Diffusion protein language model for de novo design conditioned jointly on GO terms, InterPro domains, EC numbers, motifs, and backbone structure.
No providers recorded yet. Browse all providers
Generative protein models can usually be steered by one kind of instruction at a time: a text prompt, a functional label, a sequence motif, or a backbone to thread. Real design briefs are rarely that simple. A useful enzyme candidate might need to carry a specific catalytic activity, belong to a particular structural family, retain a known binding motif, and fold onto a chosen scaffold — all at once. CFP-Gen, from KAUST and published at ICML 2025, is a diffusion language model built to accept those constraints simultaneously.
The model conditions generation on three distinct kinds of signal. Functional annotations — Gene Ontology terms, InterPro domain identifiers, and Enzyme Commission numbers — are injected through an Annotation-Guided Feature Modulation module that shifts the protein feature distribution according to whichever combination of labels is supplied. Residue-level control, such as a conserved motif or an active-site constellation, is handled by a Residue-Controlled Functional Encoding module that captures residue-wise interactions. Structural constraints enter through an off-the-shelf 3D backbone encoder, which turns the same model into a functional inverse folding system when a backbone is supplied.
Where ESM-3 exposes multimodal prompting over sequence, structure, and function tracks, and ProGen2 conditions on family-level context, CFP-Gen's contribution is composable functional conditioning: annotations can be added, removed, and combined at sampling time without retraining.
CFP-Gen is built on the pretrained DPLM-650M discrete diffusion protein language model and trained in two progressive stages, so the released checkpoints carry roughly 650 million parameters. The general-purpose training set comprises 103,939 SwissProt sequences annotated with 375 GO terms and 1,154 InterPro domains; a separate enzyme set of 139,551 sequences spanning 661 EC labels comes from intersecting SwissProt with the CARE enzyme benchmark. Backbone coordinates are drawn from the PDB and the AlphaFold Protein Structure Database.
On GO-conditioned generation scored by an external function predictor, CFP-Gen with GO, IPR, EC, and motif conditioning reaches a macro F1 of 0.550, exceeding the positive control of natural sequences at 0.522, against 0.189 for DPLM and 0.355 for ProGen2. For IPR-domain-conditioned generation it improves macro F1 over ESM-3 by roughly 30 points. In functional inverse folding it reports 76.39% amino acid recovery with an MRR of 0.882, pLDDT of 83.53, and self-consistency TM-score of 0.889 — a gain of about 9.45 percentage points in recovery over the DPLM baseline at 66.94%, and well above ProteinMPNN at 45.28% and ESM-IF at 57.39%.
The intended users are protein engineers who start from a functional specification rather than a template sequence: designing enzymes with a target EC activity, building proteins that combine domains from different families, generating binder or scaffold variants that must retain a known motif, or redesigning sequences onto an existing backbone without losing catalytic function. Because conditioning is compositional, the same checkpoint supports library-scale generation across many annotation combinations, which suits high-throughput design-build-test campaigns where hundreds of candidates are ordered per round.
CFP-Gen shows that a discrete diffusion protein language model can be steered by several heterogeneous constraints at once without a separate model per constraint type, and that doing so can push in-silico functional scores past those of natural sequences drawn from the same annotation classes. The important caveat is that all reported functionality is assessed by computational predictors and structure-based self-consistency metrics; no wet-lab characterization of generated proteins is reported, so predicted activity should not be read as demonstrated activity. The authors also note that training data inherits the family and function imbalances of curated databases, and they discuss biosecurity considerations explicitly. Official PyTorch code, training and inference scripts, and checkpoints for the general, enzyme, and inverse folding variants are released, with data and weights distributed via Google Drive.
Much of this page is generated or calculated automatically. Flag anything that looks off and we will re-run it.