Key Moments
Mini Brains, Beating Heart Organoids, and the AI That Predicts Them — Fabian Theis, Helmholtz Munich
Latent Space PodcastKey Moments
Scientists are developing 'virtual cells' that can predict cellular behavior, but their complexity and the vastness of biological data mean simple linear models still often outperform them, suggesting that our understanding of biological systems is still nascent.
Key Insights
The $1.8 billion Biohub initiative, involving Google, Meta, and NIH, aims to make biology predictable by collecting and open-sourcing multi-omics data for virtual cell development.
Despite advances, linear models and even mean predictions can be surprisingly difficult to beat when predicting cellular responses to perturbations, indicating challenges in current predictive models.
Current 'virtual cell' development is heavily focused on predicting cellular responses to 'perturbations,' which can range from gene knockouts to drug treatments or changes in cell culture conditions.
Organoids, such as mini-brains and beating heart organoids, are being used as more complex, multi-cellular systems to study biology and disease, overcoming limitations of single-cell studies.
The 'Lab Compass' is a novel reverse generative model that uses an existing forward model to suggest experimental interventions needed to reach a desired cellular state.
The development of AI in biology is moving towards integrating experimental labs into a closed-loop reinforcement learning system, aiming to continuously guide experiments based on model predictions.
The push for predictable biology and the role of virtual cells
The conversation highlights a significant push in the scientific community, exemplified by the $1.8 billion Biohub initiative, to make biological systems predictable. This ambitious goal involves generating and sharing vast amounts of multi-omics data to develop 'virtual cells'—computational models that can simulate cellular behavior. Despite advancements in AI and machine learning, researchers are finding that predicting the outcome of perturbing cells remains a significant challenge. Surprisingly, simple linear models or even basic average predictions often outperform more complex deep learning models, suggesting that our current understanding and modeling of biological complexity is still in its early stages. This underscores the need for better benchmarks and potentially new approaches to model biological systems.
Understanding cellular perturbations and the 'holy grail' of prediction
A core focus in current biological research, particularly within startups, is the concept of 'perturbations.' This refers to any change introduced to a cell or system that alters its state, such as knocking out a gene using CRISPR, adding a drug, or modifying the cell culture environment. The ultimate aim is to predict how a cell, with its complex molecular machinery, will respond to these various inputs. This predictive capability is considered by many to be the 'holy grail' of computational and systems biology. While the term 'virtual cell' is often equated with predicting perturbations, it's a simplification. The true challenge lies in understanding the cell's dynamic state across its genome, epigenome, and proteome, and then predicting how this state will change under different conditions, including disease or aging.
The evolution of data generation: from qPCR to organoid experiments
The ability to model cellular behavior is intrinsically linked to the quality and scale of data available. Early work in transcriptomics, using techniques like qPCR and later single-cell RNA sequencing, provided a foundational understanding. A significant leap occurred with technologies enabling the measurement of millions of genetic transcripts in a single experiment. This expanded the scope from simply counting cells to understanding the 'cell space' and, more importantly, the 'perturbation space.' Historically, the perturbation space was sparse, often limited to comparing control versus disease states. However, recent advancements have dramatically enriched this space, with experiments involving hundreds of cell lines and thousands of drug treatments. This data explosion is enabling more sophisticated predictions and is driving the development of new experimental systems, including organoids.
Organoids as a new frontier for biological study
Organoids, described as miniature organ-like structures grown from cells in a lab, represent a significant advancement in studying complex biological systems. Researchers can now grow 'mini-brains' or 'beating heart organoids' from stem cells, allowing them to observe multi-cellular interactions and behaviors that are impossible to study in single-cell cultures. These organoids, while still facing limitations like the absence of a full immune system or vascularization, provide a more physiologically relevant model for investigating diseases, including neurodegenerative disorders and cancer. Their development is closely tied to the concept of the perturbation space, as scientists can introduce perturbations to these organoids to observe their complex responses.
The modeling challenge: linear models still hold their ground
Despite the allure of deep learning and complex AI models, a recurring theme is the persistent performance of simpler, linear models. In tasks like cell type prediction, logistic regression often suffices. Even in perturbation prediction, linear transformations in latent space have shown remarkable efficacy. This suggests that while complex models might capture intricate patterns, the sheer scale and inherent noise in biological data mean that simpler, more interpretable models can be equally, if not more, effective. The challenge lies not just in building complex models, but in ensuring they are robust and generalizable, and that the fundamental biological questions are well-posed for the chosen metrics.
