Comparison of sample-level classification accuracy on AGP data.

Unlike sample-level classification methods that use OTU/ASV tables and k-mers (e.g. 9-mers) as features, our proposed model is trained on reads. Then, read-level results are fused by the sample-level predictor using three methods as described in this paper. By increasing the number of samples in the training data, we compare the read-level classifier’s ability to learn sample-level predictive taxa/information from limited data sizes. The training set size refers to the number of samples used for training. For each training set size, we train 5 different models on 5 sets of randomly selected training samples and the accuracies are averaged and standard deviation is measured over 5 different experiments. We show sample-level prediction for the proposed methods are competitive with prediction from OTU tables and will allow interpretable representations shown in subsequent sections.