Publications
Real PMIDs, Wrong Papers: Harness Design for Biomedical Claim-Verification Agents
Agentic AI for Biological Discovery (AgenticLS)
Mohammed Saqlain
A controlled ablation of harness design for biomedical agents on SciFact, with live PubMed as the only tool. Holding the backbone fixed, we compare closed-book answering, single-shot retrieval, a ReAct-style agent that writes its own queries, and an oracle given the gold abstracts. Query formulation triples gold recall, but a 17–24 point gap to the oracle shows retrieval is the main bottleneck. A citation audit finds closed-book models cite PubMed IDs that almost always exist yet are almost never relevant.
HyperGNNs for Multi-Modal Classification and Severity Analysis of Neurodegenerative Disorders
K. Bhavish Raju, K. Musadiq Pasha, Mohammed Saqlain, Nishaan Padanthaya, Jayashree R
Proposed a Hypergraph Neural Network (HyperGNN) framework for multi-modal classification and severity assessment of neurodegenerative disorders like Alzheimer’s. By integrating MRI imaging and clinical data through hypergraph representations, the model improves diagnostic accuracy across CN, MCI, and AD stages, incorporating SHAP and Grad-CAM for clinical interpretability.
BiasNet: A Contrastive GNN Based Framework for Classifying Political Stance in News
T. Bhat, Mohammed Saqlain, A. Muralidharan, B. Das
We introduce BiasNet, a framework using Graph Neural Networks and contrastive learning to classify political stance in news articles. By modeling relationships between articles, our approach effectively distinguishes left, right, and neutral biases, outperforming traditional baselines. This method supports media monitoring and fact-checking by detecting subtle bias in news content.
Reading Between the Lines: LLM-Powered Topic Modelling and Graph-Based Insights from Research Abstracts
Mohammed Saqlain, Nishaan Padanthaya, Fiza Haneen, Bhaskarjyoti Das
This study presents an innovative approach to research abstract analysis by combining advanced topic modeling with large language models (LLMs) and graph-based analytics. We employ BERTopic for topic extraction, LLama2 for semantic enhancement, and graph neural networks to uncover hidden patterns and relationships within academic literature. Our methodology demonstrates significant improvements in identifying emerging research trends, cross-disciplinary connections, and knowledge evolution patterns compared to traditional approaches.
Comparative Analysis of Traffic Accident Detection with Emphasis on Explainability of DL Models
Mohammed Saqlain, Nishaan Padanthaya, K. Musadiq Pasha, J. O. Shivnesh, Dhriti Rajesh Krishnan, K. S. Srinivas
This paper investigates traffic accident detection using deep learning models with a focus on explainability using SHAP. The study demonstrates how explainable AI makes traffic monitoring systems more trustworthy and interpretable for real-world applications.