Independent Computational Laboratory · Istanbul

Mechanistic AI for the variants and tumours others overlook.

We turn genomic data into evidence-informed therapeutic hypotheses — ancestry-aware variant interpretation and single-cell analysis of rare cancers, released as open-source tools.

0false-benign calls across 1,277 known-pathogenic variants
9,534germline VUS reclassified in ATM & PALB2
9,157TBXT+ chordoma cells profiled at single-cell resolution
Research programmes

Models that explain the biological why.

Rather than treating models as black boxes, we build mechanistic AI and validate it against biological and clinical evidence — with a focus on rare and neglected entities. Where AI touches a clinical call, a deterministic engine decides and the model's reasoning stays auditable.

01

Ancestry-aware variant interpretation

Deterministic ACMG/AMP engines that ask whether evidence is strong enough for a given ancestry — exposing gaps Western-centric databases hide.

VUS Lens · VUS Pipeline
02

Single-cell biology of rare tumours

Single-cell analysis of chordoma is surfacing phenotypes invisible to conventional metrics — from tumour heterogeneity to ferroptosis vulnerabilities.

msep · Dual Shield
03

Safe-by-Design toxicity prediction

A hybrid Mixture-of-Experts pairing random forests with graph neural networks, designed to fail safely under data scarcity.

HybridTox · Tox21
Featured finding · 2026
Finding
Tumour cell populations that are individually diverse yet collectively disciplined.
Method
msep pairs per-cell entropy with across-cell variation, pathway by pathway.
Data
Sacral chordoma scRNA-seq (Arrieta et al., Neuro-Oncology, 2025).
individually diverse, collectively disciplined per-cell entropy → across-cell CV →
Schematic — illustrates the two axes msep measures; not data.
Open tools

Software we build and ship.

VUS Lens audits the evidence; VUS Pipeline classifies for the tumour board. msep is method-agnostic and works on any cellular system, and HybridTox fails safely on scaffolds it has never seen.

Python package · scRNA-seq msep

Multi-scale entropy profiling — pathway-decomposed, any cellular system.

pip install msep GitHub →
Clinical decision support VUS Lens

Ancestry-aware ACMG interpretation. 0 false-benign across 1,277 variants.

Clinical variant analysis VUS Pipeline

Deterministic ACMG/AMP classification for molecular tumour boards.

Toxicity prediction · Tox21 HybridTox

Random forest + GNN Mixture-of-Experts that fails safely on unseen scaffolds.

Publications & news

Latest work.

All publications →
ESMO TAT 2026
Paris
AI-Driven Reclassification of 9,534 VUS in ATM & PALB2 Surfacing actionable targets for PARP-inhibitor therapy.
Abstract (PDF) →
ESHG 2026
Gothenburg
Ferroptosis Resistance in Chordoma at Single-Cell Resolution Extreme GPX4 dependency and a dormant SREBF1–ACSL4 axis.
Abstract →
Research
2026
The Inverse Entropy Paradox in Chordoma Individually diverse, collectively disciplined tumour cells.
Read →
Release
2026
VUS Lens — Built with Claude Open source, and live for researchers to use and audit.
Live tool →
People & collaborators

Founded in 2024, Istanbul.

Ozge A. Cavus Founder & Principal Investigator
Prof. Ayşegül Kuşkucu Scientific collaborator · Department of Medical Genetics, Yeditepe University
Presented at ESHG 2025ESMO 2025ESHG 2026ESMO TAT 2026