Research map / Vision-language and multimodal / Human and embodied modeling
Understanding protein function with a multimodal retrieval-augmented foundation model
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.
Introduces PoET-2, a 182M-parameter multimodal, retrieval-augmented protein foundation model with an order-equivariant hierarchical encoder and dual causal/masked decoders, achieving SOTA zero-shot indel/higher-order variant effect prediction and data-efficient supervised function prediction.
Problem
Protein language models struggle with indels, higher-order (epistatic) mutations, data-efficient supervised function prediction, and scaling does not improve mutation understanding; existing PLMs do not combine multimodal inputs with retrieval augmentation.
Approach
PoET-2 conditions on a set of homologous sequences (context) and optional structure/partial query in an order-equivariant hierarchical encoder, with a causal decoder for exact likelihoods (including indels) and a masked decoder for embeddings, trained jointly with three losses.
Key results
- Zero-shot DMS indels Spearman 0.566 vs 0.517 for PoET-1 and about 0.466 for the best non-PoET model; PoET-2 is SOTA on indels.
- Zero-shot DMS substitutions: PoET-2 + VenusREM ensemble reaches 0.543 Spearman overall vs 0.519 for VenusREM; PoET-2 alone 0.500.
- Supervised ProteinGym DMS: PoET-2 GP average Spearman 0.693 vs 0.664 for Kermut, and it matches Kermut with at most 250 training points in contiguous split.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add predicted structures to the context (ablation) | PoET-2 zero-shot scoring · ProteinGym DMS indels | Spearman ρ: 0.55589 -> 0.56666 | Helped |
| Add predicted structures to the context (ablation) | PoET-2 zero-shot scoring · ProteinGym DMS substitutions | Spearman ρ: 0.47534 -> 0.48374 | Helped |
| Add WT structure as the query (ablation) | PoET-2 zero-shot scoring · ProteinGym DMS substitutions | Spearman ρ: 0.47534 -> 0.49128 | Helped |
| Ensemble sequence-only context and structure-context predictions (ablation) | PoET-2 zero-shot scoring · ProteinGym DMS substitutions | Spearman ρ: 0.47534 -> 0.48260 | Helped |
| Ensemble WT-structure query strategies with and without structure context (ablation) | PoET-2 zero-shot scoring · ProteinGym DMS substitutions | Spearman ρ: 0.49128 -> 0.49256 | Helped |
| Ensemble sequence-only context with different queries, including WT structure (ablation) | PoET-2 zero-shot scoring · ProteinGym DMS substitutions | Spearman ρ: 0.47534 -> 0.49632 | Helped |
| Ensemble structure-context strategies with different queries (ablation) | PoET-2 zero-shot scoring · ProteinGym DMS substitutions | Spearman ρ: 0.47534 -> 0.49987 | Helped |
| Ensemble all four structure-conditioning strategies (ablation) | PoET-2 zero-shot scoring · ProteinGym DMS substitutions | Spearman ρ: 0.47534 -> 0.49989 | Helped |
| Add explicit structure conditioning (ablation) | PoET-2 supervised function prediction · ProteinGym supervised DMS substitutions | predictive performance | No clear effect |
| Use WT structure as an inverse-folding query (ablation) | PoET-2 clinical variant prediction · ProteinGym clinical variants | predictive performance | Hurt |
| Use PoET-2 rather than PoET-1 | Zero-shot variant scoring · ProteinGym DMS indels | Spearman ρ: 0.517 -> 0.566 | Helped |
| Use PoET-2 rather than PoET-1 | Zero-shot variant scoring · ProteinGym clinical substitutions | AUROC: +0.008 AUROC | Helped |
Surprising findings
- Structure conditioning helps zero-shot stability prediction but gives little or no benefit for clinical variant prediction and supervised function prediction.
- PoET-2 alone exceeds the PoET-2 + VenusREM ensemble on higher-order (3+) substitutions.
Limitations
- Lags slightly behind structure-based masked models (VenusREM) on single-substitution DMS, driven by stability datasets.
- Structure conditioning offers little benefit outside stability prediction.
- Supervised GP kernel may not be optimal for multi-mutation variants.
Experimental setup
Datasets: UniRef50 v2304 AlphaFoldDB predicted structures ProteinGym DMS substitutions and indels ProteinGym clinical benchmark ProteinGym supervised DMS benchmark
Benchmarks: ProteinGym DMS substitutions ProteinGym DMS indels ProteinGym clinical substitutions and indels ProteinGym supervised DMS benchmark
Metrics: Spearman rho AUROC NDCG mean squared error
Baselines: ESM-2 ESM C ProGen2 SaProt ESM-3 Open ProSST MSA Transformer TranceptEVE L GEMME PoET-1 S3F-MSA VenusREM
Builds on: PoET-1 MSA Transformer Kermut T5 relative position bias ProteinGym
Closest papers by meaning
- Prot2Text-V2: Protein Function Prediction with Multimodal Contrastive Alignment (NeurIPS 2025 · similarity 0.91)
- DPLM-2: A Multimodal Diffusion Protein Language Model (ICLR 2025 · similarity 0.89)
- From Mechanistic Interpretability to Mechanistic Biology: Training, Evaluating, and Interpreting Sparse Autoencoders on Protein Language Models (ICML 2025 · similarity 0.88)
- LEMON-ZEST: Evolution-Informed Tokenization for Efficient Protein Language Modeling (NeurIPS 2026 · similarity 0.88)
- Prot2Text: Multimodal Protein’s Function Generation with GNNs and Transformers (AAAI 2024 · similarity 0.88)
- MutaPLM: Protein Language Modeling for Mutation Explanation and Engineering (NeurIPS 2024 · similarity 0.88)
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.