Host Phenotype Prediction from Differentially Abundant Microbes Using RoDEO

Abstract

Metagenomics is the study of metagenomes which are mixtures of genetic material from several organisms. Metagenomic sequencing is increasingly used in human and animal health, food safety, and environmental studies. In these high-dimensional (metagenomic) data, the phenotype of the host organism, e.g., human, may not be obvious to detect and then the ability to predict it becomes a powerful analytic tool. For example, consider predicting the disease status of an individual from their gut microbiome.

In this study, we compare various normalization methods for metagenomic count data and their impact on phenotype prediction. The methods include RoDEO, Robust Differential Expression Operator, originally developed for gene expression studies. The best prediction accuracy is observed for RoDEO-processed count data with linear kernel support vector machines in most cases, for a variety of real datasets including human, mouse, and environmental samples.

We also address the problem of identifying the most relevant microbial features that could give insight into the structure and function of the differential communities observed between phenotypes. Interestingly, we obtain similar or better phenotype prediction accuracy with a small subset of features as with the complete set of sequenced features.

A.P. Carrieri and N. Haiminen contributed equally to this work.

Appendix: Experimental Details

1.1 RoDEO Projection Details on Full Datasets

For each of the 96 human samples with 134 OTUs, we run RoDEO for 100 independent re-sampling simulations, with $P = 7$ number of segments, $10^{6}$ number of reads for the re-sampling and gap parameter equal to 1. For each of the samples we compute the average of projected values for each OTU (average of the 100 iterations), and combine all the obtained values in a single matrix.

Similarly, we apply RoDEO to the 139 mouse samples and 10,172 OTUs for 100 independent re-sampling simulations, with $P = 10$ number of segments, $10^{7}$ number of reads for the re-sampling and gap equal to 1, and we compute the average of projected OTU values.

Finally, we run RoDEO for each of the 213 corpse samples with 17,803 OTUs for 100 independent re-sampling simulations, with $P = 10$ number of segments, $10^{7}$ number of reads for the re-sampling and gap between the samples equal to 2. In the same way as described before, we compute the average of projected OTU values for each sample.

1.2 Feature Selection Details

We start the feature selection process deleting duplicated OTUs from each of the three initial raw count datasets described in Sect. 2.7. Removing identical OTUs allow us to deal with smaller datasets and apply Random Forests as an alternative prediction method to SVM. More precisely, for the corpse data we remove about 3000 OTUs passing from an original dataset of 213 samples and 17804 OTUs to a new dataset with 213 samples and 14789 OTUs. For the mouse data we pass from 139 samples described by 10172 OTUs to 139 samples described by only 4411 features. Finally, in the human data we find only 4 OTUs identical in the count and we obtain a new human dataset with 97 samples and 130 OTUs.

We proceed to run DESeq2 on this duplicate-removed data, including the DESeq2 normalization and subsequent DE computation, in order to obtain a ranked list of differentially abundant OTUs. For RoDEO, projection and scaling is required before the DE computation, in order to make the samples directly comparable across phenotypes. Below is a detailed description of the RoDEO scaling process described in Sect. 2.1.

For the greatest human sample, i.e. the one with smallest number of zeros, we run RoDEO for 100 independent re-sampling simulations, with $P_g = 7$ number of segments, $10^{6}$ number of reads for the re-sampling and gap parameter 1. The number of segments we use to run RoDEO for all the other 96 human samples varies and depends on the result obtained from the scaling process for a given sample. All the other required parameters are instead equal to the ones used for the greatest sample. We then compute the average of projected values for each OTU (average of the 100 iterations), combine all the obtained values in a single matrix and we add to each row i, representing sample i, the difference between the number of segments $P_g$ used to run RoDEO on the greatest sample g and the number of segment $P_i$ used to run RoDEO on sample i.

Similarly, we apply RoDEO projection and the scaling algorithm to the mouse dataset running 100 independent re-sampling simulations, with $P = 10$ number of segments, $10^{7}$ number of reads for the re-sampling and gap 1, for the greatest mouse sample.

Finally, we run RoDEO on the greatest corpse sample for 100 independent re-sampling simulations, with $P = 10$ number of segments, $10^{7}$ number of reads for the re-sampling and gap between the samples equal to 2. In the same way as described before, we compute the averages of projected OTU values for each sample and we add the difference values from the scaling.