Syed Salman Ali Zaidi

LLM-Powered Protein Classification: A Deep Learning Framework for Kinases, Phosphatases, and Transmembrane Proteins

Accurate functional classification of proteins is essential in biomedicine and therapeutic development. Traditional computational approaches often face challenges in scalability and identifying novel sequences. This study evaluates the performance of several pre-trained Large Language Models (LLMs) on two key classification tasks: kinase/phosphatase differentiation and transmembrane protein identification. We fine-tuned and assessed CAFA-5, ChatterjeLab/PepMLM, and ChatterjeLab/MeMDLM using UniProtKB for kinase/phosphatase data and ActTRANS for transmembrane proteins. In contrast to earlier studies such as ActTRANS which focused on binary classification using frozen embeddings and SVMs. The proposed approach employs an end-to-end neural classifier in a broader multi-class setting, demonstrating greater generalizability across protein families. The integration of LLMs with protein sequence analysis offers a scalable and accurate solution that can accelerate drug discovery and optimize therapeutic strategies. CAFA-5 consistently outperformed other models, ranking first in both tasks. It achieved 90% accuracy, 89.74% balanced accuracy, and an Matthews correlation coefficient (MCC) of 0.7995 in kinase/phosphatase classification, and 98.9% accuracy with a perfect MCC of 1.0 for transmembrane proteins. While PepMLM also performed well (91.89% accuracy, MCC: 0.84), MeMDLM scored lower (80.53% accuracy, MCC: 0.63). These results highlight the capability of LLMs to detect complex sequence patterns and outperform traditional methods. This work not only advances computational biology but also contributes to precision medicine, particularly in areas like oncology, where understanding kinase-phosphatase dynamics is critical.