Research map / Vision-language and multimodal / Human and embodied modeling
MMSite: A Multi-modal Framework for the Identification of Active Sites in Proteins
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.
Builds the ProTAD dataset of 570k protein sequence and multi-attribute text pairs and proposes MMSite, a two-stage "First Align, Then Fuse" framework that uses soft-label cross-modal alignment and fusion to improve residue-level active site identification by PLMs.
Problem
Identifying residue-level active sites in proteins is understudied because precise per-residue annotations are scarce, limiting protein language model performance; textual protein descriptions could provide complementary context but are unavailable at inference time.
Approach
Uses frozen PLM and biomedical language model (BLM) features with manual prompting that turns attribute tables into sentences, a MACross module for inter-attribute and cross-attention, soft-label alignment (KL between semantic similarity distributions) in an Align stage, then fusion attention for prediction; at inference a text generator (Prot2Text) supplies the missing text modality.
Key results
- MMSite reaches Fmax 0.8250, AUPRC 0.8909, MCC 0.8319 on the 10% clustering split, outperforming 21 protein representation learning baselines (best sequence-only S-PLM Fmax 0.7262).
- Adding a BLM to ESM-1v improves Fmax by about 0.16-0.17 over PLM-only input.
- Removing the text modality at training lowers average performance by 0.105 across metrics, and two-stage Align-then-Fuse beats single-stage training.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add MMSite multimodal sequence-text framework | ESM-1b · ProTAD, 10% clustering | Fmax: 0.7052(±0.02) -> 0.8250(±0.02) | Helped |
| add Align (ablation) | MMSite · ProTAD, 10% clustering | Fmax: 0.8021 -> 0.8250 | Helped |
| add MACross (ablation) | MMSite · ProTAD, 10% clustering | Fmax: 0.8152 -> 0.8250 | Helped |
| add STEnc (ablation) | MMSite · ProTAD, 10% clustering | Fmax: 0.8037 -> 0.8250 | Helped |
| add text modality (ablation) | ESM-1b · ProTAD, 10% clustering | Fmax: 0.7052 -> 0.8250 | Helped |
| use two-stage Align-then-Fuse instead of single-stage joint training (ablation) | MMSite with ESM-1b and PubMedBERT-abs · ProTAD, 10% clustering | Fmax: 0.8086 -> 0.8250 | Helped |
| use all text attributes instead of only Function (ablation) | MMSite · ProTAD, 10% clustering | Fmax: 0.8152 -> 0.8250 | Helped |
| add manual prompting (ablation) | MMSite · ProTAD, 10% clustering | Fmax: 0.8157 -> 0.8250 | Helped |
| use Function and remaining attributes as key/value and query, respectively, rather than swapping their roles (ablation) | MMSite MACross · ProTAD, 10% clustering | Avg. (token): 0.8394 -> 0.8493 | Helped |
| use soft-label alignment instead of hard-label alignment (ablation) | MMSite · ProTAD, 10% clustering | Avg. (token): 0.8334 -> 0.8493 | Helped |
| increase sequence-identity clustering threshold from 30% to 50% (ablation) | MMSite · ProTAD | Fmax | Helped |
| add PubMedBERT-abs text features to ESM-1b | ESM-1b · ProTAD | Fmax: +0.120 | Helped |
Surprising findings
- Adding a text modality via a biomedical language model yields large gains even though text is absent at inference, with the missing text generated by Prot2Text.
- Manual prompting improves most but not all metrics; AUPRC is slightly higher without it (0.8911 vs 0.8909).
Limitations
- At inference the text modality is generated by an agent model (Prot2Text), so performance depends on generated text quality
- Evaluation uses a subset of Swiss-Prot proteins with AlphaFold structures and a small test set (696 samples at 10% threshold)
Experimental setup
Datasets: ProTAD (from Swiss-Prot/UniProt) AlphaFold DB structures
Benchmarks: ProTAD clustered splits (10%-90% identity)
Metrics: Fmax AUPRC MCC OS FPR
Baselines: ESM-1b ESM-1v ESM-2-650M ProtT5 ProtBert S-PLM PST ProtST
Builds on: ProtST CLEAN InfoNCE Prot2Text
Closest papers by meaning
- ProtCLIP: Function-Informed Protein Multi-Modal Learning (AAAI 2025 · similarity 0.89)
- Prot2Text-V2: Protein Function Prediction with Multimodal Contrastive Alignment (NeurIPS 2025 · similarity 0.88)
- Prot2Text: Multimodal Protein’s Function Generation with GNNs and Transformers (AAAI 2024 · similarity 0.88)
- From Mechanistic Interpretability to Mechanistic Biology: Training, Evaluating, and Interpreting Sparse Autoencoders on Protein Language Models (ICML 2025 · similarity 0.87)
- LEMON-ZEST: Evolution-Informed Tokenization for Efficient Protein Language Modeling (NeurIPS 2026 · similarity 0.87)
- ProtGO: Function-Guided Protein Modeling for Unified Representation Learning (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.