Paul Schwerd-Kleine

AI research scientist for health — deep learning, foundation-model adaptation, and rigorous statistical evaluation.

I build and rigorously evaluate machine-learning and statistical models on real clinical and biological data. Data Science Innovation Fellow at Novartis (Cambridge, MA); PhD from the German Cancer Research Center (DKFZ).

View CV Contact
Portrait of Paul Schwerd-Kleine
Novartis

About me

I develop and stress-test machine-learning and statistical models for health — from adapting single-cell foundation models to building interpretable survival models that predict clinical outcomes. I care as much about honest evaluation as about headline accuracy: strong baselines, uncertainty quantification, and clinically meaningful metrics.

My work spans deep learning and foundation-model adaptation, statistical and survival modelling, multi-omics integration, and single-cell & spatial transcriptomics — all aimed at translation into decisions that help patients. My domain roots are in oncology, but the methods travel.

Paul in profile

Selected projects

Simulated vs observed Kaplan-Meier curves for treatment and control arms

popTK — trial outcomes from preclinical data

A hierarchical PK/PD model with three free parameters per drug that predicts clinical combination efficacy before the trial.

Recovered 4 phase III trials and predicted biomarker-dependent benefit →

 UMAP of single-cell states mapped to simulated vs observed survival curves

Cell2Trial

Benchmarking single-cell foundation models against simple baselines for predicting patient survival from unmatched single-cell data.

A simple baseline beat zero-shot foundation models — a study in honest evaluation →

 Schematic of NEUROD1-driven neuroendocrine transdifferentiation in HER2+ breast cancer

Neuroendocrine transdifferentiation in breast cancer

Multi-omics integration identifying a master regulator and a targetable therapy-escape state.

Co-first author · under review at Cancer Discovery →

 Ball-and-stick model of the macrocyclic host molecule cucurbit[7]uril (CB7)

ML for molecular binding affinity

Graph neural networks predicting host–guest binding, with leakage-safe chemical evaluation.

Scaled to screen ~9–10M PubChem compounds →

Selected publications

  1. A latent sensitivity framework translates preclinical drug response to clinical combination efficacy. First author

    Schwerd-Kleine P, Wagner JP, Palmer AC. Submitted, Cancer Research.

  2. Neuroendocrine transdifferentiation defines a monitorable and targetable therapy-escape state in HER2-positive breast cancer. Co-first author

    Schwerd-Kleine P*, Würth R*, Cheytan T*, et al. Under review, Cancer Discovery. (*equal contribution)

  3. An in-depth single-cell and spatial map of inflammatory breast cancer reveals an immunosuppressive and inflammation-like microenvironment. First author

    Schwerd-Kleine P, et al. Manuscript.

  4. Biopsy-derived organoids in personalised early breast cancer care. First author

    Schwerd-Kleine P, et al. Int. J. Cancer 2025;156(11):2200–2209. doi:10.1002/ijc.35386

  5. Circulating tumor cell plasticity determines breast cancer therapy resistance via NRG1–HER3 signalling. Co-author

    Würth R, et al. Nature Cancer 2024. doi:10.1038/s43018-024-00882-2

  6. Induction of a metabolic switch from glucose to ketone metabolism programs ketogenic diet–induced therapeutic vulnerability in lung cancer. Co-author

    Wu Y, et al. Cell Metabolism 2025;37:2233–2249.e9. doi:10.1016/j.cmet.2025.10.001