Supervisor: Prof. Md. Shamsuzzoha Bayzid (BUET)

Keywords: Protein-Protein Interaction, 3D Representation Learning, Geometric Deep Learning, Graph Neural Network

• We developed EGRET, a novel method for effective protein 3D representation Learning that outperforms the previous state-of-the-art methods for protein-protein interaction site (PPIS) prediction from unbound proteins, a crucial research problem in drugs and vaccines design. (first authored publication in Briefings in Bioinformatics [doi] [github])

• Currently we are developing a Siamese architecture (with EGRET backbone) to address the pairwise-PPISP problem, where the unbound structures of both partner proteins are known. To improve 3D molecular representation learning, we are exploring the potential of Hypergraph Convolutional Networks and their variants. In this project, I am mentoring one undergraduate student.

• In these projects, I took the leading role by designing and developing the methods, running experiments, and contributing to writing the paper for EGRET.

Supervisor: Prof. Md. Shamsuzzoha Bayzid (BUET)

Keywords: Computational Phylogenomics, Data Imputation, Self-supervised Learning

• We developed an automated and specially tailored unsupervised deep learning technique (named QT-GILD), accompanied by cues from natural language processing (NLP), which learns the data distribution in a given set of incomplete gene trees and effectively imputes the dataset by generating a complete set of quartets accordingly. QT-Gild led to two first co-authored publications in RECOMB 2022 conference [doi] [github] and Journal of Computational Biology [doi].

• In this project, I took the lead by ideating and developing the framework, contributing to the experiments and writing the paper.

Supervisors: Prof. Md. Shamsuzzoha Bayzid (BUET), Prof. Clara Li (Icahn School of Med),  and Dr. Yuichiro Miyaoka (TMIMS)

Keywords: Alzheimer’s Disease, Genomics, Cognitive data, Multimodal Machine Learning, Interpretability

• In this project, we are working on developing a machine learning (ML) method for Alzheimer’s Disease prediction and analysis. We are experimenting with both genomic and cognitive features.  

• We propose using PRS as features, which obtained 72-74% accuracy, resulting in around 17% improvement over the baseline ML models using cognitive features only. Apart from these, we leveraged the widely used gradient-based algorithm for interpretability study -- Grad-Shap -- for interpreting the decision process of our neural network models.

• I contributed to this project by developing an interpretable multi-modal ML framework and running the necessary experiments. I am currently voluntarily mentoring two undergraduate students so that they can take the lead in future endeavors of this project.

Supervisors: Prof. Md. Shamsuzzoha Bayzid (BUET), Prof. Mohammad Saifur Rahman (BUET)

Keywords: Protein Secondary Structure, Deep Learning, Self-Attention

• We built a highly accurate method for Q8 protein structure prediction (SAINT), which incorporates self-attention mechanism (a concept from natural language processing) with the Deep Inception-Inside-Inception network in order to effectively capture both the short- and long-range interactions among the amino acid residues. SAINT offers a more interpretable framework than the typical black-box deep neural network methods.

• Here, I came up with the design of the framework of SAINT, and led the project by developing the method, contributing in running experiments and writing the paper. (published in Bioinformatics [doi] [github])

• Inspired by the success of SAINT, we also developed a protein torsion angle prediction framework, named SAINT-Angle, that achieved state-of-the-art accuracy for this problem. I am also mentoring two undergraduate students in this project. (published in Bioinformatics Advances [doi] [github])

Supervisor: Prof. Ehsan Hoque (University of Rochester)

Keywords: Interpretable Deep Learning, Fairness in AI, Multimodal Learning

• We built an ensemble model—by combining Multiple-Instance-Learning and Language-Modeling based models—that can predict whether an interviewee should be hired or not. Using both model-specific and model-agnostic interpretation techniques, we can decipher the most informative time-segments and features driving the model’s decision making. (first co-authored publication in ACII Workshop on Applied Multimodal Affect Recognition 2021 [doi])

Supervisor: Prof. Anindya Iqbal (BUET)

Keywords: Deep Learning, Natural Language Processing, Automated Code Repair, Code Review

• In this project, we leveraged a state-of-the-art text-summarization method and code reviews to develop an automated tool for code-repairing. We also built a large dataset with 55,060 code reviews and associated code changes. (co-authored publication in Information and Software Technology journal [doi] [github])

Supervisor: Prof. A. B. M. Alim Al Islam (BUET)

Keywords: Natural Language Processing, Textual Abuse, Automated System

• In this project, we built an automated system for detecting and preventing textual abuse in emails. (co-authored publication in NSysS 2021 conference)