Global transcriptome analysis reveals circadian control of splicing events in Arabidopsis thaliana

Abstract

SUMMARYThe circadian clock of Arabidopsis thaliana controls many physiological and molecular processes, allowing plants to anticipate daily changes in their environment. However, developing a detailed understanding of how oscillations in mRNA levels are connected to oscillations in post-transcriptional processes, such as splicing, has remained a challenge.Here we applied a combined approach using deep transcriptome sequencing and bioinformatics tools to identify novel circadian regulated genes and splicing events.Using a stringent approach, we identified 300 intron retention, 8 exon skipping, 79 alternative 3’ splice site usage, 48 alternative 5’ splice site usage, and 350 multiple (more than one event type) annotated events under circadian regulation. We also found 7 and 721 novel alternative exonic and intronic events. Depletion of the circadian regulated splicing factor AtSPF30 homolog, resulted in the disruption of a subset of clock controlled splicing events.Altogether, our global circadian RNA-seq coupled with an in silico, event centred, splicing analysis tool offers a new approach for studying the interplay between the circadian clock and the splicing machinery at a global scale. The identification of many circadian regulated splicing events broadens our current understanding of the level of control that the circadian clock has over this posttranscriptional regulatory layer.

Interpreting machine learning models to investigate circadian regulation and facilitate exploration of clock function

“…MetaCycle is one of the most well-maintained and accessible tools within the community incorporating a variety of the most widely used methods ARSER [35], JTK_CYCLE [36] and Lomb-Scargle [37] and integrating their results so that rhythmic prediction is a cumulation of different statistical approaches. We ran MetaCycle (see Methods) on a published Arabidopsis time-series transcriptomic dataset generated by [8], which was sampled every 4-hours for 48-hours, starting 24-hours after transfer to constant conditions (LL) (Table S1). The data was processed to produce normalized counts per transcript (see Methods).…”

“…We trained a series of ML classifiers to predict if a transcript was circadian or non-circadian in a binary classification system using the derived k-mer profiles for the same set of transcripts and MetaCycle derived labels used previously (for the transcriptomic ML model). Across the range of k-mers the best models were consistently generated with the classifier LightGBM and the most accurate model used a k-mer length of 6 to generate separate feature sets for the promoter and mRNA regions (8,192 features of k-mer counts per transcript) that were both inputted into the model (see Methods). This best optimized model showed ( Figure 3a, Table S2): a mean F1 score of 0.766 on cross validation (standard deviation 0.006) and a test F1 score of 0.751 on class 0 (non-circadian) and 0.804 on class 1 (circadian).…”

“…This has yielded a plethora of public multi-omic resources [5,6,7] that can be re-analysed to give new insights into the roles and functions of complex regulatory networks. In this study, we use newly generated datasets, published temporal datasets [8,9,10] (Table S1) and…”

“…This has yielded a plethora of public multi-omic resources [5,6,7] that can be re-analysed to give new insights into the roles and functions of complex regulatory networks. In this study, we use newly generated datasets, published temporal datasets [8,9,10] (Table S1) and Arabidopsis genomes, in combination with machine learning (ML) approaches (see Glossary for definitions of terms), to make predictions about circadian gene regulation and expression patterns. Critically, we advance existing approaches using explainable AI algorithms and interpretation of our models (Glossary), such methods help us to understand the predictions made by ML models.…”

“…These additional unrelated test datasets represent different sampling strategies and experimental setups (see Methods). Both test datasets were processed bioinformatically as per our original [8] dataset (see Methods). For the Arabidopsis [9] dataset, the timepoints did not match those used to train our model; sampling started 2 hours after exposure to constant light (rather than 24 hours after) and samples were taken every 3 hours instead of every 4.…”