LoMuS: Low-Rank Adaptation with Sequence Multi-representation Improves Protein Stability Prediction.

Area of Science:

• Computational Biology
• Biophysics
• Machine Learning in Biology

Background:

• Protein folding stability is vital for protein function, dynamics, and engineering.
• Accurate prediction of protein stability is challenging, especially with limited structural data.
• Existing methods struggle with data variability and sequence-only predictions.

Purpose of the Study:

• Introduce LoMuS, a novel deep learning model for predicting protein stability directly from primary sequences.
• Integrate physicochemical descriptors with sequence-derived embeddings for enhanced prediction accuracy.
• Improve protein engineering by enabling better prediction and ranking of stability scores.

Main Methods:

• Developed LoMuS, a multi-representation deep learning model.
• Fused explicit physicochemical descriptors with low-rank adapted protein language model embeddings.
• Evaluated model performance across diverse settings including absolute scoring, mutation landscapes, and out-of-distribution data.

Main Results:

• LoMuS consistently outperforms sequence-only baselines across multiple benchmarks.
• Achieved at least a 10% absolute performance gain in Spearman's rank correlation.
• Ablation studies confirmed the critical contribution of both physicochemical descriptors and sequence embeddings.

Conclusions:

• LoMuS advances the prediction of protein stability using a multi-representation deep learning approach.
• The model's ability to leverage sequence information enhances protein engineering applications.
• Open-source availability of code facilitates further research and development.