ToxTrans: A Deep Learning Method to Predict Protein Toxicity
Published in B.Sc. Thesis, Shahid Beheshti University (under review as a preprint), 2023
Extracted embeddings from protein sequences using protein language models (ProtT5), selected top features via ANOVA F-value, and classified using a tuned, GPU-accelerated LightGBM. Trained and evaluated on 8,233 protein sequences from UniProt, achieving 93.76% accuracy, 0.98 AUC, and 0.87 MCC on independent testing — outperforming all prior methods, including ToxMVA (+2.16% accuracy, +0.04 MCC).
Awarded Best B.Sc. Thesis Project by Shahid Beheshti University’s Computer Science and Engineering Faculty.
Recommended citation: Zahra Khotanlou. "ToxTrans: A Deep Learning Method to Predict Protein Toxicity." B.Sc. Thesis, Shahid Beheshti University. Supervisor: Dr. Hamed Malek.
