Researchers publish roadmap for AI-powered virtual cells

The paper, Multimodal AI for predictive and programmable virtual cells, authored by Mr Krish W. Ramadurai and Professor Abhirup Banerjee, has been published in Trends Open, a Cell Press journal.

For decades, scientists have sought to build computational models of living cells capable of predicting how a cell will respond to a drug, gene edit, or other intervention before laboratory experiments are performed. Such models, often referred to as virtual cells, could help accelerate biological research and therapeutic discovery by reducing reliance on trial-and-error experimentation.

In their review, the authors examine recent progress in the field, driven by developments in single-cell atlases, perturbation screens, spatial transcriptomics, and multimodal artificial intelligence (AI.) They also assess the limitations of current approaches and outline the key scientific advances needed to transform virtual cells from promising research tools into robust and experimentally grounded predictive systems.

The review argues that while existing models can learn valuable representations of cellular states, many remain largely correlative, require task-specific training, and do not consistently outperform simpler approaches. The authors suggest that future progress will depend not only on larger AI models, but also on richer multimodal datasets, causal perturbation information, stronger biological understanding, and continuous experimental validation.

The work reflects the broader vision of the Multimodal Medical Data Integration & Analysis (MultiMeDIA) research programme, which seeks to integrate multiple forms of biological and medical data, including molecular structures, single-cell measurements, imaging, spatial information, and perturbation responses, to develop more comprehensive and predictive AI models.

Krish Ramadurai, first author of the paper, said: “Virtual cells could fundamentally change how biological research and therapeutic discovery are conducted. Instead of relying solely on experimental trial and error, researchers could use multimodal AI models to integrate cellular, spatial, imaging and perturbation data, predict how a cell will respond to a drug or gene edit, and then use targeted experiments to test and improve those predictions."

“This paper defines the path from today’s largely correlative models toward experimentally grounded systems that can predict and ultimately help program cellular behaviour. Its acceptance is especially meaningful, as it captures the ambition at the heart of my doctoral research and the MultiMeDIA programme: to build multimodal AI that does not simply observe biology, but helps us understand, predict and rationally engineer it.”

Readers interested in learning more about the wider MultiMeDIA programme and its research into multimodal medical AI can visit: https://eng.ox.ac.uk/multimedia

Read the publication here.