Abstract:
A protein is a large, complex macromolecule, and it has many crucial roles in the human body as it performs most of the work in cells and tissues. It consists of one or multiple extended sequences of amino acid components. Another important biomolecule that comes after DNA and proteins is carbohydrates. Carbohydrates interact with proteins to facilitate various biological processes. Sev- eral biochemical experiments exist to study protein-carbohydrate interactions, but they are expensive, time-consuming, and challenging. As a result of the swift advancements in sequencing technologies, the quantity of recognized protein sequences has surged exponentially. Therefore, developing a com- putational technique from known protein sequences for effectively predicting protein-carbohydrate binding interactions has led to the emergence of a prominent new area of study.
Most of the computational approaches for protein-carbohydrate binding sites prediction are biased towards the negative class. This is due to the fact that the count of carbohydrate-binding residues is considerably lower compared to non-carbohydrate-binding residues in the benchmark datasets. In this thesis, we introduce a proficient ensemble machine learning model called ‘StackCBEmbed’ for the accurate classification of protein-carbohydrate binding interactions at the residue level within established protein sequences. StackCBEmbed demonstrates a more balanced behavior compared to the state-of-the-art methods in terms of accurately predicting both the positive and negative data points.
Our research used a benchmark training dataset and two separate independent test sets. Through the use of the Incremental Feature Selection method, we identified crucial sequence-based features and picked the most impactful ones. Furthermore, we integrated embedding characteristics from a pre-trained transformer-based language model known as ‘ProtT5-XL-Uniref50.’ To the best of our knowledge, this is the initial endeavor to utilize a protein language model for predicting protein- carbohydrate binding interactions. StackCBEmbed achieved sensitivity, specificity, and balanced accuracy scores of 0.691, 0.849, 0.769 and 0.627, 0.835, 0.731 in the two independent test sets respec- tively. Compared to the earlier prediction models that were benchmarked in the same datasets, our reported results are significantly superior. Thus, we hope the StackCBEmbed will help discover novel protein-carbohydrate interactions and advance the related research fields. StackCBEmbed is freely available as python scripts at https://github.com/farah5112github/StackCBEmbed.
