guided protein model had targeted.
• SwissProtDis: A released corpus of roughly 540,000 protein sequences, each paired with an LLM-decomposed structural description and functional description derived from UniProt annotations.

#Technical Details

Each text channel is encoded by SciBERT followed by a modality-specific MLP projection; protein sequences are encoded by ProtBERT from the ProtTrans suite, also with an MLP projection head, and an autoregressive T5 decoder reconstructs sequences from the latent. The angular split is set to an even allocation between the structural and functional subspaces, with the uniformity and disentanglement loss weights tuned to 0.2 and 1.0 respectively. On the single-attribute editing benchmark, DisProtEdit raises alpha-helix content in 56.14% of cases and beta-sheet content in 31.58%, against 28.27% and 9.16% for ProteinDT. On a new multi-attribute benchmark of 196 bootstrapped sequences with paired structure–function instructions, it reaches a both-hit success rate of 61.7%. On the TAPE benchmark it is competitive with ProteinDT and OntoProtein on secondary structure and remote homology, and substantially better on the fluorescence and stability regressions (Spearman 0.5373 and 0.8258).

#Applications

DisProtEdit is aimed at protein engineers who want interpretable, incremental control rather than unconditional generation: nudging a scaffold toward more helical content while leaving a catalytic function intact, or altering a stability profile without rewriting the fold. Because the interface is natural language, the same checkpoint serves biologists who can describe a desired change in words but cannot specify it as a structural constraint. The learned representations also transfer to standard property-prediction tasks, so the model doubles as a general-purpose protein encoder.

#Impact

DisProtEdit's contribution is less a raw performance jump than a demonstration that semantic factors in protein–text models can be separated and controlled independently, plus the SwissProtDis corpus that makes that supervision reproducible. The caveats are real and acknowledged by the authors: the LLM decomposition can introduce biologically imprecise text, the T5 decoder tends to memorize frequently occurring fragments and so limits sequence diversity, edits are scored by pretrained oracle predictors rather than experiment, and the multi-attribute benchmark is the authors' own. The work has appeared only at ICML workshops, and while the code is released under an MIT license, no standalone pretrained weights are distributed.