RNA Therapeutic-Design Platform
Mechanism Arbitration and Candidate Generation Across Nine Therapeutic Goals
| Year | 2026 |
|---|---|
| Location | NCBS |
| Tags | ASO, siRNA, SpliceAI, ViennaRNA |
A single-pass pipeline arbitrating across 41 RNA therapeutic mechanisms & 9 goals, with 19 mechanism-specific features (SpliceAI, RNAplfold, miRNA, RBP sites), benchmarked on 165,237 rows and 7 ML models across 12 experiments.
Jun 2026 – Sep 2026 at NCBS, Bangalore, supervised by Dr. Tanay Bhatt.
- 01Mechanism arbitration
A single-pass pipeline that, for a therapeutic goal, weighs 41 RNA therapeutic mechanisms across 9 goals and decides which ones apply before designing anything.
Python, Mechanism-arbitration API - 02Target & transcript validation
Resolved genes and transcripts per organism from reference annotation, with provenance tracking, confidence caps and explicit Unresolved / Not-applicable states so the pipeline never over-claims.
Ensembl, NCBI, Genome-scale reference data - 03Mechanism-specific features
Built 19 features covering splice prediction, RNA accessibility, miRNA targeting, promoter accessibility, RNA-binding protein sites, alternative polyadenylation, promoter switching and intron retention, then scored and ranked candidate oligos.
SpliceAI, ViennaRNA RNAplfold, miRNA targeting, RBP sites - 04Benchmarking
Analysed 165,237 benchmark rows across RNase H (159,109), splice-switching (2,181) and siRNA (3,947) modalities to compare mechanism-specific performance.
Pandas, Benchmark datasets - 05Machine-learning experiments
Evaluated seven ML models across 12 controlled experiments: class imbalance, gene-grouped splitting, ranking, cross-chemistry transfer, feature attribution, thermodynamics, RNA-FM embeddings and ML versus heuristics.
Scikit-learn, RNA-FM, Feature attribution - 06Testing
Validated the API and workflow with 84 mechanism-arbitration tests and 11 feature-level tests, including regression tests for defects found along the way.
pytest, Regression testing