Models that explain the biological why.

Overview

Deterministic ACMG/AMP engines that ask whether evidence is strong enough for a given ancestry — exposing gaps Western-centric databases hide. The same rigour extends to the molecular logic of treatment failure: from reclassifying Variants of Uncertain Significance (VUS) in DDR genes to uncovering intrinsic resistance to Antibody-Drug Conjugates (ADCs), we turn genomic complexity into actionable targets. Where AI takes part in a clinical call, the class is set by rules; models advise and are audited.

Recent work

• VUS Lens — ancestry-confidence auditing: A deterministic ACMG rule engine that flags where variant evidence is unreliable for under-represented ancestries. Validated against 1,277 known-pathogenic variants with zero false-benign calls; open source and live.
• VUS Pipeline — decision support for tumour boards: ACMG/AMP 2015 + ClinGen SVI + Tavtigian points applied deterministically, with an advisory mechanistic interpretation layer that never changes the class and source-cited reports. How it works.
• AI-Driven Variant Resolution (ESMO TAT 2026, Paris): Proteome-wide AI reclassification of 9,534 germline VUS in the DDR genes ATM and PALB2, identifying hidden candidates for PARP-inhibitor therapy.
• Decoding Intrinsic Resistance: Uncovering the "1p32 Co-Deletion" syndrome in lung cancer — a genetic blind spot driving resistance to TROP2-ADCs.
• Synthetic Lethality: Transforming molecular liabilities into therapeutic assets by validating CDK4/6 inhibitors as a rescue strategy for drug-resistant tumours.

Overview

Single-cell analysis of chordoma, a rare and treatment-resistant bone tumour, is surfacing phenotypes invisible to conventional metrics — from tumour heterogeneity to ferroptosis vulnerabilities. Where data are scarce, we use 'Biological Twins': mechanistic AI that simulates the tumour microenvironment and maps phenotypic convergences with common cancers to repurpose validated therapies.

Recent work

• The Inverse Entropy Paradox in Chordoma: Applying our multi-scale entropy profiling package (msep) to sacral chordoma single-cell RNA-seq data (Arrieta et al., Neuro-Oncology, 2025) reveals tumour cell populations that are individually diverse yet collectively disciplined — a structural signature that single-scale entropy metrics miss entirely.
• Ferroptosis Resistance Architecture (ESHG 2026, Gothenburg): Single-cell profiling of 9,157 TBXT⁺ chordoma cells uncovered transcriptional "metabolic lockdown" — a dormant SREBF1-ACSL4 axis (Pearson r=0.017) and extreme GPX4 stoichiometric dominance in 74.9% of tumour cells, with a functionally silent system xc⁻ antiporter. Chordoma sits on a precarious metabolic edge: resistant to conventional ferroptosis inducers at baseline, yet vulnerable if static ER/lipid homeostasis is disrupted.
• The "Dual Shield" Model in Chordoma: A single-cell RNA-seq and LLM-agent pipeline describing two layers of defence — an outer immune-evasion shield (HLA-E/NKG2A) and an inner structural-resistance shield (vimentin-driven partial EMT). Combination strategies that dismantle both layers are designated for validation in 3D organoid co-cultures. Code is open source; outputs are research hypotheses, not clinical guidance.
• The "Biological Twin" Strategy: Overcoming data scarcity by mapping phenotypic overlaps between rare chordoma and pancreatic cancer (PDAC) to accelerate drug repurposing.
• Quadruple-Action Protocol: A systemic attack on the CXCR4 hub to dismantle the tumour's "fortress", simultaneously targeting radioresistance, fibrosis, and immune exclusion.
• TechBio Integration: Leveraging the TxGemma-27B model for mechanistic reasoning to simulate complex WNT5A/TGF-β feedback loops in silico.

Overview

A hybrid Mixture-of-Experts pairing random forests with graph neural networks, designed to fail safely under data scarcity. Deep models can fail silently on unseen chemical scaffolds; we engineer 'fail-safe' architectures and validate them with rigorous, reproducible protocols.

Recent work

• Hybrid Mixture of Experts (MoE): Fusing the topological intuition of Graph Neural Networks (GNNs) with the explicit memory of Random Forests through a transparent stacking ensemble.
• Epistemic Uncertainty Management: A "Safe-by-Design" framework engineered to prevent silent failures on unseen chemical scaffolds.
• Tox21 Benchmarking: 12 toxicity endpoints, a 5-seed protocol and nested scaffold split. The preprint is on ChemRxiv and under review at the Journal of Chemical Information and Modeling.