Johannes
Schimunek
Research Associate in the Computational and Biological Learning group at the University of Cambridge. I build data-efficient models for small-molecule discovery: few-shot prediction, generative modelling, and chemical reasoning models.
About
- model
- Johannes Schimunek, Research Associate
- deployed at
- Computational and Biological Learning group, University of Cambridge, with José Miguel Hernández-Lobato
- intended use
- Few-shot and in-context prediction for molecules, generative models for small molecules, and chemical reasoning models
- trained on
- Mathematics and physics at the University of Augsburg. Doctorate in machine learning at JKU Linz (2025), supervised by Günter Klambauer and Sepp Hochreiter.
- earlier versions
- Postdoc at JKU Linz. Maths and physics teacher in Munich. Data scientist in industry.
- known for
- MHNfs, Bio-xLSTM, MolecularIQ
I work on machine learning for small-molecule discovery, with a focus on settings where data is scarce: predicting what a molecule does when only a handful of measurements exist. During my PhD I developed MHNfs, a few-shot model for bioactivity prediction. Since then I have worked on generative models for chemical and biological sequences, and on measuring how well language models reason about chemistry.
I joined the Computational and Biological Learning group in Cambridge in September 2026, working with José Miguel Hernández-Lobato. Before that I did my PhD and a postdoc at the Institute for Machine Learning at JKU Linz with Günter Klambauer and Sepp Hochreiter. I originally trained as a mathematics and physics teacher.
Selected publications
- ICLR 2026
- ICLR 2025
- J. Chem. Inf. Model. 2025
- ICLR 2024 workshops
Autoregressive Activity Prediction for Low-Data Drug Discovery
PaperAI for Good spotlight - ICLR 2023
Show all publications (11)Hide full list
- ICLR 2026
- Nucleic Acids Res. 2026
Machine learning-driven optimization of specific, compact, and efficient base editors via single-round diversification
Nucleic Acids Research - ESWC 2026
Structure is the Signal: Graph Encodings and GNNs for Constraint Repair in Collaborative Knowledge Graphs
European Semantic Web Conference (ESWC) - ICLR 2025
- J. Chem. Inf. Model. 2025
- Book chapter 2025
Few-Shot Learning in Drug Discovery
Book chapter in Applied Artificial Intelligence for Drug Discovery, Springer Nature - Mol. Inform. 2024
- ICLR 2024 workshops
Autoregressive Activity Prediction for Low-Data Drug Discovery
ICLR 2024 workshops: Practical ML for Low Resource Settings, and Data-centric Machine Learning Research (DMLR)PaperAI for Good spotlight - ICLR 2023
- ELLIS ML4Molecules 2021
A Generalized Framework for Embedding-Based Few-Shot Learning Methods in Drug Discovery
ELLIS Machine Learning for Molecule Discovery Workshop - QSAR 2021
Comparative Assessment of Interpretability Methods of Deep Activity Models for hERG
International Workshop on (Q)SAR in Environmental and Health Sciences
Talks & awards
- Invited talk · 2024AI for Good Global Summit
- Invited talk · 2024ELLIS Health Symposium on Machine Learning for Drug Discovery
- Award · 2024Best paper, ELLIS ML4Molecules workshop, for Bio-xLSTM
- Award · 2024AI for Good spotlight, ICLR DMLR workshop
- Award · 2024Top Reviewer, NeurIPS
Community & teaching
- Area ChairICLR 2027
- ReviewerICLR, NeurIPS, ICML, TMLR, Journal of Chemical Information and Modeling, Journal of Medicinal Chemistry
- OrganiserELLIS ML4Molecules workshop, 2024 and 2025. Weekly Drug Discovery Journal Club, since 2020.
- TeachingAI for Life Sciences, Deep Learning and Neural Networks, Hands on AI and Programming in Python at JKU Linz. Supervised more than ten bachelor's and master's theses.
For organisers
Introducing me at an event? Feel free to use this bio and photo.
Johannes Schimunek is a Research Associate in the Computational and Biological Learning group at the University of Cambridge, working with José Miguel Hernández-Lobato on machine learning for small-molecule discovery when data is scarce: few-shot bioactivity prediction, generative models for chemical and biological sequences, and chemical reasoning in language models. Johannes completed a PhD and a postdoc at the Institute for Machine Learning at JKU Linz with Günter Klambauer and Sepp Hochreiter and developed MHNfs, a few-shot model for bioactivity prediction.
