Explainable Deep-Learning on condition specific expression profiles reveals critical cytosines for gene regulation in plants

Abstract

Compared to other nucleotides, cytosines occupy a unique regulatory position in plant genomes through methylation-based epigenetic control. Measuring the influence level of each cytosine on downstream gene’s expression is a highly promising question which would result in precise control intervention which still remains hardly attempted. A deep learning framework, CritiCal-C, is being presented here which addresses this question while building upon a large volume of multimodal data including 232 WGB-sequencing and 260 corresponding RNA-seq datasets. The model successfully identifies the critical cytosines in the 2kb promoter regions which exert significant influence on downstream gene’s expression. A systematic knockout and Explainable AI (XAI) attribution approach applied to the deep-learning model helped here. Strong condition-specific relationships between promoter methylation patterns at the critical cytosine and downstream gene’s expression levels were established (Spearman ρ > 0.90), enabling systematic annotation of critical cytosines genome-wide in the considered species. During its cross-species application this was found that instead of sequence homology it is GC% similarity as the more appropriate of similar regulatory patterns. With the same philosophy the model worked at the similar level of consistent performance in cross-species manner, establishing it as a universal model to identify critical cytosines in any plant species. The reported criticality discovery by the raised deep model was also validated experimentally. It will be very helpful in designing far efficient, frugal, targeted methylation and editing based molecular interventions. The raised tool and resource are freely available at: https://hichicob.ihbt.res.in/critical/