An Effective Framework for the Prediction of Protein Folds using Natural Language Processing and Evolutionary Features

  • Implemented advanced NLP techniques to improve protein fold recognition for low similarity datasets such as DD, EDD, TG, and SCOPe baseline datasets encompassing diverse amino acid-based protein sequences and their corresponding folds.
  • Extracted features by utilizing evolutionary PSSM and HMM profiles of protein sequences, and concatenating them with global Convolutional and Skip Bi-gram features.
  • Implemented BERT and ESM by Meta transformer-based models for classification and achieving an impressive accuracy exceeding 93% across all datasets, surpassing the previous 85% accuracy.
  • The project has been documented and has been sent to IEEE Transaction in Computational Biology and Bioinformatics where it is currently under review.

SQR: High Performance Secure QR System to Filter Against Malicious QR Codes Using Deep Learning Techniques

  • Developed a high-performance system for detecting malicious and benign QR codes.
  • The models used in the system are Vision Transformer (ViT), VGG16 and EfficientNet.
  • Paper is currently under review in Expert Systems with Applications Journal.
  • Preprint available here.

Interestingness from COVID-19 Data: Ontology and Transformer-Based Methods

  • Identify interesting patterns using ontology-based mining techniques and process them with transformer models for identifying the interesting rules from the mined corpora.
  • Paper is accepted for publication at ICON 2022, proceedings in ACL Anthology.
  • Preprint available here.

Ontology-Based Semantic Data Interestingness Using BERT Models

  • An effective data preprocessing technique that introduces semantics at the level of data curation.
  • Semantic Interestingness Framework (SIF) for COVID-19 Data.
  • Enhanced apriori approach with constraints (ConstApriori algorithm), which employs interestingness measures for semantic facts extracted from RDF data.
  • Implementation of Clinical BERT and Bio BERT model for identifying the most interesting rules using a cosine similarity measure - Semantic Interestingness.
  • Paper is under review in Connection Science Journal.
  • Preprint available here.

Affect Aware Tutoring System in E-Learning systems

  • Worked on a novel idea in the field of education technology, named ”Affect Aware Tutoring System Using Video Bots”
  • Built a learning management system that collects the click-stream log data of the student, simultaneously captures video and then predicts the user’s affect state for realtime feedback.
  • Developed an optimized transformer-based deep learning model Vision Transformers to predict the user’s affect state. The model was trained on the huge DAiSEE dataset for approximately 300 hours.
  • Paper documented and to be submitted in IEEE Transactions in Learning Technologies.