A content-addressed audit engine for drug-target direction of effect — inhibit versus activate — built on Open Targets colocalization and benchmarked against approved drug–target pairs. Rules are SHA-locked so a verdict cannot silently change, and the engine is allowed to refuse when the genetics are ambiguous. The production layer adds batch auditing, Nextflow and Snakemake DAGs, and a Docker image.

Tests whether gnomAD LOEUF constraint works as a drug-target safety oracle, as it is routinely used. It does not: LOEUF remains a solid predictor of genetic loss-of-function tolerance, the thing it was built to measure, but is near-useless for predicting clinical safety of inhibition. A negative result, shipped as one.

A one-page, reproducible target-confidence summary that resolves the underlying evidence into an explicit verdict — including refusal — rather than a score whose provenance is unrecoverable.

When can you trust ESM-C zero-shot variant-effect rankings, and when can you not? A reproducible protein-language-model benchmark showing that a model’s self-consistency across sizes does not predict its accuracy, plus a calibration tool for deciding when to rely on it.

One fixed nine-stage QSAR and conformal-prediction pipeline — curation, noise-floor estimation, random-forest QSAR on ECFP4, scaffold-split cross-validation, conformal prediction, applicability domain, freedom-to-operate, design, synthesis triage — run unchanged across ENPP1, NLRP3, TYK2 and IRAK4, spanning a 23× range in public data. Its most valuable outputs are the two refusals.

Why the best genetically validated pain target keeps failing in the clinic. Argues a genetic–pharmacological asymmetry: lifelong genetic loss of Nav1.7 is cardiovascular-silent, while acute pharmacological block causes on-target autonomic toxicity. Single-cell atlas analysis, a leaky homeostatic-compensation model, and gnomAD / ClinVar / Open Targets human genetics.

A structure-based paralog selectivity counter-screen for the immuno-oncology target ENPP1: dock candidates into ENPP1 and its close relatives ENPP2 (autotaxin) and ENPP3 in one identical zinc-centered box, then rank by cross-paralog margin rather than raw affinity. The top affinity binder reversed to an off-target liability. Ships native controls, bootstrap CIs and two honest negative benchmarks; runs on a free Kaggle T4.

Maps type 1 diabetes GWAS loci to the pancreatic cell types they likely act in, using τ-based cell-type specificity across the HPAP single-cell RNA-seq atlas, cross-validated against autoantibody-positive pre-clinical transcriptional change.

Runs a standard ChEMBL-driven target-discovery pipeline on Madurella mycetomatis, a neglected fungal pathogen nobody had mapped — and catches the pipeline’s own artifact: the top-ranked gene was 384 duplicate records of HDAC4. Triaged to a hardened shortlist, with the audit that caught the error shipped alongside.

An eight-stage open drug-discovery pipeline for Chagas disease targeting the T. cruzi cysteine protease cruzain: three parallel scoring tracks fused into a consensus, a selectivity counter-screen against three human cathepsins, and ADMET filtering. Runs on a free Kaggle T4.

An empirical evaluation of which public-data features actually predict PROTAC E3-ligase tractability — and which are just correlated with how well-studied a ligase is.

A layered ablation of rule-based models for therapeutic exon-skipping targets in Duchenne muscular dystrophy. Starts from the reading-frame rule, then adds splice-topology and domain constraints and measures what each one buys. The published version corrects a substantive modelling error in the first draft — including one found by a reader whose child carries the deletion in question — and that correction is the scientific content.

Where do disease mutations sit on the sodium–potassium pump, and can you tell without arguing in a circle? A reproducible integration of cross-species conservation, experimental structures, ClinVar curation, gnomAD constraint and AlphaMissense predictions across the four human Na⁺/K⁺-ATPase α-subunit paralogues.