Predicting Forward and Reverse Mutations from Protein Sequences Using Embeddings and Deep Learning

Forward mutations alter DNA, potentially causing diseases, drug resistance, or driving evolution, while reverse mutations restore normal gene function. Predicting these mutations is essential for understanding genetic disorders and developing effective therapies. This study introduces a novel deep learning approach to classify point mutations as forward or reverse using protein embeddings generated by the ESM-2 protein language model. These embeddings were processed using a Bi-directional Long Short-Term Memory (Bi-LSTM) network with an attention mechanism, leveraging both the representational power of ESM-2 (Evolutionary Scale Modeling 2) and the sequential modeling capabilities of Bi-LSTM. The classifier was trained and tested using an 80-20 data split, with performance consistency ensured by employing 10 different random seeds. Our analysis identified a 9-amino-acid window as the optimal size, achieving an accuracy of 0.89 ± 0.00, precision of 0.88 ± 0.09, recall of 0.89 ± 0.09, F1-score of 0.89 ± 0.09, AUC-ROC of 0.95 ± 0.05, and specificity of 0.88 ± 0.09 on the test dataset. Further investigation demonstrated that altering the sequence order of wild-type and mutated proteins did not significantly impact classification performance, likely due to the bidirectional nature of the ESM-2 model. Additionally, the classifier was validated on a larger independent dataset (S9028) containing 9,028 single mutations across 153 proteins, where it maintained robust performance with an AUC-ROC of 0.88 ± 0.04, an F1-score of 0.80 ± 0.05, and an accuracy of 0.82 ± 0.04. These results confirm the effectiveness of our approach and highlight the robustness of the model in mutation classification across diverse datasets.