Development and evaluation of a colorectal cancer screening method using machine learning‐based gut microbiota analysis

Abstract

Accumulating evidence indicates that alterations of gut microbiota are associated with colorectal cancer (CRC). Therefore, the use of gut microbiota for the diagnosis of CRC has received attention. Recently, several studies have been conducted to detect the differences in the gut microbiota between healthy individuals and CRC patients using machine learning‐based gut bacterial DNA meta‐sequencing analysis, and to use this information for the development of CRC diagnostic model. However, to date, most studies had small sample sizes and/or only cross‐validated using the training dataset that was used to create the diagnostic model, rather than validated using an independent test dataset. Since machine learning‐based diagnostic models cause overfitting if the sample size is small and/or an independent test dataset is not used for validation, the reliability of these diagnostic models needs to be interpreted with caution. To circumvent these problems, here we have established a new machine learning‐based CRC diagnostic model using the gut microbiota as an indicator. Validation using independent test datasets showed that the true positive rate of our CRC diagnostic model increased substantially as CRC progressed from Stage I to more than 60% for CRC patients more advanced than Stage II when the false positive rate was set around 8%. Moreover, there was no statistically significant difference in the true positive rate between samples collected in different cities or in any part of the colorectum. These results reveal the possibility of the practical application of gut microbiota‐based CRC screening tests.

A systematic review of microbiome-derived biomarkers for early colorectal cancer detection

“…For microbiome-derived biomarkers based on bacteria for the detection of early CRC, 26 articles were found and two articles were found investigating the co-metabolome solely as a biomarker. Seventeen of the studies were conducted in East Asia [ 19 , [27] , [28] , [29] , [30] , [31] , [32] , [33] , [34] , [35] , [36] , [37] , [38] , [39] , [40] , [41] , [42] ], one in South Asia [43] , eight in Europe [44] , [45] , [46] , [47] , [48] , [49] , [50] , [51] and two in North America [ 52 , 53 ]. The number of inclusions of these studies ranged between 2 and 462 for CRC, 18 and 665 for adenoma and/or polyp and 24 and 788 for healthy controls.…”

“…The number of inclusions of these studies ranged between 2 and 462 for CRC, 18 and 665 for adenoma and/or polyp and 24 and 788 for healthy controls. Regarding the stage used in the articles for early CRC detection, five studies were based on stage I or II [ 29 , 30 , 37 , 40 , 50 ] and two used stage 0 [ 19 , 38 ]. The other studies investigated a precursor lesion of CRC, either advanced adenomas or polyps (AP) [ 28 , [31] , [32] , [33] , [34] , [35] , 39 , 41 , [44] , [45] , [46] , [47] , [48] , [49] , [51] , [52] , [53] ].…”

“…The other studies investigated a precursor lesion of CRC, either advanced adenomas or polyps (AP) [ 28 , [31] , [32] , [33] , [34] , [35] , 39 , 41 , [44] , [45] , [46] , [47] , [48] , [49] , [51] , [52] , [53] ]. Only five articles evaluated both precursor lesions of CRC and stages I and II [ 27 , 36 , 38 , 43 , 54 ]. All articles used faecal samples, except for four: two studies used the microbiota collected from the oral mucosa instead of or next to faecal samples [ 31 , 46 ] and two used serum-based samples next to faecal samples [ 36 , 54 ].…”

“…and Konishi et al. were the only ones that included stage 0 in their analyses and reported an AUC of 0.73 and 0.54, respectively [ 19 , 38 ]. For stage I, an AUC range between 0.74 and 0.97 was seen, with a sensitivity ranging from 90 to 100 per cent, and stage I/II or only stage II showed a range between 0.82-0.89, with a sensitivity ranging from 60 to 100 per cent.…”

Do we need to change our perspective about gut biomarkers? A public data mining approach to identify differentially abundant bacteria in intestinal inflammatory diseases

“…Due to the following reasons, some articles were discarded: the study did not meet the inclusion criteria, the metadata provided by the study did not enable the differentiation of case and control samples, the accession number corresponding to the sequencing data of the study was not found in the repositories, and the quality of the sequences was too low for downstream analysis. After inclusion criteria evaluation, quality verification, and taxonomic assignment with DADA2, the total of analyzed studies were: five for CRC ( Lu et al., 2016 ; Clos-Garcia et al., 2020 ; Uchino et al., 2021 ; Du et al., 2022 ; Konishi et al., 2022 ), eight for CD ( Forbes et al., 2018 ; Zhou et al., 2018a ; Braun et al., 2019 ; Lloyd-Price et al., 2019 ; Weng et al., 2019 ; Alam et al., 2020 ; Bourgonje et al., 2022 ; Pisani et al., 2022 ), nine for UC ( Bajer et al., 2017 ; Halfvarson et al., 2017 ; Forbes et al., 2018 ; Zhou et al., 2018a ; Lloyd-Price et al., 2019 ; Weng et al., 2019 ; Alam et al., 2020 ; Quraishi et al., 2020 ; Bourgonje et al., 2022 ; Pisani et al., 2022 ), four for IBS ( Zhuang et al., 2018 ; Zhu et al., 2019 ; Wang et al., 2020 ; Rogers et al., 2021 ), and one for CDI Duan et al, (2020) . Additional information on the mentioned studies is described in Supplementary Data 1 .…”