Virtual Cells Need Context, Not Just Scale.

Area of Science:

• Computational Biology
• Artificial Intelligence
• Genomics

Background:

• The field of AI in biology is rapidly advancing, aiming to create computational models of cells ('Virtual Cells') capable of predicting cellular responses.
• Current approaches focus on training large, high-capacity models using extensive single-cell data, inspired by successes in structural biology and large language models.

Purpose of the Study:

• This paper argues that simply scaling model capacity is inadequate for solving the Virtual Cell problem.
• The primary limitation identified is insufficient coverage across diverse biological contexts, rather than a lack of model expressivity.

Main Methods:

• Reviewing recent studies comparing simple baselines with sophisticated architectures within specific biological contexts.
• Analyzing the generalization capabilities of current models across different biological contexts.
• Connecting findings to causal inference literature on transportability.
• Examining a state-of-the-art model using a large-scale (22 million cells) immunology dataset.

Main Results:

• Simple baseline models perform comparably to complex architectures when evaluated within a single biological context.
• Existing models demonstrate poor generalization performance when applied to new or different biological contexts.
• Analysis of a large immunology dataset reveals limitations in current state-of-the-art models.

Conclusions:

• The Virtual Cell problem is fundamentally a 'causal transport problem', requiring more than just increased data from similar distributions.
• Future progress necessitates a greater emphasis on contextual diversity and causal representation learning.
• These approaches should complement, not replace, ongoing efforts to scale model capacity and data volume.