GenBio proposes a multilevel AI model of an entire organism
Nature Medicine has published a plan for an AI system designed to trace the effects of interventions from molecules to the whole organism
On 13 August, Nature Medicine published an article outlining AIDO, a system of connected AI models for different levels of biology. The authors propose linking models of DNA, proteins, cells, tissues, and organism traits to trace how a drug or gene alteration might affect each level in this sequence.
When a drug or gene alteration acts on an organism, it first affects molecules. It may then alter gene networks and cell states, followed by tissues and organism traits. The data at each level have different structures. DNA is a sequence of symbols, a protein is defined in part by its three-dimensional shape, and tissue depends on the spatial arrangement of neighboring cells. The authors therefore propose a separate model for each type of data.
AIDO is based on specialized foundation models. Each model is first trained on a large collection of one type of data and then adapted to a specific task. Some models are intended to read DNA and RNA sequences. Others relate protein structure to protein properties. A third group describes cell states or changes in organism measurements over time.
The authors then plan to connect these models using known biological relationships. A cell uses DNA to produce RNA, and the instructions in RNA guide protein synthesis. A gene regulatory network describes how the activity of some genes changes the activity of others. When this network is linked to cell state data, the model receives both information about an intervention and a possible path through which that intervention could alter the cell. The authors plan to extend the same chain from cells to tissues and organism traits.
The authors propose training the connected models together. A prediction at the organism level should modify the settings of the cellular and molecular models, while data from those models should refine the higher-level prediction. To model a change in cell state, the system receives the initial state and an intervention, specifically a drug, a gene alteration, or a combination of the two. It then calculates a possible response through the gene regulatory network. The protein workflow uses a different cycle. The system proposes protein variants and evaluates their structure, stability, and ability to bind other molecules. A laboratory then tests selected variants and returns the results for the next training cycle.
In the open preprint of the study from 2024, the authors report that they evaluated individual AIDO components on more than 300 tasks in total. The new article describes the project's next step: connecting these components into a system in which a prediction about an intervention moves from the molecular level to cells, tissues, and organism traits.
In a test of AI “virtual cell” models that predict a cell's response after a gene is disabled, most approaches did not outperform a simple baseline model in predicting the cell's complete response. Some models did, however, capture the direction of changes and identify key genes. The article presents AIDO as an engineering program whose aim is to connect specialized models and test whether they can produce an integrated prediction from an intervention to organism traits.