Research
Transcriptomics and machine learning for autoimmune disease, cancer expression and survival analysis, and computational drug safety.
M.Tech thesis, IIT Bombay
RNA-seq analysis and AI models for diagnosis of autoimmune diseases
Supervisor: Prof. Prakriti Tayalia
- Problem
- Autoimmune diseases share overlapping symptoms and immune signatures, which complicates early and specific diagnosis. The thesis asked whether blood transcriptomes carry disease-specific signals that a classifier can learn.
- Data
Public bulk RNA-seq from peripheral blood mononuclear cells (PBMCs).
Samples after quality control| Group | Samples after QC |
|---|
| Ankylosing spondylitis (AS) | 23 |
| Rheumatoid arthritis (RA) | 63 |
| Systemic lupus erythematosus (SLE) | 173 |
| Type 1 diabetes (T1D) | 137 |
| Healthy | 43 |
- Methods
Pipeline: Trimmomatic, HISAT2 (hg38), samtools, featureCounts on the IIT Bombay Spacetime HPC cluster (Linux).
- DESeq2 normalization and differential expression
- Gene set enrichment analysis (GSEA)
- Protein-protein interaction and gene regulatory networks
- Principal component analysis (PCA)
- Classifiers: Logistic regression, SVM, Ensemble model, ANN. Evaluated on an external test set.
- Results
Best AUC by disease (ANN):
AS 0.75RA 0.97SLE 0.80T1D 0.80
Manuscript
An integrative transcriptomic approach to identify shared and disease-specific molecular signatures and predictive gene signatures across autoimmune diseases
Submitted 5th author
Builds on the M.Tech thesis above.
MTP-1, IIT Bombay, Oct 2022
LMNA (Lamin A/C) and SUN2 expression in TCGA breast cancer
- Problem
- Nuclear envelope proteins Lamin A/C and SUN2 are linked to nuclear mechanics. The project asked how their expression relates to clinical features and survival in breast cancer.
- Data
- TCGA breast cancer cohort (public).
- Methods
- Expression analysis of LMNA and SUN2
- Survival cut-off analysis
- Comparison across nodal stage
- Comparison across molecular subtypes
B.Tech thesis, NIT Rourkela, May 2021
2D QSAR classification model for hERG channel inhibition
Supervisor: Prof. J Sivaraman
- Problem
- Blocking the hERG potassium channel can cause fatal cardiac arrhythmia, and it is a common reason drug candidates fail. The thesis built a screen to flag likely hERG blockers early.
- Data
- Compound structures with hERG inhibition labels, encoded as 2D molecular descriptors.
- Methods
- 2D molecular descriptors
- QSAR classification model
- Model selection prioritizing sensitivity and AUC over accuracy
- Rationale
- A missed blocker (false negative) carries a cardiotoxicity risk downstream, so sensitivity and AUC were prioritized over raw accuracy.