01
Computational Biology / Biotechnology
AI-Driven Optimization of CRISPR Delivery Systems
Developed through the Lumiere Research Scholar Program.
A machine-learning framework that scores and prioritises guide RNA and nanocarrier pairings for targeted gene-editing delivery.
- CRISPR
- Machine Learning
- Nanomedicine
Overview
I built a computational framework that treats CRISPR guide selection and kidney-targeted delivery as one decision problem. Using 3 PKD1 guides and 73 literature-curated nanocarriers, the system evaluated 219 possible guide–carrier combinations and prioritized the most experimentally promising pairs for follow-up testing.
Research Problem
CRISPR guides and delivery systems are usually optimized separately, even though both must succeed for gene editing to work. My project closes that gap by jointly ranking guide safety and renal-targeting evidence, creating a more practical and reproducible way to decide which combinations deserve laboratory validation first.
Motivation
Kidney delivery is especially difficult because many nanoparticles naturally accumulate in the liver, while PKD primarily affects renal tubular cells. I expanded the model to include tubular-targeting systems so the ranked candidates would better reflect the actual biological compartments relevant to PKD.
Methodology
I compared logistic regression, XGBoost, ExtraTrees, and random forest models, then selected the strongest overall performer. To improve reliability, I used repeated grouped cross-validation, leakage-safe feature selection, threshold optimization, and 2,000-resample bootstrap confidence intervals rather than relying on a single accuracy score.
Computational Workflow
The pipeline first scored PKD1 guides for off-target risk using CRISPOR, then predicted renal tropism for 73 nanocarriers. Only carriers exceeding the learned threshold entered the final ranking, where guide safety and delivery performance were combined and tested across multiple weighting schemes.
Results & Outputs
The final model achieved AUROC 0.856, balanced accuracy 0.843, and 100% recall on primary positives. It narrowed 219 combinations to 99 threshold-eligible candidates, with PKD1 404 fw + TRX6 C6 LP remaining the top-ranked pair across all four weighting schemes.
Limitations
The framework is constrained by the coverage and heterogeneity of publicly available delivery data, and its predictions are computational hypotheses that still require experimental validation.
What I Learned
This project taught me that a dependable biomedical model is defined not just by performance, but by how carefully it is validated. I learned to test for leakage, class imbalance, threshold sensitivity, uncertainty, reproducibility, and failure cases—and to distinguish computational prioritization from true therapeutic validation.
Recognition and related iterations
- IRIS National Fair Finalist, 2026
- Padma Shri Dr. B. S. Chaubey Memorial Trust Special Award, 2026
- CREST Gold Award, 2026 — later iteration: Thresholded ML Framework Prioritizing Guide-Nanocarrier Pairs for PKD1
