MetaTIS: A tool to predict eukaryotic translation initiation sites

As the ribosome makes its foray unto a sequence of mRNA, ribosomes typically commence translation at a methionine-encoding AUG codon flanked by a so-called Kozak region, a short nucleic acid motif serving as an initiation site in many eukaryotes. Though, the characteristic AUG start codon of an mRNA is not always effective in initiating translation. In more seldom cases, near-cognate codon sequences may also be recognized as start sites. Ribosome profiling techniques, characterized by the stymieing of mRNA-ribosome complex translation function via chemical treatment, are able to elucidate active translation start sites. Historically, several bioinformatics classifiers have been trained on translation start site data, gleaned from ribosomal profiling, to predict putative translation initiation sites from mRNA sequence features. A stacking approach was formulated for the MetaTIS tool that can differentiate spurious and true translation initiation sites. The tool was trained on experimental data for translation initiation in HEK293 cells produced by the TISCA protocol, a method allowing for accurate translation initiation site identification. Our classifier delivers a notable ROC-AUC of 0.93 while performing on its own test set, as well as multiple external validation sets. Moreover, it was able to almost quantitively predict whether overlapping open-reading frames suppress translation from the main ORF for 11 genes in HeLa cells, as validated by experimental luciferase assays. The MetaTIS tool is publicly available as a webserver

The FlanksERF, KmersERF, and MetaTIS models with their training data can be found below. For information on how these models are utilized please refer to github. The datasets are composed of 229 columns. Whereby, the first 168 represent the upstream (U) and downstream (D) kmers of sizes 1 till 3. Then comes the start codon used and the normalized Noderer et al. efficiency values based on the flanking region. Next, 40 columns representing the 20 upsteam (U) and 20 downstream (D) nucleotides with respect to the initiation site. The final 19 features represent the relative binding scores of the 9 RNA binding proteins (RBPs) considered. Note that some RBPs have multiple binding motifs.

The FlanksERF and KmersERF are each composed of 40 random forest classifiers which are stored as a dictionary. scikit-learn version 1.5.1 was used to create these models. The DownstreamNegatives and Positives datasets were used to train the KmersERF model, while the UpstreamNegatives and Positives datasets were used to train the FlanksERF model.