Till Richter

I am Till Richter, a MUDS Ph.D. researcher at the Technical University of Munich (TUM) and Helmholtz Munich, advised by Fabian Theis, Niki Kilbertus, and Yoshua Bengio. I am also a member of the Munich Center for Machine Learning (MCML) and the Munich Data Science Institute (MDSI), and I collaborate with the Causal Cell Dynamics Lab (Helmholtz Munich · MILA Montreal). In 2025 I was a research intern at Microsoft Research in Cambridge, MA, working with Lorin Crawford on multimodal cell foundation models.

What I work on — self-supervised learning as a design principle for biology

Biological systems are multimodal, multiscale, and dynamic, and most of their structure is conditional rather than marginal: which genes are co-regulated, how morphology constrains transcription, how a cell state evolves into the next. Traditional unsupervised methods optimize the marginal likelihood of the data and end up allocating capacity to whatever dominates variance — often library size, batch, or donor — rather than to the conditional dependencies that define biological mechanism.

My research argues for a shift from describing the data distribution to predicting within it: structured self-supervised learning (SSL) as a unifying inductive bias for cellular biology. Concretely, I build models across three increasingly complex scales:

• Static representations. Masked-prediction SSL learns transcriptomic representations that transfer across donors, tissues, and technologies because the objective forces the model to encode gene–gene regulatory dependencies rather than technical correlations (Nature Machine Intelligence, 2024).
• Compositional multimodal models. Paired multimodal data is orders of magnitude scarcer than unimodal atlases. I work on compositional foundation models that fuse frozen unimodal experts through learned interfaces (Cell Systems, 2026), and show theoretically and empirically that standard alignment objectives collapse to linear redundancy — motivating synergy-aware integration measured with the Synergistic Information Score (SIS) (preprint, 2026).
• Cellular dynamics. I extend self-supervision into time with hybrid symbolic-neural models and flow matching, where orthogonality and sparsity constraints recover identifiable, mechanistically meaningful vector fields from snapshot data (Communications Biology, 2026).

Together, these projects target a single thesis: representations that capture predictive structure transfer; representations that merely compress distributional structure do not. This is the foundation I am building toward predictive, in silico models of the cell.

Beyond research

I co-organize the Learning Meaningful Representations of Life (LMRL) workshop at ICLR (2025 Singapore, 2026 Rio de Janeiro), and I have co-organized the Explainable ML in Bioinformatics workshop at ECCB. At TUM I teach the Statistical Learning master’s course (SS24, WS25) and organize the Deep Learning Seminar (since WS24, currently SS26). I mentor interns, working students, and thesis students at Helmholtz Munich. Through the Manage&More scholarship at UnternehmerTUM I work on innovation projects at the intersection of AI and industry, and I volunteer with KI macht Schule! to teach AI in high schools.

Feel free to reach out about research, collaborations, or PhD/internship questions — you can find my email and socials below.