Background: Reading changes before aggregates appear

Early population images of the soil bacterium Myxococcus xanthus contained clues that could predict whether it would later aggregate. Published in the Proceedings of the National Academy of Sciences (PNAS), the study used deep learning to quantify subtle spatial patterns in microscopy images and extract information associated with developmental outcomes before visible aggregates appeared.

When nutrients are depleted, this bacterium begins a multicellular developmental process in which cells gather to form fruiting bodies. Earlier work quantified aggregate structure and formation timing, but it was difficult to compare pre-aggregation cell flows, immature aggregates with indistinct boundaries, and mutants that did not aggregate at all using the same criteria. Because changes emerged at different times in different strains, comparing entire developmental trajectories rather than appearances at a single time point became the study’s starting point.

Key findings: Compressing population images into 13-dimensional features

The researchers collected 937 bright-field time-lapse image sequences from 292 genetically distinct strains. Imaging began within 15 minutes of the onset of nutrient deprivation and continued at one-minute intervals for 24 hours. The unit of analysis was the population’s image pattern, not individual cells’ movement trajectories.

The model combined a residual neural network (ResNet) for extracting image features and StyleGAN2, an image-generation model, within a variational autoencoder (VAE). A Siamese network was trained alongside it to learn phenotypic similarity between pairs of images. Each image was compressed into a 13-dimensional feature vector, and changes in the vector over time represented a developmental trajectory. The learned features were also associated with aggregate area and number.

Could these features distinguish different motility defects? The researchers used a support vector machine (SVM) classifier to distinguish strains with A-motility defects from those with S-motility defects. In three-fold cross-validation split by strain, the area under the receiver operating characteristic curve (ROC-AUC) using the first observed image was 0.91 ± 0.06. This measures how well the two defect groups were distinguished; it does not mean that every genetic variant can be identified.

Among 352 wild-type image sequences, 220 showed visible aggregates by 24 hours, while 132 did not. In three-fold cross-validation split by experimental replicate, the ROC-AUC for distinguishing subsequent aggregation outcomes using only the first observed image was 0.82 ± 0.04. Values in both analyses were reported as means with 95% confidence intervals. ROC-AUC measures discrimination between groups and is not the same as accuracy.

Meaning and outlook: Predictive power in early patterns, and open questions

Images reconstructed from early features revealed subtle local differences in cell density and arrangement. The researchers interpreted these as early information associated with subsequent development, but whether the patterns directly induce aggregation or serve as markers of pre-existing cell states or environmental differences remains unresolved.

Wild-type aggregation outcomes tended to be similar among experiments that used the same agar batch and culture stock on the same day. The conditions responsible for the differences have not been identified, making it difficult to conclude that early patterns are a causal mechanism determining developmental fate. Nor does changing latent features to alter the shapes in generated images demonstrate that desired populations have been designed through actual genetic modification.

This method may serve as a research tool for comparing mutant developmental trajectories and detecting abnormal phenotypes that are easy to miss by eye. Extending it to images of other biofilms or tissues will require validation for each target. The research methods and results are available in the full PNAS article.