Fabian Theis discusses the integration of advanced AI models, such as inverse generative models and reinforcement learning, to predict and guide biological experiments, particularly in complex systems like organoids, enhancing the understanding of cellular behavior through perturbations. He also highlights the evolving biotech ecosystem in Europe, emphasizing interdisciplinary collaboration, robust modeling approaches, and the potential of AI-driven automation to transform biological research.
In this insightful discussion, Fabian Theis, a leading figure in computational biology at Helmholtz Munich, explores the evolving landscape of virtual cells, perturbation prediction, and the integration of AI in biological research. He highlights the transition from merely mapping human cell atlases to predictive modeling and now to actively steering experiments through advanced AI models. Theis introduces the concept of “lab compass,” an inverse generative model designed to guide experimental steps based on forward world models, emphasizing the potential of reinforcement learning to optimize biological experimentation cycles.
The conversation delves into the significance of perturbations in understanding cellular behavior. Perturbations, such as gene knockouts via CRISPR or environmental changes, serve as inputs that alter cell states, enabling researchers to observe resultant phenotypic changes. Theis underscores the complexity of modeling these effects, especially as research moves beyond single cells to multicellular systems like organoids—miniature, simplified versions of organs grown in vitro. These organoids provide richer, more physiologically relevant data, crucial for studying diseases and drug responses in a controlled environment.
On the modeling front, Theis discusses the challenges and advancements in representing complex biological data. While early models often relied on linear assumptions, recent efforts incorporate nonlinear and generative approaches, such as flow matching, to better capture the dynamics of cellular responses to perturbations. He stresses the importance of robust baselines and appropriate metrics in evaluating model performance, referencing competitions like the Virtual Cell Challenge that have shaped the field’s understanding of effective modeling strategies.
Theis also reflects on the broader research ecosystem, contrasting European and US funding and startup environments. He notes Europe’s strength in early-stage startup creation but acknowledges challenges in scaling ventures due to limited late-stage venture capital. Despite bureaucratic hurdles, regions like Munich are emerging as vibrant hubs for biotech innovation, supported by initiatives that foster interdisciplinary training—exemplified by the Munich School for Data Science’s “Y model,” which pairs students with both computational and wet lab mentors to bridge domain knowledge and methodological expertise.
Finally, the discussion touches on the cultural and practical aspects of integrating computational and experimental biology. Theis emphasizes the need for close collaboration and trust between modelers and experimentalists to realize the full potential of AI-guided research. While automation and robotics hold promise for accelerating experimental cycles, current assay durations pose challenges. Nonetheless, the vision of fully integrated, reinforcement learning-driven lab environments is gradually becoming attainable, promising a future where AI not only predicts but actively directs biological discovery.