Tilak Sharma

Bioinformatics analyst — computational drug discovery & reproducible omics pipelines

I try to break a result before I report it. Every project below states its controls first — what the negative control was, where the leakage could hide, what the number looks like once you close that door. Usually the honest number is lower, and that is the point.

Selected work

inha-tb

Can a QSAR model trained on public InhA data pick new chemotypes — or is it just recognising analogues it has already seen?

2107 curated bioactivity records, 1417 unique molecules, across the InhA enzyme and M. tuberculosis whole-cell assays. Three model families, two split strategies. Docking validated first by redocking the co-crystallised ligand into 4TZK — RMSD 0.23 Å — because a docking score means nothing if the pose protocol cannot reproduce a known pose.

Held-out ROC-AUC	scaffold split	random split	inflation
InhA enzyme (XGBoost)	0.837	0.921	+0.084
M. tb whole-cell (RF)	0.841	0.911	+0.07

Across all six model×dataset combinations a random split overstates ROC-AUC by 0.069–0.119 (median 0.08). The cause is concrete: 29 and 89 Bemis–Murcko scaffolds appear on both sides of the random split. Every number quoted outside the full table is the scaffold-split one.

Scaffold-split test molecules are singletons; ROC-AUC by test-set composition — Why the gap exists. a Every scaffold-split test molecule is a singleton — no scaffold-mate anywhere in training. b Same model, same training set: accuracy falls monotonically as the test set loses its analogues. The random split's advantage is the analogues, not better learning.

ROC and PR curves, model comparison, accuracy vs chemical distance, enzyme-to-cell correlation — Full validation panel: ROC and precision–recall under both splits, accuracy binned by Tanimoto distance to training, and enzyme potency vs whole-cell kill (Spearman ρ = 0.91, n = 35) — the assay-translation check that justifies modelling the two endpoints together.

Docking score versus measured potency, ROC curves, enrichment factors, and the score-size correlation for InhA — Docking, tested rather than assumed. Score barely tracks potency (b); ROC-AUC reaches 0.65 on the enzyme and 0.58 whole-cell (c); enrichment stays below its attainable ceiling (d); and the score rewards molecular weight (Spearman ρ = −0.41, f). On the same molecules, neither fitted, QSAR reaches 0.83 and 0.84 (e). Reported because it is the result.

Re-run comparison of ROC-AUC inflation per model, and metric drift across dependency versions — Reproducibility as a measurement, not a claim. a The random-versus-scaffold gap reproduces on an independent re-run for every model. b Dependency-version drift stays under the 0.02 gate for ranking metrics; threshold metrics drift more, and varying both at once understates it.

Curation cascade from 4,450 ChEMBL records to 1,417 unique molecules; potency distributions; replicate-assay noise floor; InhA structure resolutions; binding-pocket contacts; docked subset chemistry — The dataset before any model. a 4,450 retrieved records reduce to 1,417 unique molecules once data validity, exact relations and a potency endpoint are required. c Replicate assays of the same molecule disagree by a median 0.23 log units — that is the floor any model error is measured against. d 4TZK (1.62 Å) chosen as receptor from ten screened structures.

QSARscaffold splitdocking RDKitXGBoostM. tuberculosis

omics-ml-leakage-audit

How much does each cross-validation leak actually inflate a reported AUC?

Four protocols — leaky feature selection, sample-level splits on patient-clustered data, preprocessing before the split, and a clean nested pipeline — run over three tasks and three public lung-cancer cohorts. The point is not that leakage exists; it is the size of the effect, measured per protocol, so a reviewer can tell which shortcut costs what.

cross-validationdata leakagescikit-learn nested CV

clinvar-vus-audit

How do variants of uncertain significance actually get resolved?

Two ClinVar releases 19 months apart, joined on variant identity, to measure which VUS move and in which direction. A reclassification rate is only interpretable against the denominator of variants that had the opportunity to move — so that denominator is constructed explicitly rather than assumed.

ClinVarvariant interpretationclinical genomics

protein-domain-burden

Are pathogenic missense variants enriched inside annotated conserved domains?

ClinVar missense variants mapped onto InterPro domain boundaries. The enrichment is tested against a length-matched background, because domains are not a random slice of the protein and an unmatched comparison would find enrichment whether or not it is there.

InterProClinVarenrichment testing

nf-fastq-qc

A QC pipeline that fails loudly instead of producing a plausible-looking report.

Nextflow, nf-core conventions, 42 tests covering the failure modes that matter — truncated reads, mismatched pair counts, empty inputs — because a QC pipeline that silently passes bad data is worse than no pipeline.

Nextflownf-corepytestcontainers

Also on GitHub

rnaseq-de-GSE60450 · scRNAseq-PBMC-Analysis · 16s-amplicon-report · inha-ai-docking-benchmark · oss-contributions — upstream work on nf-core/tools and Galaxy training material.

Background

M.Sc. Botany. Research internship at Inha University (South Korea) on the computational TB work above. nf-core member. Based in Baran, Rajasthan, India (IST).

Every number on this page is generated from the source data in the linked repositories — the inha-tb figures and table come from docs/facts.json, which the analysis writes and the write-up reads. No number here is typed by hand.