<p>Comparison of data processing methods for three “metagenomics-to-metabolomics” models: MelonnPan, MiMeNet and SparseNED.</p>

We use the word “default” to refer to the data processing approach applied internally by the model, on normalized data. Some experiments were not performed. For example, MelonnPan already uses an arcsine and quantile transformation in its default pipeline, so we omitted that comparison. For each model, we highlight the data processing approach selected to report results for the model. Performance was measured using the average Spearman’s rank correlation coefficient for the 50 best predicted features. (XLSX)