MSc, Koç University · Research intern, DELTA — Inria Nancy
I build generative models for peptides and protein binders, and I care as much about how we evaluate those designs as about how we generate them.
I am an MSc student in Computational Sciences and Engineering at Koç University, advised by Prof. Attila Gürsoy and Prof. Özlem Keskin, where I work on antimicrobial peptide (AMP) design against carbapenem-resistant targets. Since June 2026 I have also been a research intern with the DELTA team at Inria Nancy, with Hamed Khakzad and Yasaman Karami, working on peptide binder design and on evaluating binders with molecular dynamics rather than static scores alone.
Recent work includes AMP-DiT, a classifier-conditioned discrete diffusion transformer that generates peptides directly in sequence space (GenBio Workshop at ICML 2026); AMP-BindGen, a guided design pipeline for peptide inhibitors of the carbapenemases NDM-5 and KPC-3, now in experimental testing; and IDPepDesign, a target-conditioned insertion–deletion diffusion model for peptide binder design.
Before Koç I was an AI Engineer in Computational Biology at 310.ai (2024–2025), working on protein design, taxonomy prediction and binder pipelines. I hold a BSc in Computer Engineering from Sharif University, with a thesis on paratope prediction using protein language models (Scientific Reports, 2024).
Four threads, all pointed at the same question: how do we design short binders that are genuinely selective, and how do we know before we go to the bench?
Existing AMP generators lean on protein language models trained on full-length proteins, which mismatches the peptide distribution. AMP-DiT instead runs a DiT-style denoiser directly over sequence space, conditioned on Macrel-derived activity and non-hemolytic safety scores. Conditioning on one predictor transfers to others, which suggests it learns antimicrobial structure rather than overfitting a scorer.
A guided design pipeline for peptide-like binders against two clinically important carbapenemases. Candidates are ranked by a composite scheme combining interface confidence, catalytic-site engagement, predicted antimicrobial activity and toxicity filters. Designed peptides are in activity testing, and the first results have now come back from the lab.
Many promising binders are simply too large to synthesise or deliver. IDPepDesign builds on SCISOR, an insertion–deletion diffusion model, and adds two conditioning paths — a pooled target representation injected through denoising, and residue-level cross-attention — turning an unconditional peptide generator into a target-aware binder designer. Perplexity 17.1 → 11.3 and FPD 1.84 → 0.75, with competitive interface quality across 11 benchmark targets.
Binder design is still judged mostly by interface pLDDT, PAE and energy terms on a single predicted complex — none of which capture induced fit or show whether a binder stays selective across a conformational ensemble. BinderEval scores designs using dynamics data and COMPASS, and a new version of COMPASS is now available.
Preprint, 2026 · ML4SM Workshop, NeurIPS 2026
A target-conditioned insertion–deletion diffusion model that turns an unconditional peptide generator into a binder designer.
Preprint, 2026
Replaces single-structure confidence scores with ensemble- and dynamics-based evidence for whether a designed binder actually holds.
Merck Challenge submission, 2026
GenBio Workshop at ICML 2026 (poster)
IEEE Journal of Biomedical and Health Informatics, 2025
Scientific Reports 14, 29141 (2024)