Correction for potentially confounding features of MT

The confounding features of 3′UTR length, local AU content, TA, and SPS might also be potentially confounded with the RBP binding so need to be carefully controlled. To rigorously correct these confounding effects, we have devised an in-house algorithm facilitating to sample 3′UTRs into four different groups only according to d_MTS-RBS while keeping the confounding effect similar across the groups. 3′UTRs containing a single 7, 8mer MTS and one or more RBSs were defined as the input of sampling process. The sampling process comprises of multiple rounds of selection. For a round of selection, a set of four 3′UTRs was selected with the following criteria: (i) In a same set, the site type of MTSs should be identical to one of the 8mer, 7mer-m8, and 7mer-A1, (ii) The min–max range of confounding effect should fall within the specified cutoffs, that were determined by the Gaussian process implemented in the ‘bayes_opt’ Python package to maximize the number of selected 3′UTRs, (iii) One 3′UTR should have an overlapping MTS with the nearest RBS (d_MTS-RBS = 0) and other three should have MTS separated from the nearest RBS (d_MTS-RBS > 0). 3′UTRs which met the criteria above were then assigned to the first to fourth groups in ascending order of the value of d_MTS-RBS. Next rounds of selection of four 3′UTRs proceeded until the inputs were exhausted. If any confounding factor exhibited significant difference across the groups, the sampling process was retried with the narrower cutoffs mentioned at the second criterion. After the sampling process finished, comparison of the log₂(fold change) or the confounding features across the groups was performed using Wilcoxon’s rank-sum test of ‘SciPy’ package in Python. Accordingly, we can clearly divide 3′UTRs solely dependent on d_MTS-RBS without any concern that the confounding effect involves across groups. By using the dataset of DICER or DROSHA knockout in HCT116 cell lines³⁵, we performed the similar analysis to detect the derepression of target mRNAs. To select miRNAs whose targets show the strongest derepression in response to miRNA removal, a 2 × 2 contingency table was constructed for each miRNA by examining whether its MTS was included in the 3′UTR and whether the 3′UTR was highly de-repressed. The most significant five miRNAs after χ² tests were chosen and used for the analysis.