Oz Kilim

About

I am a computational pathology researcher. My work asks a simple question: how much can a haematoxylin-and-eosin slide tell us about what will happen to a patient? I train and validate deep-learning models — pathology foundation models, multiple-instance learning, multi-centre survival analysis — that turn whole-slide images into risk scores for cancer recurrence and treatment response, and I build the open-source infrastructure that lets collaborating hospitals run those models on their own data.

Before Boston I did a PhD in the group of István Csabai at Eötvös Loránd University in Budapest, working on deep learning, complex systems and bioinformatics — from domain shift in medical imaging to AlphaFold2 mutational scans of the SARS-CoV-2 spike. Before that I read Chemistry at Lincoln College, University of Oxford; my Part II research year (2017–18) was in Justin Benesch's group, studying small heat-shock protein chaperones with gas-phase HDX and ion-mobility mass spectrometry.

Outside the lab I am CTO of Tropiflo, and I spend my evenings on systems problems: streaming mixture-of-experts models from disk so that trillion-parameter networks run on a single machine, pooling idle office Macs into one inference cluster, and making coding agents controllable by voice.

Research

Multimodal biomarkers for early-stage lung adenocarcinoma

PATH-ORACLE combines whole-slide-image embeddings with a 46-gene RNA signature to predict recurrence in stage I lung adenocarcinoma, validated across international cohorts. Weights, cut-offs and inference code are public in PATH-ORACLE-INFERENCE. The PRELUDE consortium pairs histology with omics to predict relapse in stage IA/IB disease.

Platinum response in high-grade serous ovarian cancer

Our npj Precision Oncology paper showed that histopathology and proteomics are synergistic for predicting platinum response in HGSOC. The KAIROS consortium extends this across centres.

Robustness, domain shift and interpretability

Physical imaging parameters — not just biology — drive domain shift in medical images (Scientific Reports, 2022). Spatial statistics borrowed from cosmology give interpretable prognostic features in histopathology (ICCV Workshops, 2025). And transfer learning may explain why pigeons can be trained to detect cancer on slides (Bioinspiration & Biomimetics, 2024).

Selected publications

Full list on Google Scholar (111+ citations, h-index 5) and the Harvard Catalyst profile. Red bar = first author.

• A multimodal AI biomarker PATH-ORACLE improves prediction of recurrence in stage I lung adenocarcinoma. O Kilim, O Pipek, Z Sztupinszki, M Diossy, A Prosz, C Naceur-Lombardelli, et al. medRxiv, 2026. doi:10.64898/2026.01.28.26344973
• Zero-shot biological reasoning with open-weights large language models reproduces CRISPR screen based prediction of synthetic lethal interactions. A Prosz, Z Sztupinszki, M Diossy, O Kilim, B Zimon, I Csabai, Z Szallasi. bioRxiv, 2026. doi:10.64898/2026.01.28.702211
• Deep-learning-based breast cancer stage prediction from H&E-stained whole-slide images in resource-constrained settings. Z Bedőházi, A Biricz, O Kilim, et al. Journal of Pathology Informatics, 2026. doi:10.1016/j.jpi.2026.100644
• Histopathology and proteomics are synergistic for high-grade serous ovarian cancer platinum response prediction. O Kilim, A Olar, A Biricz, L Madaras, P Pollner, Z Szallasi, Z Sztupinszki, et al. npj Precision Oncology, 2025. doi:10.1038/s41698-025-00808-w
• From Cosmos to Clinic: Interpretable Spatial Statistics for Histopathology Prognosis. A Barna, O Kilim, I Csabai. IEEE/CVF ICCV Workshops, 2025. doi:10.1109/ICCVW69036.2025.00100
• Transfer learning may explain pigeons' ability to detect cancer in histopathology. O Kilim, J Bóskay, A Biricz, Z Bedőházi, P Pollner, I Csabai. Bioinspiration & Biomimetics, 2024. doi:10.1088/1748-3190/ad6825
• SARS-CoV-2 receptor-binding domain deep mutational AlphaFold2 structures. O Kilim, A Mentes, B Pál, I Csabai, Á Gellért. Scientific Data, 2023. doi:10.1038/s41597-023-02035-z
• Impact evaluation of score classes and annotation regions in deep learning-based dairy cow body condition prediction. SÁ Nagy, O Kilim, I Csabai, G Gábor, N Solymosi. Animals, 2023. doi:10.3390/ani13020194
• Ixodes ricinus tick bacteriome alterations based on a climatically representative survey in Hungary. R Farkas, M Papp, O Kilim, et al. Microbiology Spectrum, 2023. doi:10.1128/spectrum.01243-23
• Physical imaging parameter variation drives domain shift. O Kilim, A Olar, T Joó, T Palicz, P Pollner, I Csabai. Scientific Reports, 2022. doi:10.1038/s41598-022-23990-4

Open-source software

PATH-ORACLE-INFERENCE — weights, cut-offs and inference code for the PATH-ORACLE lung adenocarcinoma recurrence models (TITAN embeddings ± 46-gene RNA signature → risk). Apache-2.0.

SlideCrush — patch and embed whole-slide images with every histopathology ViT foundation model in one package.

onco_run — model-agnostic WSI inference runner shipped as a single Docker image; collaborators just run ./run.sh.

colibri — pure-C mixture-of-experts inference engine that streams experts from disk; running ~1-trillion-parameter models off a spinning hard drive.

sovereign-mesh — simulator for pooling idle office Macs into one machine big enough for frontier open-weight models.

voice-claude — "Hey Jarvis": talk to your coding agent through a Bluetooth headset, Whisper on the Apple Silicon GPU.

runway — live cloud-credit runway (AWS / Azure / GCP) as a coding-agent skill.

Quick facts

Who is Oz Kilim?

A computational pathology researcher and engineer. Postdoctoral Research Fellow at CHIP, Boston Children's Hospital / Harvard Medical School, and CTO of Tropiflo.

What does he work on?

Deep-learning models that predict cancer recurrence and treatment response from whole-slide histopathology images — especially early-stage lung adenocarcinoma (PATH-ORACLE, PRELUDE) and platinum response in high-grade serous ovarian cancer (KAIROS) — plus efficient inference of very large AI models on modest hardware.

Where did he study?

MChem in Chemistry at Lincoln College, University of Oxford (Part II in the Benesch Group, 2017–18); PhD at Eötvös Loránd University (ELTE), Budapest, in the group of István Csabai.

How to cite or contact?

Publications are on Google Scholar; email ozsamkilim@gmail.com.