Predicting Mutation-Induced Alterations in Protein Solubility Using an Attention-Guided Multimodal Learning Method
Protein solubility plays a crucial role in recombinant protein production and protein developability. However, the vast mutational space makes exhaustive experimental screening costly and time-consuming, motivating computational approaches to prioritize candidate mutations. Predicting mutation-induced solubility changes remains challenging because mutation effects are associated with both the sequence context and the local three-dimensional physicochemical environment. Although multimodal models can integrate complementary sequence- and structure-related representations, the predictive contribution and biological meaning of individual architectural components cannot be inferred directly from the learned fusion behavior. Herein, we developed AttenSol, an attention-guided multimodal framework that integrates mutation-sensitive sequence representations with topology-informed graph processing through a learned fusion mechanism. On the held-out benchmark, AttenSol achieved a balanced accuracy of 52.8%. Control experiments showed that AttenSol did not exhibit a statistically significant balanced accuracy (BACC) advantage over the evaluated direct-fusion and topology-informed GCN comparators, while the learned fusion gate showed a strong but nearly constant preference for the graph-processed branch. Fixed-gate and topology-perturbation analyses indicated that this weighting should not be interpreted as evidence of structural information dominance or strong sample-adaptive routing. Forward–reverse analysis showed strong directional consistency when reverse-mutation augmentation was used, whereas matched retraining without this augmentation reversed the correlation pattern and substantially increased the zero-sum error. Cross-data set evaluation on SoluProtMutDB yielded a BACC of 0.397, revealing a substantial performance reduction under data set and domain shift. Together, these results position AttenSol as a multimodal framework for mutation effect prediction and computational prioritization while emphasizing the need for explicit control analyses when interpreting learned fusion behavior.
Authors
- Fufeng Liu (ORCID: https://orcid.org/0000-0002-9454-2382)
- Fuping Lu (ORCID: https://orcid.org/0000-0002-0610-8827)
- Xinming Wu
- Qingxu Deng (ORCID: https://orcid.org/0009-0003-8403-9462)
- Yao Song
Institutions
- Tianjin University of Science and Technology (CN)
Publication Details
- Journal
- Journal of Chemical Information and Modeling
- Published
- 2026-11-18
- DOI
- https://doi.org/10.1021/acs.jcim.6c03622
- Primary Topic
- Machine Learning in Bioinformatics
- Type
- article
- Field-Weighted Citation Impact
- 0.00
Funders
- National Natural Science Foundation of China
- Tianjin Municipal Science and Technology Bureau
- National Key Research and Development Program of China