AI Uses a 4D Map of Mitochondria to Predict Function From Shape Alone

A new technology uses AI to predict a cell’s energy state based solely on the four-dimensional (4D) shape of its mitochondria.

According to BioOn on the 30th (local time), researchers developed a self-supervised learning framework to analyze complex changes in mitochondria responding to drugs without human annotations. The findings were published in the international journal Cell.

The researchers applied the framework to 4D data on drug-treated mitochondria to create a trained model called MitoSpace. The model represents mitochondrial phenotypes in response to various drug stimuli in a latent space.

Notably, drug-induced phenotypes clustered on their own without separate ground-truth labels. A small neural network probe trained on these embeddings accurately predicted image-based morphological features and functional measures.

In particular, the researchers used a separate fluorescence channel excluded from model training to predict mitochondrial membrane potential using only 4D spatiotemporal information. The coefficient of determination (R²) was 0.91, quantitatively demonstrating the relationship between morphology and function at the single-cell level.

MitoSpace showed zero-shot performance, making predictions for drugs it had not been trained on, and also analyzed human lung organoid (LO) data. The researchers said this generalizability is consistent with the characteristics of a foundation model.

The researchers also trained models on 4D, 3D and 2D data separately and compared them. The 4D model achieved the highest classification performance in 22 of 26 conditions. As dimensionality decreased, cluster purity and separability declined.

The researchers said, “This achievement is a first step toward a more general-purpose 4D intracellular reference space,” adding that it “could be extended to disease models and CRISPR-based screening.” The pretrained model, code and 4D dataset have been made public.