The authors state that they adhere with COPE guidelines on publishing ethics as described elsewhere at https://publicationethics.org/. The authors also undertake that they are not associated with any other third party (governmental or non-governmental agencies) linking with any form of unethical issues connecting to this publication. The authors also declare that they are not withholding any information that is misleading to the publisher in regard to this article.

The composition of microbial communities has been known to be location-specific. Investigating the microbial composition across different cities enables us to unravel city-specific microbial signatures and further predict the origin of unknown samples. As part of the CAMDA 2020 Metagenomic Geolocation Challenge, MetaSUB provided the whole genome shotgun (WGS) metagenomics data from samples across 28 cities along with non-microbial city data for 23 of these cities. In our solution to this challenge, we implemented feature selection, normalization, clustering and three methods of machine learning to classify the cities based on their microbial compositions. Of the three methods, multilayer perceptron obtained the best performance with an error rate of 19.60% based on whether the correct city received the highest or second highest number of votes for the test data contained in the main dataset. We then trained the model to predict the origins of samples from the mystery dataset by including these samples with the additional group label of “mystery.” The mystery dataset compromised of samples collected from a subset of the cities in the main dataset as well as samples collected from new cities. For samples from cities that belonged to the main dataset, error rates ranged from 18.18 to 72.7%. For samples from new cities that did not belong to the main dataset, 57.7% of the test samples could be correctly labeled as “mystery” samples. Furthermore, we also predicted some of the non-microbial features for the mystery samples from the cities that did not belong to main dataset to draw inferences and narrow the range of the possible sample origins using a multi-output multilayer perceptron algorithm.

No abstract

<30 14 (23% 30-34 12 (20%) 35-39 24 (39%) 40-45 11 (18%) Race White 9 (15%) Black 12 (20%) Asian 11 (18%) Other/Unknown 29 (48%) Ethnicity Non-Hispanic 36 (59%)