Abstract
Protein aggregation is an obstacle for engineering effective recombinant proteins for biotechnology and therapeutic applications. Predicting protein aggregation propensity remains challenging due to the complex interplay of sequence, structure, environmental factors, and external stress conditions, particularly for globular proteins. To understand the determinants of aggregation and improve its prediction, we quantified insoluble aggregation following high temperature and acidic stress in custom libraries of small protein domains (40-72 amino acids) using a high-throughput, in vitro, mass spectrometry-based method. In total, we quantified aggregation for 18,987 small protein domains, revealing diverse stress-dependent aggregation phenotypes that were consistent in different library contexts. We also found that aggregation measurements on individually purified proteins strongly correlated with high-throughput mixed-pool data (Pearson's r = 0.65-0.79), supporting the use of multiplexed approaches to study aggregation. Using machine learning, we identified sequence and structural features that correlate with aggregation and fine-tuned the protein language model SaProt, which explained 43-55% of the observed variation in a held-out test set of unrelated protein domains. Our model shows promising utility for engineering aggregation-resistant proteins, and our dataset serves as an important resource for developing improved models of protein aggregation.