VISH-Pred: Protein Toxicity Prediction
Revolutionary ensemble framework using fine-tuned ESM2 transformer models for accurate protein toxicity prediction. Achieves Matthews correlation coefficient of 0.737 and F1-score of 0.759, outperforming existing methods by over 10%.
Key Features:
- Handles massive class-imbalance in toxicity data
- Uses protein sequence as sole input
- Employs LightGBM and XGBoost techniques
- Available as easy-to-use web server
DeepRepurpose: COVID-19 Drug Discovery
Framework Performance
0.916
Mean Pearson Correlation
0.840
Mean R²
0.313
Mean Squared Error
Identified Compounds
Discovered 47 potential compounds targeting SARS-COV-2 proteins including 21 antivirals, 15 anticancer drugs, 5 antibiotics, and 6 investigational compounds.
Molecular docking simulations confirmed low binding energies and high binding affinity for virus proteins.
DeepSol: Protein Solubility Prediction
Novel deep learning framework using convolutional neural networks for sequence-based protein solubility predictions, crucial for pharmaceutical research and production yield.
1
Input Processing
Exploits frequent k-mer and additional sequence/structural features from protein sequences.
2
CNN Analysis
Convolutional neural network backbone processes complex sequence patterns.
3
Superior Results
Achieved 0.77 accuracy and 0.55 Matthew's correlation coefficient, outperforming all existing methods.
Master Regulator Identification
Developed consensus framework using RGBM and ARACNE techniques to identify transcription regulators associated with immune-silent cancer phenotypes.
Key Findings:
- TGFB1I1 emerged as main negative immune modulator
- Enriched pathways: NOTCH1, TGF-β, Interleukin-1, TNF-α
- Validated across TCGA and PRECOG datasets
RGBM: Gene Regulatory Networks
Regularized Gradient Boosting Machines framework for identifying transcriptional regulators of discrete glioma subtypes, converting network inference into feature selection.
1
Feature Selection
Converts GRN inference into feature selection problem handling diverse data sources.
2
L-curve Regularization
Tikonov regularization with optimal L-curve criterion determines optimal TF sets.
3
Superior Performance
Outperformed DREAM challenge winners and established methods on synthetic and real datasets.
Network Evolution & Visualization Tools
Differential Community Detection
Three-stage approach identifying differential sub-networks in paired biological networks through community detection in denoised topological graphs.
a) Control Differential Sub-network for Ovarian Cancer
b) Case Differential Sub-network for Ovarian Cancer
Netgram Visualization
Line-based visualization tool tracking community evolution in dynamic networks, maintaining aesthetic guidelines with minimal crossovers.
MHKSC Clustering
Multilevel Hierarchical Kernel Spectral Clustering exploiting eigenspace projections for large-scale network analysis.
Research Impact & Future Directions
Publications & Tools
Extensive portfolio of peer-reviewed publications in top-tier journals including Nature Medicine, Nature, Cell, Science Immunology, Bioinformatics, Nature Communications, and BMC Systems Biology.
Open-source tools and web servers available for global research community, with code repositories on GitHub and CRAN.
Key Achievements:
- 15+ major software frameworks developed
- Multiple web servers deployed for public use
- Datasets publicly available on Mendeley
"Bridging artificial intelligence and biomedicine to unlock new therapeutic possibilities and improve patient outcomes through innovative computational approaches."