We defined the task of learning SP regions as a multilabel classification problem at each sequence position. Multilabel differs from multiclass in the sense that more than one label can be true at a given position. This approach was motivated by the fact that there is no strict definition of region borders that is commonly agreed upon, making it impossible to establish ground-truth region labels for models to train on. We thus used the multilabel framework as a method for training with weak supervision, allowing us to use overlapping region labels during the learning phase that could be generated from the sequence data using rules. For inference, we did not make use of the multilabel framework, as we only predicted the single most probable label at each position using Viterbi decoding, yielding a single unambiguous solution.
We defined a set of three rules based on known properties of the n-, h-, and c-regions. The initial n-region must have a minimum length of two residues and the terminal c-region a minimum length of three residues. The most hydrophobic position, which is identified by sliding a seven-amino-acid window across the SP and computing the hydrophobicity using the Kyte–Doolittle scale29 (link), belongs to the h-region. All positions between these six labeled positions are labeled as either both n and h or h and c, yielding multitag labels.
This procedure was adapted for different SP classes, with only Sec/SPI completely following it. For Tat SPs, the n–h border was identified using the twin-arginine motif. All positions before the motif were labeled n, followed by two dedicated labels for the motif, again followed by a single position labeled n. For SPII SPs, we did not label a c-region, as the C-terminal positions cannot be considered as such30 (link). The last three positions were labeled as the lipobox, all positions before that as h only. For SPIII SPs, no region labels were generated within the SP.
We defined a set of three rules based on known properties of the n-, h-, and c-regions. The initial n-region must have a minimum length of two residues and the terminal c-region a minimum length of three residues. The most hydrophobic position, which is identified by sliding a seven-amino-acid window across the SP and computing the hydrophobicity using the Kyte–Doolittle scale29 (link), belongs to the h-region. All positions between these six labeled positions are labeled as either both n and h or h and c, yielding multitag labels.
This procedure was adapted for different SP classes, with only Sec/SPI completely following it. For Tat SPs, the n–h border was identified using the twin-arginine motif. All positions before the motif were labeled n, followed by two dedicated labels for the motif, again followed by a single position labeled n. For SPII SPs, we did not label a c-region, as the C-terminal positions cannot be considered as such30 (link). The last three positions were labeled as the lipobox, all positions before that as h only. For SPIII SPs, no region labels were generated within the SP.
Free full text: Click here
