Software we build and ship.

Multi-Scale Entropy Profiling integrates per-cell Shannon entropy with across-cells coefficient of variation (CV), decomposed by biological pathway, to characterise population-level transcriptomic coordination in single-cell RNA-seq data. The framework reveals states invisible to single-scale analyses — such as populations that are individually diverse (high per-cell entropy) yet collectively disciplined (low across-cells CV) — and is applicable to any cancer type or cellular system.

• Pathway-decomposed entropy, not a single global score
• 33,000+ curated gene sets via MSigDB out of the box
• Bootstrap confidence intervals reported for every estimate
• Built for standard scRNA-seq workflows

VUS Lens is a deterministic ACMG classification engine that does something most variant interpretation tools do not: it audits whether the evidence behind a call is actually strong enough for the patient's ancestry, rather than assuming Western-centric reference databases generalize across populations. Its reasoning layer is built with Claude.

• Deterministic ACMG rule engine — same input, same output, every time
• Ancestry-confidence auditing surfaces where reference data falls short
• Validated against 1,277 known-pathogenic variants with zero false-benign calls
• Reasoning layer built with Claude

VUS Pipeline is a production-grade clinical decision-support system that applies the ACMG/AMP 2015 + ClinGen SVI + Tavtigian point framework deterministically, gathers evidence in parallel from authoritative sources, and layers advisory mechanistic interpretation on top — every step traceable and source-cited. The class is set by rules, never by an LLM.

• Deterministic ACMG / SVI decision engine (Tavtigian points)
• 8-layer parallel evidence collection (gnomAD v4, ClinVar, SpliceAI, ClinGen, CIViC)
• Advisory mechanistic interpretation layer — never changes the class
• Source-cited reports (HTML / PDF / DOCX) with a citation-verified assistant

HybridTox integrates the explicit structural memory of Random Forests with the topological intuition of Graph Neural Networks through a transparent stacking ensemble that acts as a Mixture of Experts. It is built to prevent deep learning's silent failures — high-confidence errors on out-of-distribution chemistry — and to act as a fail-safe in data-scarce regimes.

• Random Forest + GNN experts, combined by a transparent logistic-regression stacker
• Evaluated on 12 Tox21 endpoints with a 5-seed protocol
• Nested scaffold split — tested on chemistry the model has not seen
• Preprint on ChemRxiv; under review at the Journal of Chemical Information and Modeling

Dual Shield is a single-cell RNA-seq and LLM-agent pipeline for chordoma. It describes an outer immune-evasion shield (HLA-E/NKG2A) and an inner structural-resistance shield (vimentin-driven partial EMT), and searches for combination strategies that could dismantle both. Agent-generated hypotheses pass through an adversarial review before they are reported.

• Single-cell evidence for both shields from chordoma scRNA-seq
• LLM-agent hypothesis generation with adversarial review
• Candidate combinations designated for 3D organoid co-culture validation
• Apache-2.0 code; restricted compound-level outputs kept private

Research hypotheses only — not clinical guidance.