Peleg, T. | bioRxiv Science

bioRxiv2026

Protify: A Benchmarking Platform Revealing Task Dependence and Homology Leakage in Protein Language Model Evaluation

Protein Language Models (pLMs) have become essential tools for computational protein analysis, and the rapid proliferation of architectures, training objectives, and pretraining corpora has produced a diverse and expanding model landscape. Evaluation practices, however, remain fragmented: studies use different datasets, preprocessing pipelines, and reporting conventions, so results rarely transfer between publications. Here we introduce Protify, an open-source platform that standardizes pLM benchmarking across zero-shot, supervised, and transferlearning settings through a unified, low-code interface. Protify currently supports 35 pretrained models and 61 curated datasets and it is easy to add custom versions of each. Leveraging Protify's high throughput pipelines, we benchmarked 23 models on ProteinGym, discovering that DPLM3B is currently the top-performing self-supervised encoder. Another large-scale supervised evaluation across 32 diverse tasks, spanning enzyme classification, gene ontology annotation, protein-protein interaction, subcellular localization, and other functional properties, found that no single pLM dominated: 14 different models took the top score on at least one dataset, and the margins between the best and second-best models fell within run-to-run variability. On enzyme optimal pH prediction, a general-purpose probe pipeline matched or exceeded EpHod, a specialized model selected from 11,550 training runs, after 50 hyperparameter trials (R2 = 0.482 versus 0.425). Under a homology-controlled split of the same data, every pooled-embedding probe collapsed to near-zero R2, and only a transformer probe over full residue embeddings recovered appreciable signal (R2 = 0.248). Model rankings therefore depend as much on evaluation design as on the models themselves. Protify makes those design choices explicit, controlled, and replayable, and is freely available at https://github.com/Gleghorn-Lab/Protify.

bioinformatics↗

bioRxiv2025

Protein Language Models are Accidental Taxonomists

Protein-protein interactions (PPIs) are fundamental to nearly all biological processes, yet their experimental characterization remains costly and time-consuming. While computational methods, particularly those using protein language models (pLMs), offer higher-throughput solutions, they often report unexpectedly high performance on multi-species datasets. Here, we introduce the accidental taxonomist hypothesis, proposing that neural networks can exploit the phylogenetic distances across labels in protein datasets rather than genuine interaction features. We show that in PPI datasets with random negative sampling, protein pairs for real PPIs are almost exclusively from the same species, while negatives almost always originate from different species. We then demonstrate that pLM embeddings can be used to accurately distinguish whether two proteins share a taxonomic origin, allowing models to "cheat" by learning phylogeny instead of genuine PPI features. By employing a strategic sampling strategy that restricts negative examples to protein pairs from the same species, we reveal a marked drop in model performance, confirming our hypothesis. Compellingly, these strategically trained models still outperform single-species models, suggesting that multi-species data can improve performance if carefully curated. These findings suggest that accidental taxonomist behavior is a particularly influential confounder for PPI, and it is also broadly applicable to any supervised-learning protein dataset.

bioinformatics↗