Research map / Label-efficient and robust learning / Recognition and long-tail
A Standardized Benchmark for Multilabel Antimicrobial Peptide Classification
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 ESCAPE, a standardized multilabel AMP dataset compiled from 27 repositories with a benchmark of seven AMP classifiers and a transformer baseline fusing peptide sequence and predicted 3D distance-matrix inputs via bidirectional cross-attention.
Problem
Antimicrobial peptide (AMP) prediction lacks a standardized, multilabel benchmark: data is fragmented, annotations are inconsistent, and prior models are mostly binary and evaluated on non-comparable splits.
Approach
Merge 27 AMP databases plus curated non-AMP negatives into a five-label multilabel hierarchy (antibacterial, antifungal, antiviral, antiparasitic, antimicrobial), benchmark seven existing methods on it, and propose a dual-branch transformer that fuses sequence tokens with CA-distance-matrix patches through bidirectional cross-attention.
Key results
- ESCAPE Baseline reaches 72.1% mAP and 69.8% F1 averaged over three seeds, versus 70.3% mAP and 68.5% F1 for the second-best AMPlify.
- Antiparasitic AP improves to 37.6% from 27.7% for AMPlify (about 35.7% relative).
- Sequence-only input gives 69.4 mAP vs 47.7 for structure-only; combining via cross-attention gives 72.7 mAP on seed 42.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Combine sequence and structural representations with bidirectional cross-attention | ESCAPE Baseline · ESCAPE Benchmark | F1-score: 68.5\pm0.77 -> 69.8\pm0.43 | Helped |
| Combine sequence and structural representations with bidirectional cross-attention | ESCAPE Baseline · ESCAPE Benchmark | mAP: 70.3\pm0.87 -> 72.1\pm0.60 | Helped |
| Add the sequence module and cross-attention to a structure-only model (ablation) | ESCAPE Baseline · ESCAPE Benchmark, seed 42 | mAP: 47.7 -> 72.7 | Helped |
| Add the sequence module and cross-attention to a structure-only model (ablation) | ESCAPE Baseline · ESCAPE Benchmark, seed 42 | F1-score: 46.9 -> 69.5 | Helped |
| Add the structural module and cross-attention to a sequence-only model (ablation) | ESCAPE Baseline · ESCAPE Benchmark, seed 42 | mAP: 69.4 -> 72.7 | Helped |
| Add the structural module and cross-attention to a sequence-only model (ablation) | ESCAPE Baseline · ESCAPE Benchmark, seed 42 | F1-score: 67.6 -> 69.5 | Helped |
| Replace experimental + generated structures with only generated structures (ablation) | ESCAPE Baseline · ESCAPE Benchmark, seed 42 | mAP: 72.7 -> 71.2 | Hurt |
| Replace experimental + generated structures with only generated structures (ablation) | ESCAPE Baseline · ESCAPE Benchmark, seed 42 | F1-score: 69.4 -> 67.5 | Hurt |
Surprising findings
- Model size does not correlate with performance; the second-best method is a Random Forest (amPEPpy), and BERT-based models do not rank in the top three.
- Structure-only input performs much worse than sequence-only input (47.7 vs 69.4 mAP).
Limitations
- Dataset cannot capture the full natural diversity of peptides.
- Experimental structures are available for only 2.5% of peptides, so most inputs use predicted structures that reduce performance (mAP 72.7 to 71.2).
- Sequence length distributions differ between AMPs and non-AMPs, which may introduce a shortcut.
Experimental setup
Datasets: BIOPEP-UWM CPPsite 2.0 CAMPR3 TumorHoPe APD3 SPdb ParaPep CancerPPD BrainPreps Quorumpeps YADAMP LAMP2
Benchmarks: ESCAPE benchmark
Metrics: mAP F1-score
Baselines: AMPlify AVP-IFT amPEPpy PEP-Net AMP-BERT TransImbAMP AMPs-Net
Builds on: TransImbAMP AVP-IFT AMPlify AMP-BERT
Closest papers by meaning
- MAC-AMP: A Closed-Loop Multi-Agent Collaboration System for Multi-Objective Antimicrobial Peptide Design (ICLR 2026 · similarity 0.84)
- PepSpecBench: A Unified Evaluation Benchmark for Peptide Tandem Mass Spectrometry Prediction (NeurIPS 2026 · similarity 0.83)
- Full-Atom Peptide Design with Geometric Latent Diffusion (NeurIPS 2024 · similarity 0.82)
- Universal Biological Sequence Reranking for Improved De Novo Peptide Sequencing (ICML 2025 · similarity 0.81)
- Full-Atom Peptide Design based on Multi-modal Flow Matching (ICML 2024 · similarity 0.81)
- CL-MFAP: A Contrastive Learning-Based Multimodal Foundation Model for Molecular Property Prediction and Antibiotic Screening (ICLR 2025 · similarity 0.81)
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.