Controlling Biology - Fabian Theis, Helmholtz Munich

Judgment

Fabian Theis’s decision to move from mapping to controlling biology should shift his lab’s modeling priority toward multi-scale, co-registered data and scale interfaces, accepting slower validation and intervention risk rather than chasing perturbation-prediction benchmarks.

Basis

Theis argues that perturbation prediction on existing datasets is not a hard enough problem to justify complex algorithms. The source notes that in the 2025 Arc Virtual Cell Challenge, constant baselines often beat sophisticated models for perturbation prediction, though poorly calibrated metrics can make uninformative mean predictions look better. He instead points to Biohub’s $1.8B commitment from Biohub, NIH, DOE, Google and Meta to generate large, high-precision, multi-scale, co-registered biological data across molecules→cells→niches→tissues→organs→organisms. The mechanism is that richer data and better-calibrated metrics could let models learn scale interfaces rather than overfit weak perturbation benchmarks, but the source also says we lack understanding of those interfaces and the data.