Scalable embedding fusion with protein language models: insights from benchmarking text-integrated representations

Protein language models (pLMs) have become essential tools in computational biology, powering diverse applications from variant effect prediction to protein engineering. Central to their success is the use of pretrained embeddings-contextualized representations of amino acid sequences-which enable effective transfer learning, especially in data-scarce settings. However, recent studies have revealed that standard masked language modeling objectives used to train these models often produce representations that are misaligned with the needs of downstream tasks. While scaling up model size improves performance in some cases, it does not universally yield better representations. In this study, we investigate two complementary strategies for improving pLM representations: (i) integrating text annotations through contrastive learning, and (ii) combining multiple embeddings via embedding fusion. We benchmark six text-integrated pLMs (tpLMs) and three large-scale pLMs across six biologically diverse tasks, showing that no single model dominates across settings. Fusion of multiple tpLMs embeddings improves performance on most tasks but presents a computational bottleneck due to the combinatorial number of possible combinations. To overcome this, we propose greedier forward selection, a linear-time algorithm that efficiently identifies near-optimal embedding subsets. We validate its utility through two case studies, homologous sequence recovery and protein-protein interaction prediction, demonstrating new state-of-the-art results in both. Our work highlights embedding fusion as a practical and scalable strategy for improving protein representations.