Fabian Theis on Why Hardcoded Biology Loses to Pure Data
Fabian Theis explains why biological inductive biases fail at scale and when hybrid architectures still matter in perturbation biology.
Mini Brains, Beating Heart Organoids, and the AI That Predicts Them — Fabian Theis, Helmholtz Munich
Fabian Theis, Director of the Computational Health Center at Helmholtz Munich, joins co-hosts RJ Honicky and Brandon Anderson to discuss the evolution of virtual cells, single-cell transcriptomics, and perturbation biology. They explore how large-scale multiscale biological datasets, foundation models like scConcept and Nicheformer, and generative modeling techniques like CellFlow are enabling predictive cellular world models. The conversation also examines the challenges of baseline models, closed-loop active learning with LabCompass, the European biotech startup ecosystem, and interdisciplinary graduate training frameworks.
Key takeaways
- Hardcoded biological assumptions hurt model performance once data expands. Theis points out that decades of machine learning show rigid a priori constraints consistently lose predictive power to flexible architectures. Read more →
- Software-only biology platforms struggle to survive because venture investors eventually force them to pivot into internal drug discovery pipelines to prove financial returns. Read more →
- Simple linear models and mean predictors routinely beat complex neural networks on noisy single-cell biology benchmarks. Read more →
- A virtual cell functions as an input-output predictor rather than an observational atlas. Read more →
4 articles from this episode
Virtual Cell Platforms Fail Without Wet-Lab Data
Fabian Theis explains why computational biology platforms pivot to drug assets and how proprietary wet-lab data creates defensible value.
Simple Baselines Beat Deep Models in Cell Biology
Fabian Theis explains why simple linear baselines beat deep learning in the Arc Virtual Cell Challenge, and when complex models actually work.
Fabian Theis on Why Virtual Cells Mean Perturbation Prediction
Fabian Theis explains why pharma needs perturbation prediction models across 1,000 drugs and CRISPR screens instead of passive cell atlases.
More Latent Space episodes
- Reducto Final YouTube 1080p Sandwich TitleNov 8 · 4 articles
- ⚡️ The 2026 Software Factory Standard — Zach Lloyd, Warp.devNov 8 · 6 articles
- Deep Dive: Anthropic S-1 Myths and Mythos — Jamin Ball, AltimeterNov 8 · 4 articles
- Behind Latent Space, ft AIEi Shanghai preview — with swyx and Galen GaoNov 8 · 5 articles