Research map / Image generation / Autoregressive and latent generation
CFP-Gen: Combinatorial Functional Protein Generation via Diffusion Language Models
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes CFP-Gen, a diffusion protein language model built on DPLM-650M that composes GO/IPR/EC annotation tags, sequence motifs, and backbone structure as conditions for functional protein generation and inverse folding.
Problem
Existing protein language models condition generation on a single constraint from one modality, so they struggle to satisfy functional, sequence-motif, and structural constraints jointly, especially for multi-functional proteins.
Approach
Adds an Annotation-Guided Feature Modulation (AGFM) module that embeds GO, IPR and EC tags and uses them to scale, shift and gate features in each ESM block, plus a ControlNet-style Residue-Controlled Functional Encoding (RCFE) branch for motifs, and reuses a pretrained GVP-Transformer structure adapter, all on top of DPLM with only the new modules trained.
Key results
- On GO-conditioned generation, CFP-Gen (GO, IPR and Motif) reaches macro F1 0.550 vs 0.522 for positive control, with macro AUC 0.795 vs 0.775.
- In functional inverse folding, CFP-Gen (SFT, GO and IPR) reaches AAR 76.39 vs 66.94 for DPLM (+9.45%) and MRR 0.882 vs 0.721.
- For EC-conditioned generation, CFP-Gen (EC, GO, IPR and Motif) reaches macro F1 0.925 vs 0.924 for natural enzymes, against ZymCTRL at 0.774 macro F1 in the subset evaluated.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add IPR conditioning to GO (ablation) | CFP-Gen · GO function generation | macro F1: 0.370 -> 0.458 | Helped |
| add motif conditioning to GO and IPR (ablation) | CFP-Gen · GO function generation | macro F1: 0.458 -> 0.550 | Helped |
| add GO conditioning to IPR (ablation) | CFP-Gen · IPR domain generation | macro F1: 0.826 -> 0.858 | Helped |
| add motif conditioning to GO and IPR (ablation) | CFP-Gen · IPR domain generation | macro F1: 0.858 -> 0.975 | Helped |
| add GO and IPR conditioning to EC (ablation) | CFP-Gen · EC number generation | macro F1: 0.673 -> 0.743 | Helped |
| add motif conditioning to EC, GO and IPR (ablation) | CFP-Gen · EC number generation | macro F1: 0.743 -> 0.925 | Helped |
| add GO and IPR conditioning to EC (ablation) | CFP-Gen · EC number generation | MRR: 0.567 -> 0.774 | Helped |
| use motif conditioning together with EC, GO and IPR (ablation) | CFP-Gen · EC number generation | macro F1: 0.925 vs. 0.924 | Helped |
| fine-tune the structural adapter on the general protein dataset (ablation) | CFP-Gen inverse folding · Functional protein inverse folding | AAR: 72.05 -> 76.39 | Helped |
| use SFT rather than zero-shot (ablation) | CFP-Gen inverse folding · Functional protein inverse folding | MRR: 0.866 -> 0.882 | Helped |
| set condition dropout probability to 0.5 (ablation) | CFP-Gen · Condition dropout evaluation | MRR | Helped |
| use 16 ESM block copies in RCFE (ablation) | CFP-Gen · Motif-conditioned sequence generation | performance | Helped |
Surprising findings
- Generated sequences with GO, IPR and motif conditions score slightly above the positive control (natural sequences) on macro F1 and AUC for GO prediction.
- Adding functional annotations reduces repetitive n-gram patterns compared to DPLM, which the authors interpret as reduced mode collapse.
Limitations
- Supports only a subset of GO, IPR and EC labels, limiting generality.
- Does not yet include physicochemical property conditions or sequence-structure co-design.
- Evaluation relies on automated function predictors (DeepGO-SE, InterProScan, CLEAN) rather than experimental validation.
Experimental setup
Datasets: SwissProt (UniProtKB) InterPro CARE PDB AlphaFold DB CATH
Benchmarks: general protein dataset GO validation split IPR validation subset enzyme EC validation split CATH-based inverse folding
Metrics: MRR MMD MMD-G micro F1 macro F1 AUPR AUC AAR scTM pLDDT
Baselines: DPLM ProGen2 ProteoGAN ESM3 ZymCTRL ProteinMPNN ESM-IF LM-DESIGN
Builds on: DPLM ControlNet ESM2 GVP-Transformer
Closest papers by meaning
- Diffusion Language Models Are Versatile Protein Learners (ICML 2024 · similarity 0.88)
- Diffusion on Language Model Encodings for Protein Sequence Generation (ICML 2025 · similarity 0.88)
- DPLM-2: A Multimodal Diffusion Protein Language Model (ICLR 2025 · similarity 0.87)
- Protein Design with Dynamic Protein Vocabulary (NeurIPS 2025 · similarity 0.86)
- Controllable Protein Sequence Generation with LLM Preference Optimization (AAAI 2025 · similarity 0.86)
- Sequence-Augmented SE(3)-Flow Matching For Conditional Protein Generation (NeurIPS 2024 · similarity 0.86)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.