Poster #85: Hameed Sanusi MS

tatCNN: A Multi-Input Convolutional Neural Network for Predicting mRNA Targets of Transfer RNA-Derived Fragments (tRFs)

Hameed Sanusi, MS¹; Sathyanarayanan Vaidhyanathan¹; Sarthak Chandervanshi¹; Andrey Grigoriev, PhD¹ ¹Center for Computational and Integrative Biology, Rutgers University-Camden, Camden, New Jersey, USA

Transfer RNA-derived fragments (tRFs) are a class of small non-coding RNAs with growing interests into their functions. tRFs are generated small non-coding RNAs increasingly recognized as functional molecules rather than random products of tRNA degradation. CLASH (crosslinking, ligation and sequencing of hybrids) enables direct, high-throughput recovery of RNA-RNA hybrids associated with Argonaute complexes, providing evidence of guide-target interactions [. Some tRFs associate with Argonaute proteins and can guide sequence-dependent post-transcriptional regulation through mechanisms akin to miRNA-mediated silencing. Through CLASH experiments, chimeric reads provide evidence of small RNAs and their interactions with their targets. Previous work from our laboratory developed a computational framework for identifying and characterizing AGO1-associated tRF-target hybrids in CLASH data. Using this framework, we assembled 119,738 unique CLASH-supported tRF-mRNA hybrids containing guide and target sequences, guide lengths, transcript-region annotations, orientation, pairing information and read support. We then developed a CNN model to investigate whether CLASH-supported mRNA targets can be distinguished from presumed non-target sequences. Motivated by experimental evidence of tRF-mediated regulation through complementary sites in 3’ UTRs, and by previous observations that CDS and 3’ UTR regions are frequent tRF target locations, we compared two constructed-negative sourcing strategies. In the global strategy, presumed negative target sequences were sampled from all mapped transcript regions, including the 5’ UTR, 3′ UTR, coding sequence and introns. In the region-restricted strategy, presumed negative target sequences were sampled exclusively from sequences mapped to the 3′ UTR. Experimentally observed positive interactions remained unchanged in both experiments. For tRFs found represented in the CLASH dataset, tatCNN distinguished observed interactions from constructed guide and target pairs. We found that 3’-UTR negative sourcing improved AUPRC and F1-score for 13 of 14 individually modeled lengths and 10 of 11 pooled-model length subsets.