TERRA: a 112.6M-cell spatial model turns gene inactivation into testable tissue hypotheses

The TERRA preprint appeared on bioRxiv on August 4. The model was trained on spatial measurements from 112.6 million cells across 636 sections of 20 human tissues. Spatial transcriptomics captures which genes are active in a cell and where that cell sits in an organ section. The same gene can behave differently depending on which cells surround it. TERRA takes the active genes of a central cell together with up to ten neighboring cells ordered by distance and reconstructs a stable latent representation of the tissue region. During training, the model masks some genes and predicts the latent vector of the masked portion from the remaining genes and neighboring cells. By "predicting changes," the authors mean shifts in the space of these latent representations.

TERRA-96M, trained without 215 sections, achieved the highest agreement with annotated cellular niches (recurring combinations of neighboring cells) among the spatial models the authors included in their analysis. The evaluation used four held-out datasets: new samples, new donors, a new dataset, and a new measurement technology.

To test the approach, the authors computationally inactivated genes with established roles in different kidney regions. The strongest changes appeared in the corresponding cells and their neighborhoods. They then inactivated CTLA4 and PDCD1, which encode the targets of ipilimumab and nivolumab (cancer immunotherapy drugs), in nine kidney sections from five patients collected before treatment. After the intervention, the spatial gene representations in those sections were closer to three real post-therapy kidney sections than the original representations (without virtual inactivation) or representations after control interventions (inactivation of random or housekeeping genes). From the predicted shifts, the authors selected 23 genes that also differed between tissue collected before and after therapy. In the post-treatment sections, this gene program was concentrated in regions containing clusters of immune cells. In public data from blood immune cells of six lung cancer patients who received pembrolizumab, the program was more strongly expressed in the three with documented kidney injury associated with this therapy.

Comparison with real tissue is a stricter test than analysis of similarity between the model's internal numerical representations. An analysis of errors in virtual cell models showed that an evaluation method can give high scores to the prediction of an average profile even when that prediction misses the changes that motivated the experiment. TERRA makes it possible to formulate a testable hypothesis about how a change in one gene alters the immediate environment of nearby cells and to check that hypothesis against real spatial data. The weights and code for the full TERRA-112M model are public.

Open the related Eternal Search page

Sources
[1] t.me
[2] doi.org
[3] t.me