ProtSATT: An Advanced Protein Solubility Predictor Based on Attention Mechanism

Protein solubility is an important factor affecting the production efficiency and downstream utility in biologics development. However, conventional expression systems often suffer from insoluble aggregation, and current optimization strategies still rely heavily on time-consuming empirical screening. Existing computational approaches do not always make full use of the complementary information captured by modern protein language models (PLMs), which limits their performance and practical utility. To address this issue, we developed ProtSATT (Protein Solubility Attention Network), a sequence-based computational framework for solubility-related prediction. ProtSATT integrates sequence-level embeddings from three PLMs, namely, UniRep, ESM-2, and ProtT5, and applies attention-based feature extraction and fusion in the latent space of these embeddings.This design enables the model to learn interactions among complementary PLM-derived representations for the downstream regression and classification tasks. Across three public benchmarks covering solubility regression and expression-related classification (eSOL, S. cerevisiae, and E. coli), ProtSATT achieved competitive and, in several cases, improved performance relative to existing methods. On the eSOL data set, ProtSATT achieved an R² of 0.5450 and an accuracy of 81.21%. On the external S. cerevisiae benchmark, it reached an accuracy of 83.33% without additional fine-tuning. On the homology-aware E. coli benchmark, ProtSATT also showed competitive performance in expression-related classification, achieving an average accuracy of 72.21% in the binary setting. After PLM embeddings were precomputed, the downstream ProtSATT predictor contained only 6.0 M trainable parameters and processed more than 11,400 sequences per second on a single GPU. These results suggest that integrating multiple PLM-derived representations with attention-based downstream modeling can provide useful computational support for protein engineering and related applications. All source code and pretrained models are available at https://github.com/quietbamboo/ProtSATT.