Analysing alternative polyadenylation in ALS using a predictive neural network
Summary
The predictive neural network model APARENT2 was used to analyse the effect of variants in amyotrophic lateral sclerosis (ALS) patients on alternative polyadenylation (APA). Variants from whole-genome sequencing data of both ALS patients and healthy controls (n=6538, n=2415, resp.) were scored with the APARENT2 model. No difference was found between the frequency of PAS-affecting variants in ALS patients versus healthy controls. A key limitation was the complex regulatory system surrounding pA, making it difficult to predict the effect up- or downregulation of a particular polyadenylation site (PAS) might have on the overall transcripts. This means we cannot exclude the absence of ALS associated PAS altering variants. Attempts to train an ALS specific APA-prediction model on TDP-43 knockdown PAS data did not result in a better-performing model with the training-data available. APARENT2 was used to predict the strength of newly identified PAS’s in motor cortex tissue. The model scored these supposed brain-specific PAS’s similarly to previously identified ones and the sequence around these PAS’s showed enrichment for the CUX1 RNA binding site, which is a transcription factor mainly used in the brain. The outcomes of the APARENT2 model can be used to decide which brain-specific PAS’s are to be used in further research on polyadenylation in the brain.