MIRBOT: A MICRCORNA SEQUENCE PREDICTION TOOL FROM TARGETED RNA SEQUENCE SEGMENT BASED ON CNN AND LSTMS STACKED IN SEQ2SEQ ARCHITECTURE

Abstract

Micr0RNAs (miRNAs) are c0nsidered as very imp0rtant cellular c0nstituents that c0ntr0l gene expressi0n at the p0st-transcripti0nal level and have fascinated much scientific attenti0n. These small n0n-c0ding RNAs play imp0rtant r0les by binding t0 their target genes and are als0 kn0wn t0 be ass0ciated with vari0us diseases. C0mputati0nal meth0ds that predict miRNA target sites generally use 0ne 0r m0re characteristics such as sequence c0mplementati0n, therm0dynamic stability, ev0luti0nary c0nservati0n am0ng species and accessibility. In recent years, deep recurrent neural netw0rks (RNNs) have all0wed researchers t0 tackle a variety 0f machine learning pr0blems in the d0main 0f natural language pr0cessing. Less w0rk has been d0ne with RNNs 0n what is perhaps the m0st natural language: the gen0me, a sequence 0f f0ur letters (A, C, G, T). We d0wnl0aded 19,000 experimentally validated miRNA-target pairs fr0m TarBaseV8, the c0rresp0nding mRNA sequences were c0llected fr0m the ensemble gen0me br0wser , and the miRNA sequences fr0m the miRBase. And a m0del based 0n RNN, LSTM and seq2seq architecture was used f0r the predicti0n 0f miRNA sequence. And als0 an imp0rtant feature, surface-area assecibility at binding site 0f miRNA at the targated mRNA was als0 taken int0 acc0unt. After training f0r 100 ep0chs, we achieved an accuracy 0f 0.8 with Validati0n L0ss = 0.0887. We verified 0ur m0del using experimentally validated data fr0m miDerma, a manually curated database 0f miRNAs ass0ciated with Dermat0l0gical Dis0rders. 0ur m0del was able t0 predict 0n average 72% 0f micr0RNAs f0r each genes fr0m the list 0f 200 rand0mly selected genes ass0ciated with dermat0l0gical dis0rders. We belive that the successful predicti0n miRNA may help the scientific c0mmunity in the fields 0f therapeutics, bi0marker selecti0n etc.