Bridging the gap: 'Lab Compass' and guided experimentation
To address the challenge of navigating the vast biological landscape, new tools are emerging. One such innovation is 'Lab Compass,' a reverse generative model. Once a 'forward model' (a predictive model of the biological system) is established, Lab Compass can suggest specific experiments or interventions needed to steer the system towards a desired state. This moves beyond simply predicting outcomes to actively guiding the scientific process. The goal is to create a continuous learning loop, where experiments inform models, and models, in turn, suggest the next best experiment, accelerating discovery.
The future of AI in science: Reinforcement learning and closed-loop labs
The ultimate vision is to apply reinforcement learning (RL) to the entire scientific discovery process. This involves creating 'closed-loop' experimental systems where AI continuously directs research. Instead of researchers developing a model and then starting a new project, the model would suggest the next experiment, the results would be fed back to refine the model, and this cycle would repeat. While applying full RL to complex biological experiments is ambitious, the progress in areas like molecule design and material science, which already utilize iterative design-build-test-learn cycles, suggests this is an achievable long-term goal. Current efforts are focused on integrating computational models with experimental labs, aiming to make the research process more efficient and data-driven.
Funding, innovation, and the German startup ecosystem
The discussion touches upon the funding landscape for scientific research and startups, contrasting the German and US models. Germany's Helmholtz association, a large, federally funded research organization, allows for long-term, in-depth research, particularly in health and computational biology. While Germany has a growing startup scene, especially in Munich, a common bottleneck is the 'scale-up' phase, where venture capital for later-stage funding is less abundant compared to the US. The EU's efforts to standardize regulations, like the AI Act and the European Health Data Space, aim to create a more unified and data-accessible environment. The emphasis is shifting towards integrating computational expertise with wet-lab research, fostering a new generation of scientists and companies that can bridge these disciplines.
Mentioned in This Episode
●Software & Apps
●Companies
●Organizations
●Books
●Studies Cited
●Concepts
Common Questions
Virtual cells are computational models that simulate cellular behavior. They are used to predict how cells will respond to various disturbances and aim to understand complex biological systems, acting as a 'holy grail' in computational and systems biology.
Topics
Mentioned in this video
Fabian Theis leads AI and computational biology efforts here.
Mentioned as having a model similar to the Helmholtz Centers' applied approach, but distributed nationwide.
A consortium where farmers can join to gain early access to perturbation datasets.
A major project founded by the speaker, involving various disciplines beyond biology.
One of the top universities in the EU that the speaker's institution interacts with.
Announced a significant collaboration with US DOE, NIH, Meta, and Google for biological data.
Mentioned for its initiatives in structural biology and biopharma, focusing on molecule design.
Mentioned for its initiatives in structural biology and biopharma, focusing on molecule design.
Mentioned as a company that is an example of generating data, particularly with access to data.
A company where Brandon Anderson develops RNA therapies.
Collaborating with Biohub on collecting and publishing multi-scale biological data.
Launched a technology called 'MegaLab' allowing for the measurement of 10 million gene copies in a single experiment.
Mentioned as a company visited by the speaker, highlighting the startup culture in the Bay Area.
An example of a company that aligns with the virtual cell and systems biology approaches.
Hired two members of the speaker's team to work on computational and biological aspects of their development.
The company where RJ Honicky is the CTO, host of the "AI for Science" podcast.
An example of a company that aligns with the virtual cell and systems biology approaches.
Collaborating with Biohub on collecting and publishing multi-scale biological data.
A preprint server where the Lab Compass paper is expected to be available.
An open-source data structure for single-cell data, likely used by many computational biologists.
A tabular baseline model developed by the speaker's team showing that non-linearity adds value beyond the baseline.
Nobel Prize-winning factors used to convert somatic cells into induced pluripotent stem cells.
A collection of open-source tools developed from work by Fabian Theis's lab.
A gene-editing technology used to remove a gene from the genome, leading to no expression.
An open-source tool likely used by most computational biologists, developed in Fabian Theis's lab.
A non-linear model developed in Fabian Theis's lab for disease prediction using machine learning in 2019.
A novel reverse generative model that guides experimental direction by suggesting actions to reach a target state.
Software developed by the speaker's team around 2019, introducing linear transformations in latent space.
More from Latent Space
View all 275 summaries
51 minCooking with Reducto’s CEO: Why PDFs Still Break AI — Adit Abraham
73 minBehind Latent Space, ft AIEi Shanghai preview — with swyx and Galen Gao
35 min⚡️ The 2026 Software Factory Infra Stack — Zach Lloyd, Warp.dev
68 minBuilding Legal Agents When You Can't See the Data — Jacob Lauritzen, CTO of Legora
Ask anything from this episode.
Save it, chat with it, and connect it to Claude or ChatGPT. Get cited answers from the actual content — and build your own knowledge base of every podcast and video you care about.
Get Started Free