Özge Ayşe Çavuş

• Turkish clinical input: a 10,572-term diagnosis vocabulary is mapped to international disease and phenotype ontologies; the gene set for the pre-diagnosis is drawn from curated sources, and no variant is suppressed.
• Deterministic classification: an ACMG/AMP engine based on the Tavtigian Bayesian framework; compound-heterozygote assessment, cross-gene isolation, and inheritance–zygosity conflict checks. The language model has no classification authority.
• Turkish population data: 46.4 million records compiled from 3,362 individuals are used alongside international data; a separate layer audits the reliability of the frequency evidence.
• Clinician dialogue layer: free-text questions in Turkish about the result; the answer is generated only from sources gathered for that question, and an unverifiable citation cannot enter the answer.
• Structural safety and validation: absence of data and absence of finding are kept distinct at every layer, an audit trail is written, and continuous regression checks run on a 390-variant reference set; run end-to-end on a real clinical case with Yeditepe Medical Genetics.

• The pilot open-source project (MIT) of the main system; a deterministic ACMG engine, with no language model in the classification call.
• Ancestry-aware confidence auditing: a rarity claim is flagged when the allele count is insufficient under the rule of three.
• Warns at repeat-expansion loci that short-read exomes cannot see; cross-source conflicts and unreachable sources are never read as benign.
• ClinGen HBOP VCEP specifications are applied for ATM and PALB2; where a generic default is used for other genes, this is explicitly labelled.
• Validation: zero false-benign across 1,277 known-pathogenic variants; 58 of 94 definite-benign variants correctly classified; run on a 10,747-variant ATM/PALB2 cohort.
• Live: vuslens.aistanbulresearch.com/live · github.com/aistanbulresearch/vus-lens

• Combines per-cell Shannon entropy with across-cells coefficient of variation in single-cell RNA-seq data, decomposed by biological pathway.
• Surfaces states invisible to single-scale analyses: populations that are individually diverse yet collectively disciplined; produces a Coordination Score that reduces the finding to a single number.
• Access to 33,000+ gene sets across 12 MSigDB collections; bootstrap confidence intervals, the Fano factor, cross-pathway pseudo-perturbation, and XBP1 stress-consolidation analyses.
• An optional Bayesian validation layer (scVI) separates technical variance from biological variance.
• MIT-licensed, published on PyPI, validated with 48 unit tests: github.com/aistanbulresearch/msep

• Part of the same chordoma research strand as the ESHG 2026 presentation and the msep package.
• 10x single-cell RNA-seq data of 114,989 cells from six patients: quality control, doublet removal, batch-effect correction with Harmony, clustering, and cell-type annotation.
• Hypothesis testing with ligand–receptor interaction analysis (LIANA) and pseudotime; language-model agents connect to the literature via PubMed, ClinicalTrials.gov, and Open Targets to produce hypothesis validation and therapeutic reasoning.
• A counter-argument layer that audits its own results runs on a different model from the primary reasoning agent; its conditional-approval decision and unresolved objections are published openly in the repository.
• Findings are computational and await experimental validation; compound-level outputs are kept in a separate, access-restricted repository: github.com/aistanbulresearch/Dual_Shield

An independent research group without a registered legal entity.

• Design and development of decision-support architectures for VUS interpretation in clinical genetics.
• Coordination of the scientific collaboration with the Department of Medical Genetics, Yeditepe University Faculty of Medicine.
• Computational hypothesis generation in rare cancers and precision oncology.
• Running the research programme's cycle of goal-setting, resource planning, and output.

• Corporate advisory and management of intellectual-property portfolios.
• Over fifteen years working within complex regulatory frameworks, managing teams and client relationships.

European Intellectual Property Law, LL.M. (by thesis)

Master's thesis: The Tragedy of Anticommons in Biotechnology Field and Possible Solutions.

Faculty of Law, LL.B.