Single-cell and perturbation bioinformatics · machine learning MSc Bioinformatics, University of Edinburgh · Edinburgh, UK

CRISPRi perturbation screens, cell foundation models, and out-of-core pipelines for single-cell data that will not fit in memory. Two years at the bench before moving to computation, so I know how the data was made — and I measure a dataset before I model it.

📄 Portfolio · LinkedIn · mgkavya6@gmail.com

Predicting how three anonymised human cell lines respond to 300 CRISPRi gene knockdowns from their unperturbed profiles alone — zero-shot generalisation under distribution shift. 360,000 cells × 18,533 genes.

• Measured the noise floor (0.0132 MAE) from non-targeting guides and the ceiling (~0.30) by scoring control cells against each other, before building anything.
• Identified all three anonymised cell contexts from marker genes alone, and caught undocumented subsampling in the released data (400 cells per guide, SD 0.0).
• Diagnosed a metric pinned at its floor: truncation alone cost nmae −6. Predicting at full density moved the score from −0.9976 to −0.0317 and the rank from 455 to 274.
• Fine-tuned STATE (242M-parameter perturbation transformer, ESM2 embeddings) and built the count-conversion and submission path end to end.
• Found the released training config excluded competition_train — 221,273 cells, three quarters of the available in-distribution data.

Breast cancer relapse classification on GEO GSE2034 (286 tumours, 22,215 probes), with scaling and feature selection refit inside every cross-validation fold. A controlled experiment measures how much the common shortcut inflates reported performance: +0.120 ROC-AUC — enough to change which model looks best. Reports the honest 0.61–0.64 test AUC rather than the flattering one.

Modular Nextflow DSL2 pipeline with imported process modules, samplesheet-driven channels, Docker, and a Slurm profile. Proof-of-concept scope, built to demonstrate pipeline structure and portability across local, container and cluster execution.

Integrated 16S + shotgun metagenomics of kvass fermentation. Four taxonomic profilers run in parallel and compared for agreement rather than trusting one, plus functional profiling via HUMAnN3 and DRAM2. Versioned manifests and reproducible outputs throughout.

Deployed web platform for wet-lab scientists: 13 experimental calculators, a statistics module that checks its own assumptions, and an offline-capable electronic lab notebook.

Single cell & perturbation — AnnData/h5ad · CSR sparse · out-of-core HDF5 with h5py · CRISPRi Perturb-seq · differential expression (Wilcoxon, Benjamini–Hochberg) · marker-gene annotation · cell-eval

Machine learning — scikit-learn · Random Forest, SVM, logistic regression, XGBoost · leakage-safe cross-validation with Pipelines · ROC-AUC, PR-AUC · arc-state (STATE) fine-tuning and zero-shot inference · ESM2 embeddings

Statistics — multiple-testing correction and FDR · nonparametric tests · linear models and ANOVA · GLMs with Poisson and binomial outcomes

Transcriptomics — RNA-seq end to end · FastQC/MultiQC · STAR, HISAT2 · DESeq2, edgeR · TPM/TMM/VST normalisation · PCA · GO and KEGG enrichment

Engineering — Python · R · Bash · numpy, pandas, scipy.sparse · Nextflow DSL2 · Docker · Slurm/HPC · Git · Linux · RMarkdown

Wet lab — DNA/RNA extraction · PCR and qPCR · cloning · protein expression · western blotting · cell culture

In progress — Scanpy, Seurat and Cell Ranger · AWS (S3, Batch, IAM) · PyTorch from first principles · scVI

Competing in the Arc Virtual Cell Challenge 2026 (submission deadline 5 November). Bioinformatics and AI-evaluation contract work — designing deterministic computational biology tasks and assessing model outputs for factual accuracy and reproducibility.