Predicting Protein Stability of Single-Point Mutations using Ensemble Learning and ESM Embeddings

Predicting Protein Stability of Single-Point Mutations using Ensemble Learning and ESM Embeddings

Computer Science

Victoria Krushovska

2025

Within a biological system, proteins are essential; they are the building blocks of life. Their functions range from performing as antibodies to catalysing reactions as enzymes; so, proteins have an extremely varied and extensive use through a variety of industries. With mutations, their stability and functionality can be modified, with even single-point mutations having a strong effect. Therefore, being able to predict the effect in stability changes of these mutations is essential, but often computationally costly and weak when encountering unseen data. With the rise in popularity of protein language models (PLMs), including ESM, there have been improvements, but these models still encounter the same difficulties. This study explores the effect of ensemble learning on improving protein stability predictions. This is done by combining two base models: a convolutional neural network (CNN) using wild-type and mutant sequences, and a neural network trained on frozen ESM embeddings. They each form two base-learners, where each is trained on half the base training data. To combine their predictions, intermediate meta-learners for each model type and a final meta-learner are used. Results show that ensemble learning improves predictions over base models, showing improvements in all performance metrics used. Notably, even the weaker CNN-based learners contributed to improved ensemble performance, reinforcing the benefit of this type of architecture. Although it is necessary to further optimise hyperparameters and include stronger base models in the future, these results indicate that ensemble learning can lead to more accurate protein stability prediction with minimal additional computational cost.

protein stability predictions, mutation effect prediction, ensemble learning