Pralay Mitra

This study introduces RSMBind, a deep learning model for predicting RNA-small molecule binding sites, achieving high accuracy on benchmark datasets and demonstrating the utility of RNA language models in the absence of structural data.

The paper presents a framework using BERT embeddings and WGAN-GP to predict pathway associations for human dark kinase proteins, achieving high accuracy in classification and pathway clustering.

MOSANIC learns a spatial cell-cell communication network from transcriptomics data to identify network hubs and predict vulnerabilities in breast cancer.

The paper evaluates host-specific linear B-cell epitope prediction models across human, mouse, and non-human primate datasets, showing significant performance differences and highlighting the need for host-aware approaches.

This paper introduces a resource-efficient pipeline for B-cell epitope prediction using a frozen ESM-2 model and genetic algorithm compression, achieving strong performance on an IEDB benchmark.

GraphTox predicts peptide toxicity using a geometric graph transformer and LORA-based finetuning, outperforming existing methods by incorporating structural information.

The paper introduces an AI framework for protein interaction design using KAN-based PU learning, achieving improved stability and validation metrics compared to traditional methods.

This paper identifies a miRNA signature for lung adenocarcinoma using machine learning on TCGA data, achieving high accuracy and stability.

The paper develops a transformer-based method to predict 5' cap sites from DNA sequences, achieving high accuracy and identifying biologically relevant motifs.

DeepPROTECTNeo predicts immunogenic neoepitopes by integrating genomic data and TCR-epitope interactions, outperforming existing methods with high accuracy and interpretability.

The paper introduces E(Q)AGNN-PPIS, an equivariant graph neural network with attention for predicting protein-protein interaction sites, achieving improved performance on benchmark datasets.

rpcFold predicts RNA secondary structures using a residual parallel convolutional neural network, achieving improved performance on nested and pseudoknot base-pairs.

The paper develops a 20-gene molecular prognostic score (mPS) for high-grade serous ovarian cancer (HGSOC) that outperforms conventional clinical parameters in predicting 5-year survival.

The paper introduces ProFuMCell and ProModb, web services that identify and analyze spatially localized protein modules in cells using interaction networks and KEGG pathway data.

The paper introduces PINDeL, a GCN-based model for identifying disease-associated human proteins, achieving high accuracy and outperforming existing methods with experimentally verified results.

This paper introduces a method using Protein Locality Graph topology to infer species trees by analyzing inter-module interactions in protein networks.

The paper introduces MAT, a toolkit for parsing and analyzing macromolecular data files, offering faster and more accurate parsing with new data structures and algorithms.