Incorporating metabolic activity, taxonomy and community structure to improve microbiome-based predictive models for host phenotype prediction

Figures & data

Table 1. Summary of human gut microbiome datasets used for disease state prediction.

Disease	Dataset abv. [ref]	No. of samples	Healthy	Patients	Dim. of in put profile
Inflammatory Bowel Disease	IBD^Citation19	110	85	25	443
Type 2 Diabetes	EW-T2D^Citation20	96	43	53	381
Type 2 Diabetes	C-T2D^Citation21	344	174	170	572
Liver Cirrhosis	Cirrhosis^Citation22	232	114	118	542
Colorectal Cancer	Colorectal^Citation23	121	73	48	503
Obesity	Obesity^Citation24	253	89	164	465
Obesity	Obesity-multi^Citation18	648	324	324	6463

Figure 1. Impacts of the hyperparameters on the MicroKPNN performance for the EW-T2D dataset.

(a) Comparison of the performance of the models built using different numbers of fully-connected hidden nodes (with the taxonomic rank set to “order” for this comparison). In this plot, each bar represents the performance of a model built using a certain number of fully connected hidden nodes (the numbers are shown below the bar). (b) comparison of the performance of models built using different taxonomic ranks (the number of fully connected hidden nodes was set to 10 for this comparison). The taxonomic ranks are shown below the bars in the plot. The standard deviation error bars were computed using results from five different runs.

Table 2. Summary of best performing neural network architecture for each dataset and their average AUC.

	No. of nodes in different groups in the hidden layer
Dataset	All	Taxon (rank)	Metabolite	Community	Fully connected	Avg. AUC
IBD	519	176 (genus)	234	29	80	0.954
EW-T2D	309	34 (order)	234	31	10	0.858
C-T2D	365	27 (class)	240	38	60	0.755
Cirrhosis	354	40 (order)	239	35	40	0.969
Colorectal	403	65 (family)	234	34	70	0.914
Obesity	479	184 (genus)	233	32	30	0.728
Obesity-multi	822	190 (order)	276	336	20	0.891

Table 3. Comparison of MicroKPNN with different methods including NNs that are fully connected (fc-NN) in averaged AUC and standard deviation (in parenthesis).

Dataset	MicroKPNN	MicroKPNN-g100^a	fc-NN	fc-NN(Mi-croKPNN)	DeepMicro^b	EPCNN^c
IBD	0.954 (0.037)	0.885 (0.085)	0.865 (0.031)	0.678 (0.089)	0.867 (0.059)	NA^d
EW-T2D	0.858 (0.067)	0.782 (0.032)	0.595 (0.065)	0.580 (0.062)	0.779 (0.072)	0.789 (0.056)
C-T2D	0.755 (0.032)	0.735 (0.058)	0.675 (0.033)	0.723 (0.019)	0.725 (0.060)	0.813 (0.024)
Cirrhosis	0.969 (0.009)	0.908 (0.051)	0.823 (0.022)	0.947 (0.021)	0.863 (0.027)	0.953 (0.007)
Colorectal	0.914 (0.046)	0.837 (0.042)	0.624 (0.038)	0.764 (0.053)	0.639 (0.055)	0.906 (0.013)
Obesity	0.728 (0.048)	0.650 (0.046)	0.539 (0.023)	0.608 (0.022)	0.631 (0.086)	NA
Obesity-multi	0.891 (0.083)	0.833 (0.091)	0.820 (0.014)	0.826 (0.045)	0.763 (0.042)	NA

The best performance for each dataset is highlighted in bold. ^a: MicroKPNN-g100, MicroKPNN with the taxonomic rank set to genus and the number of fully connected nodes set to 100. ^b: based on the results we derived by running DeepMicro on the same inputs of species abundance for all included approaches. Using the same practice as in,⁶: combinations of different autoencoders and ML approaches were tested and the best performance was reported here. ^c: the AUCs for EPCNN were taken from.⁹: Note the reported AUCs by EPCNN were based on predictors also trained using species abundances information, but the difference is that it used both known and unknown species, and the quantification approach was different. ^d: we attempted to run EPCNN on the three datasets IBD, Obesity and Obesity-multi but were unable to get results because the program resulted in an error.

Table 4. Comparison of MicroKPNN with DeepMicro in additional metrics (MCC and AUC-PR).

	MCC		AUC-PR
Dataset	MicroKPNN	DeepMicro	MicroKPNN	DeepMicro
IBD	0.704 (0.158)	0.358 (0.192)	0.891 (0.104)	0.381 (0.097)
EW-T2D	0.518 (0.126)	0.387 (0.205)	0.879 (0.122)	0.686 (0.092)
C-T2D	0.420 (0.129)	0.317 (0.121)	0.824 (0.046)	0.602 (0.050)
Cirrhosis	0.854 (0.107)	0.641 (0.102)	0.973 (0.020)	0.786 (0.057)
Colorectal	0.417 (0.108)	0.209 (0.083)	0.850 (0.021)	0.642 (0.026)
Obesity	0.311 (0.096)	0.129 (0.073)	0.861 (0.076)	0.661 (0.009)
Obesity-multi	0.222 (0.203)	0.383 (0.047)	0.649 (0.096)	0.626 (0.016)

Figure 2. Impacts of the downsampling of samples on the different approaches for selected datasets.

