MSc student in Molecular Biotechnology and Genetics at Istanbul University, working at the intersection of machine learning, genomics, and research software engineering.

I build reproducible, provenance-tracked computational systems that turn biological data into decision-grade evidence for healthcare and diagnostics — with a particular focus on making the validation honest, not just the model accurate.

An end-to-end ML platform predicting antimicrobial resistance directly from bacterial whole-genome sequencing, delivering a versioned, queryable AMR biomarker knowledge base.

Scale: 45 models · 6 ESKAPEE organisms · 14 antibiotic classes · mean lineage-aware CV ROC-AUC 0.842 Scope: ~15.7k lines of Python across 46 modules · 220 commits · sole author

Highlights

• Out-of-core XGBoost over sparse matrices with millions of features — a custom streaming DataIter → QuantileDMatrix, plus binary histogram quantisation (max_bin=2) for a 128× training-memory reduction
• Alignment-free features — unitigs from a compacted de Bruijn graph, no reference genome required
• Lineage-aware validation (PopPUNK clusters + StratifiedGroupKFold) as the reported metric. One model scores 0.985 on a random split but 0.429 under lineage-aware CV — a live demonstration of why validation design decides whether a clinical ML result is real
• 7-layer evidence stack — CPSS stability selection with a PFER bound, TreeSHAP, permutation nulls, pyseer LMM, CARD/NCBI BLAST reverse-translation, SNP allele checks, external concordance vs. AMRFinderPlus/ResFinder
• Reproducibility engineering — three pinned Apptainer containers with committed conda lockfiles; every run stamps git commit, seed, config hash and eight tool versions into the knowledge base
• Delivered as a product — 13-table SQLite KB with semantic schema versioning, a FastAPI REST API, and a Streamlit explorer
• Registry-driven architecture — adding an organism is a registry entry, not a code change
• Operated on a national HPC cluster (SLURM), 45/45 models completed with zero failures

Python · XGBoost · Optuna · scikit-learn · SciPy sparse · FastAPI · SQLite · Apptainer · SLURM · pytest · GitHub Actions

The origin of the above: a feasibility study establishing that k-mer frequency features can classify beta-lactam resistance in E. coli, benchmarking Random Forest, Gradient Boosting and SVM under GridSearchCV.

Superseded by ML_AMR_Prediction_v2, which replaced per-chunk incremental training with full-data streaming boosting, chunk-level splits with lineage-aware grouped CV, and test-set threshold tuning with leakage-free training-derived thresholds.

An automated, schedulable ETL pipeline that monitors NCBI PubMed, generates structured bilingual (EN/TR) science summaries with an LLM, and archives them to a cloud PostgreSQL database.

Designed for unattended operation: primary-key deduplication that short-circuits before any LLM call, all-or-nothing writes, and exponential retry backoff across two independently rate-limited APIs.

Python · Biopython/Entrez · Google Gemini · Supabase (PostgreSQL)

A seven-phase statistical ML pipeline for clinical diabetes prediction, built to demonstrate the mathematics rather than the API.

Includes maximum-likelihood estimation with a hand-derived analytic gradient (OLS ≡ MLE verified to 3.16 × 10⁻¹⁶), batch gradient descent from scratch, empirical bias–variance decomposition, logistic Z-statistics and odds ratios, and BCa bootstrap confidence intervals.

• Machine learning for clinical diagnostics and biomarker discovery
• Out-of-core / memory-bounded computation on high-dimensional data
• Reproducibility, provenance and validation engineering for scientific ML
• Bioinformatics pipelines and workflow automation
• Research software that other people can actually run

Languages — Python · SQL · Bash

ML & Statistics — XGBoost · scikit-learn · Optuna · SHAP/TreeSHAP · stability selection · grouped cross-validation · bootstrap inference · hypothesis testing with FDR control

Data & Scientific Computing — pandas · NumPy · SciPy (sparse CSR) · matplotlib/seaborn

Engineering — Git · pytest · GitHub Actions CI · ruff/mypy · FastAPI · SQLite · Apptainer/Singularity · SLURM/HPC · Linux · YAML-driven configuration

Bioinformatics — WGS analysis · unitigs / de Bruijn graphs · PopPUNK population genomics · BLAST+ · pyseer · CheckM2/QUAST · CARD/ARO ontology · NCBI Entrez

• LinkedIn: linkedin.com/in/demirbaseren
• Email: eren0demirbas@gmail.com