Fig. 13

An MLP with a taxonomy-mimicking architecture does not enhance prediction accuracy. The MLP model used as baseline is configured with a pyramid-shaped architecture with one hidden layer of size half of the input dimensionality. Alternatively, we designed an MLP model that mirrors the taxonomy architecture, with each hidden layer corresponding to a specific taxonomic level, maintaining the same number of layers and neurons. The alternative MLP design significantly underperforms in prediction, as evidenced by lower AUPRC and AUROC scores. This suggests that taxonomy alone does not account for the superior predictive performance, as the increased number of parameters introduces additional challenges during training