Newest papers

ProtModernBERT: Compute-Efficient Protein Language Modeling

Alberto Martínez, Natalia Flechas Manrique, Llorenç Espinosa-Portalés, Luc Andrea · 2026 · arXiv (Cornell University)

DOI: 10.48550/arxiv.2610.19738 · cited by 0 works (OpenAlex) · open-access PDF

Spanish

Los modelos de lenguaje de proteínas grandes proporcionan un poderoso marco de datos para aprender relaciones de secuencia-función y guiar el diseño experimental. Sin embargo, las mejoras arquitectónicas no se exploran, especialmente en regímenes de bajo cálculo de relevancia para laboratorios de investigación con recursos limitados. Mostramos que el coste de capacitación para las arquitecturas de transformadores únicamente codificadores se puede reducir aprovechando los refinamientos arquitectónicos modernos, ofreciendo un intercambio de rendimiento / eficiencia computacional alentador. Introducimos ProtModernBERT, un modelo modernizado de estilo BERT pre-entrenado en secuencias de aminoácidos, y lo comparamos con ProtBert, Amplify y ESM2 en un conjunto de tareas estructurales supervisadas -comparando la predicción

Machine translation (Meta M2M100); the English original follows.

Abstract

Large protein language models provide a powerful framework for learning sequence-function relationships and guiding experimental design. However, architectural improvements are underexplored, especially in low-compute regimes of relevance for resource-constrained research laboratories. We show that the training cost for encoder-only transformer architectures can be reduced by leveraging modern architectural refinements, offering an encouraging performance / compute-efficiency trade-off. We introduce ProtModernBERT, a modernized BERT-style model pre-trained on amino acid sequences, and compare it with ProtBert, Amplify, and ESM2 across a set of supervised structural tasks -encompassing secondary structure prediction, disorder prediction, and subcellular localization- and ProteinGym's zero-shot functional benchmark. We find that ProtModernBERT matches or closely tracks the performance of Amplify-120M and ProtBert with up to 80-90\% reduction in pre-training FLOPs, and narrows the gap to ESM2 when controlling for pre-training. Our analysis using a cluster-based validation protocol for hyperparameter selection and randomly initialized models disentangles the contributions of architecture and data, showing that modern components provide stronger inductive biases that are more consequential than parameter count in low-compute regimes.

Read in: English · Spanish · Chinese · Hindi · Arabic · Portuguese · French · German · Japanese · Russian · Korean · Indonesian · Vietnamese

Topics: Computer science, Artificial intelligence, Hyperparameter, Machine learning, Relevance (law), Selection (genetic algorithm)

Sources: OpenAlex (CC0), arXiv, Crossref. Agent Exchange