A comprehensive benchmark of sequence-based subcellular localization predictors for human proteins

2026-07-06 · Source: Machine learning : nature.com subject feeds · Field: Science & Research — Life Sciences & Biology, Artificial Intelligence & Machine Learning, Mathematics & Computational Sciences · Depth: Expert, extended

Summary

A new comprehensive benchmark for sequence-based subcellular localization predictors for human proteins has been developed, integrating annotations from UniProt, Human Protein Atlas (HPA), and OpenCell. This benchmark, called HOU, features a highly validated test set of 3,814 human proteins, more than twice the size of previous benchmarks, and uses a three-level hierarchical label set for fine-grained, multilabel classification. Researchers systematically evaluated existing models like DeepLoc2, LAProtT5, and MULocDeep, alongside 16 combinations of protein language models (ESM2, ESM3-small-open, ProtT5, ProtBert) and aggregation strategies. The study found that even the best-performing model, ProtT5 with Multihead Attention (ProtT5-MHA), underperforms on fine-grained compartments, multilocalizing proteins, and pathogenic variants known to mislocalize. It also revealed that integrating protein-protein interaction (PPI) data did not substantially improve overall prediction accuracy.

Key takeaway

For AI Scientists and Machine Learning Engineers developing protein localization models, you should focus on creating new architectures that explicitly address multilocalization and fine-grained compartment prediction. Your models must integrate diverse biological data, such as structural information or stability predictors, to improve generalization to pathogenic variants. Consider multimodal protein representation models that learn from various biological signals simultaneously, moving beyond sequence-only approaches.

Key insights

Current sequence-based protein localization predictors struggle with fine-grained compartments, multilocalization, and pathogenic variants, necessitating new approaches.

Principles

• Protein localization prediction is not a solved problem.
• Label frequency significantly impacts model performance.
• Conserved sequence motifs enhance prediction accuracy.

Method

A three-level hierarchical label set was defined, integrating UniProt, HPA, and OpenCell annotations to create a 3,814-protein test set. Models were systematically benchmarked using PLMs and aggregation strategies, with performance assessed across label granularities and protein properties.

In practice

• Use MCC as a robust metric for multilabel evaluation.
• Prioritize models that explicitly handle multilocalization.
• Incorporate structural data or stability predictors for variants.

Topics

• Protein Subcellular Localization
• Protein Language Models
• Deep Learning Benchmarking
• Multilabel Classification
• Pathogenic Protein Variants
• Protein-Protein Interaction Networks

Code references

• CellProfiling/seq2loc_benchmark

Best for: AI Scientist, Research Scientist, Machine Learning Engineer

Related on AIssential

• Viral Proteins Reveal Geometry of Protein Language Models — Protein language models organize sequence embeddings by a "nativeness axis" while preserving specific biological group signals.
• New Benchmarking Shows Limited Generalization Power of TCR Antigenic Epitope Prediction Models — Existing TCR antigen prediction models lack generalization due to insufficient, unbiased benchmark datasets, limiting their utility.
• Integration of alternative fragmentation techniques into standard LC-MS workflows using a single deep learning model enhances proteome coverage — A unified deep learning model and integrated MS platform enhance proteome coverage by leveraging diverse fragmentation techniques.
• Two-dimensional geometric template diffusion for boosting single-sequence protein structure prediction — TDFold uses 2D geometric template diffusion for efficient, accurate single-sequence protein structure prediction.
• MultiMolecule: a modular ecosystem for biomolecular sequence-model workflows — MultiMolecule standardizes biomolecular sequence model reuse by providing complete, source-checked implementations with shared interfaces and provenance.
• Learning the PTM code through a coarse-to-fine mechanism-aware framework — COMPASS-PTM unifies PTM site prediction and enzyme assignment by modeling PTM patterns and catalytic regulators.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Machine learning : nature.com subject feeds.