We build virtual cell models for Turbine.

We are the research arm of Turbine, focused on understanding and predicting cell behavior beyond what current experiments and models allow.

This site is where we share how we think about the state of the field, publish what we learn and invite others to work with us on key problems.

Our thinking has been shaped by a decade of building computational models of perturbation biology (Virtual Cells) for drug discovery within Turbine. Our models are successful if and only if they improve our understanding of biology and influence real decisions in drug discovery. This gives us a different worldview from purely academic groups.

We are working towards a general model of biology.

Identical twins demonstrate how reproducible biology is in nature. Lab experiments may struggle to measure or reproduce certain effects, but we believe the biological programs encoded in DNA through evolution will eventually become learnable.

This would make it possible to predict (and eventually exploit) latent biological behaviors beyond what any individual assay can directly observe. The models capable of such predictions would be general Virtual Cells; something close to an "AGI for biology".

The state-of-the art today is Virtual Assays - narrow virtual cells grounded in specific assays.

The predictive power of current cell models is fundamentally constrained by the conditions under which their training data was generated.

Today’s models can often generalize to related biological contexts and sometimes even to unseen perturbations within an assay. But every assay still needs to be grounded individually in data.

This is why we call the models released at Turbine Virtual Assays.

We advance the field by connecting diverse datasets instead of generating large homogeneous ones.

To bridge the gap from narrow to general Virtual Cells, we believe the missing piece is understanding how cells behave across different biological and experimental contexts as opposed to scaling within a single protocol.

We therefore specialize in processing and integrating disparate data into coherent modeling architectures.

Turbine's wet-lab is also optimized for agile data generation, capable of producing new ML-grade seed data monthly instead of being locked in on one single data type.

Pushing the limits of biological understanding is a shared effort.

We believe advancing computational biology quickly through combined effort is more important than racing to be the first in isolated silos. We aim to share our findings and thinking quickly, including the rough edges.

Since many interesting results do not naturally fit into polished publication narratives, we frequently use non-traditional publication formats.

We are happy to collaborate if you share our pains (see our Open problems) or have an interesting dataset which could benefit from our integration expertise.