(a) Cirrhosis; (b) Colorectal cancer. We tried three different downsamplings, 75%, 50%, and 25%, and the results (AUCs and standard deviation distribution) are shown in the plots along with the performance when the entire dataset was used. We employed a stratified sampling approach to maintain the distribution balance between control and disease samples when downsampling.

Figure 3. Contributions of the different groups of hidden nodes to the prediction as measured by importance scores. (a) IBD; (b) EW-T2D; (c) C-T2D; (d) obesity; (e) cirrhosis; (f) colorectal cancer. The boxes in different colors with whiskers show the distribution of the importance scores of the hidden nodes in different groups.

Table 5. Ranks of taxonomic groups (orders) and metabolic activities that are potentially important for microbiome-based obesity prediction.

	Description	Edge weight^a		Diff. node weight^b
		In-group^c	Overall^d	In-group	Overall
Taxonomic	o_Pseudomonadales	1	1	1	1
	o_Rhizobiales	2	2	2	2
	o_Burkholderiales	3	3	10	10
	o_Bacillales	4	4	3	3
	o_Enterobacterales	5	5	6	6
Metabolic	NH3 (ammonia) production	1	9	1	12
	Acetate production	2	15	2	31
	L-Lactate production	3	21	3	32
	Sulfate sulfuric acid production	4	34	8	51
	CO2 production	5	39	4	35

The ranks were computed using importance scores from 50 replicated calculations. ^a: the importance scores were estimated using the outgoing edge weights of associated hidden nodes. ^b: the importance scores were estimated using the differential node weights. ^c: ranks according to the importance scores for the hidden nodes in each group. ^d: ranks according to the importance scores over all hidden nodes of the four groups.

Figure 4. The neural network structure used in MicroKPNN. It is composed of three layers (shown on the left). In the input layer, each node is a species, and the hidden layer includes nodes of four different groups: metabolites (red), taxa (blue), communities (green), and fully connected hidden nodes (gray). The links between the input nodes and the nodes in the hidden layer represent different biological meanings (shown on the right).

Qin J, Li R, Raes J, Arumugam M, Burgdorf KS, Manichanh C, Nielsen T, Pons N, Levenez F, Yamada T. A human gut microbial gene catalogue established by metagenomic sequencing. Nature. 2010;464(7285):59–65. doi:10.1038/nature08821.

 PubMed Web of Science ®Google Scholar

Karlsson FH, Tremaroli V, Nookaew I, Bergström G, Behre CJ, Fagerberg B, Nielsen J, Bäckhed F. Gut metagenome in European women with normal, impaired and diabetic glucose control. Nature. 2013;498(7452):99–103. doi:10.1038/nature12198.

 PubMed Web of Science ®Google Scholar

Qin J, Li Y, Cai Z, Li S, Zhu J, Zhang F, Liang S, Zhang W, Guan Y, Shen D, et al. A metagenome-wide association study of gut microbiota in type 2 diabetes. 2012;7418. doi:10.1038/nature11450.

 Google Scholar

Qin N, Yang F, Li A, Prifti E, Chen Y, Shao L, Guo J, Le Chatelier E, Yao J, Wu L, et al. Alterations of the human gut microbiome in liver cirrhosis. Nature. 2014;513:59–64.

 PubMed Web of Science ®Google Scholar

Zeller G, Tap J, Voigt AY, Sunagawa S, Kultima JR, Costea PI, Amiot A, Böhm J, Brunetti F, Habermann N, et al. Potential of fecal microbiota for early-stage detection of colorectal cancer. Molecul Systs Biol. 2014;10(11). doi:10.15252/msb.20145645.

 PubMed Web of Science ®Google Scholar

Le Chatelier E, Nielsen T, Qin J, Prifti E, Hildebrand F, Falony G, Almeida M, Arumugam M, Batto JM, Kennedy S. Richness of human gut microbiome correlates with metabolic markers. Nature. 2013;500(7464):541–546. doi:10.1038/nature12506.

 PubMed Web of Science ®Google Scholar

Lam TJ, Ye Y. Meta-analysis of microbiome association networks reveal patterns of dysbiosis in diseased microbiomes. Sci Rep. 2022;12(1):17482. doi:10.1038/s41598-022-22541-1.

 PubMed Web of Science ®Google Scholar

Oh M, Zhang L. DeepMicro: deep representation learning for disease prediction based on microbiome data. Sci Rep. 2020;10(1):1–9. doi:10.1038/s41598-020-63159-5.

 PubMed Web of Science ®Google Scholar

Chen X, Zhu Z, Zhang W, Wang Y, Wang F, Yang J, Wong KC. Human disease prediction from microbiome data by multiple feature fusion and deep learning. Iscience. 2022;25(4):104081. doi:10.1016/j.isci.2022.104081.

 PubMed Web of Science ®Google Scholar

Data availability statement

MicroKPNN is available as an open source repository at https://github.com/mgtools/MicroKPNN.