Building from the ground up: augmentation strategies to improve protein language model performance on protein sets

An important underlying requirement for deep learning models is the availability of large datasets, however, in sciences such as chemistry and biology, generating experimental data of sufficient size represents a significant bottleneck. In this work, we probed the extent and nature of limited input data, characterize how several protein language models (pLMs) learned on suboptimal datasets, and investigated augmentation approaches to improve predictive model performance. We found that narrowing the distribution of protein-protein complexes in a limited dataset permits an order-ofmagnitude decrease in training dataset size, if a model is only expected to generalize to in-distribution data. Similar results were found in active learning (AL), where both inand out-of-distribution Matthews Correlation Coefficient (MCC) showed a small but consistent increase upon augmentation of the training set with complexes produced by ESM-3, a generative pLM.