Gene-level structure of prior knowledge and perturbation responses informs target-level predictability in prior-informed models

Read the full article See related articles

Listed in

This article is not in any list yet, why not save it to one of your lists.
Log in to save this article

Abstract

Background

Predictive models of transcriptional responses to unseen genetic perturbations increasingly incorporate biological prior knowledge. However, it remains unclear how well prior knowledge reflects each target’s perturbation-response relationships and how closely prior-derived measures track target-level predictive performance. Here, we develop a gene-similarity framework to quantify agreement among eight sources of biological prior knowledge and between prior knowledge and co-function or co-regulation relationships derived from six Perturb-seq datasets.

Results

In genome-wide and essential-gene screens, prior agreement accounted for 30–47% of the variance in co-function agreement and 35–43% in co-regulation agreement. Agreement with Perturb-seq responses was higher for essential genes, lower for transcription factor and signaling genes, and varied with gene expression characteristics and cellular context. We compared prior agreement and prior-derived train–test similarity with measures based on the observed test response in tracking target-level predictive performance. Across three prior-informed models, predictions were more accurate for targets whose observed perturbation-response relationships better matched prior knowledge and that were more similar to training perturbations. Notably, prior agreement and prior-derived train–test similarity, calculated without the test response, tracked relative differences in target-level predictive performance. The strength of these associations varied across models and datasets and, for train–test similarity, across prior sources.

Conclusion

Our framework reveals gene- and context-level heterogeneity in how biological knowledge aligns with perturbation responses, and provides an a priori basis for assessing relative perturbation predictability before perturbation responses are measured.

Article activity feed

  1. Authors used A-549 from @cellosaurus.bsky.social in their study. Including #RRIDs will make this less ambiguous. SciScore made a table with this resource, see “Automated Services” module (download as csv, xml or #jats) #OpenScience #RRID

  2. Gene-level structure of prior knowledge and perturbation responses informs target-level predictability in prior-informed models https://www.biorxiv.org/content/10.64898/2026.11.08.765139v1