Analysis of oral microbiome in glaucoma patients using machine learning prediction models

Figures & data

Table 1. Demographic and baseline characteristics of glaucoma subjects

Variables	Glaucoma n = 96	Control subject n = 25	P-value*
Age, years	61.48 ± 13.08	62.68 ± 1.73	0.385
Male/Female	66/30	18/7	0.753
BCVA, logMAR	0.21 ± 0.32	0.00 ± 0.20	<0.0001
Diagnosis, n (%)
POAG/Secondary glaucoma	62 (64.58)/34 (35.42)	-
Glaucoma medication period (yr)	2.74 ± 3.09	-
Baseline IOP, mmHg	18.49 ± 6.60	13.76 ± 1.01	<0.0001
RNFL thickness, µm	73.44 ± 14.52	105.36 ± 5.62	<0.0001
Visual field (MD), dB	−8.93 ± 8.44	0.03 ± 0.17	<0.0001
Visual field (VFI), %	76.97 ± 26.52	99.84 ± 0.47	<0.0001
Number of lead counts/sample	77,140.05 ± 23,087.21	83,310.64 ± 8,812.47	0.038

P-value* was calculated by independent T-test for continuous variables, Chi-square test was performed in case of numeric variable

IOP: intraocular pressure, BCVA: best-corrected visual acuity, logMAR: logarithm of the minimum angle of resolution, MD: mean deviation. VFI: visual field index, RNFL: retinal nerve fiber layer, POAG: primary open angle glaucoma

Table 2. Gini index for accuracy condition in genus levels using random forest machine learning

Taxon Level	Ntree	Ratio	mtry	Times	Accuracy	95% CI	P-Value^†
*genus	200	0.1	193	100	0.5714 ~ 0.7429	0.3935 ~ 0.8751	2.81e^−08 ~ 0.0003783
genus	200	0.2	385	100	0.5714 ~ 0.7143	0.3935 ~ 0.8536	1.869e^−07 ~ 0.0003783
genus	200	0.3	578	100	0.6 ~ 0.7143	0.4211 ~ 0.8536	1.869e^−07 ~ 0.0001039
genus	200	0.4	771	100	0.6 ~ 0.7143	0.4211 ~ 0.8536	1.869e^−07 ~ 0.0001039
genus	200	0.5	964	100	0.5714 ~ 0.7143	0.3935 ~ 0.8536	1.869e^−07 ~ 0.0003783
genus	200	0.6	1156	100	0.6 ~ 0.7143	0.4211 ~ 0.8536	1.869e^−07 ~ 0.0001039
genus	200	0.7	1349	100	0.6 ~ 0.6857	0.4211 ~ 0.8315	1.089e^−06 ~ 0.0001039
genus	200	0.8	1542	100	0.5714 ~ 0.7143	0.3935 ~ 0.8536	1.869e^−07 ~ 0.0003783
genus	200	0.9	1734	100	0.6 ~ 0.7143	0.4211 ~ 0.8536	1.869e^−07 ~ 0.0001039
genus	200	1	1927	100	0.5714 ~ 0.7429	0.3935 ~ 0.8751	2.81e^−08 ~ 0.0003783

*Best accuracy condition in genus level, P-Value^† was calculated by machine learning using random forest model

95% CI: confidential Interval

Figure 1. Metagenome analysis using machine learning and statistical models

a: Schematic of machine learning for the study of algorithms and statistical models [23–25] TCC, tag count comparison; CPAR, Classification based on predictive association rules

b: Using Random Forest method, the significant taxa and genus levels in the oral microbiome were identified.

c: The top 10 species of the oral microbiome as identified from the Ribosomal Database Project (RDP) Data Base; Lactococcus was the most significant species in glaucoma patients and control subjects.

Figure 2. Comparison of alpha-diversity in glaucoma patients as compared with the control subjects

(a) Observed OTUs was higher in the control group than in the glaucoma (P = 0.003). It was also higher in the control group than in the two subgroups (P = 0.011)

(b) Shannon index was higher in the control group than in the glaucoma group (P = 6.1e⁻⁰⁶). It was also higher in the control group than in the two subgroups (P = 3.6e⁻⁰⁵). POAG, primary open-angle glaucoma; SG, secondary glaucoma

Figure 3. Differences in the oral microbiome of glaucoma patients compared with control subjects using principal component analysis for beta-diversity

Beta-diversity was calculated in two ways – RDP and NCBI library. PCoA shows the different clusters in glaucoma vs. control groups, and POAG and SG subgroups vs. control group.

RDP, Ribosomal Database Project Classifier; NCBI, National Center for Biotechnology Information; POAG, Primary open-angle glaucoma; SG, secondary glaucoma

Figure 4. Beta-Diversity differences in the microbiome of glaucoma patients and control subjects

Beta-diversity (estimated by RDP method) was significantly higher in glaucoma group than in control groups (P < 2.2e⁻¹⁶); however, the beta-diversity using NCBI was not different between control and glaucoma groups (P = 0.52).

RDP, Ribosomal Database Project Classifier; NCBI, National Center for Biotechnology Information; POAG, Primary open-angle glaucoma; SG, secondary glaucoma

Figure 5. MA and volcano plots generated using analysis with tag count comparison for oral microbiome of glaucoma patients and control subjects

a: MA plot of glaucoma vs control subjects (DE number of the upper panels 166 vs. that of the lower panels 171; P < 0.05)

b: Volcano plot of DEG in glaucoma vs control subjects Lactococcus and Candidatus Pelagibacter were downregulated in the glaucoma group DEG, differential expression of genes

Figure 6. MA and volcano plots generated using subgroup analysis with tag count comparison for oral microbiome in the POAG subgroups compared with the control group

a: MA plot of POAG vs control subjects (DE number of the upper panels 175 vs. that of the lower panels 142; P < 0.05)

b: Volcano plot of DEG in POAG vs control subjects DEG, differential expression of genes; POAG, primary open-angle glaucoma

Table 3. Multinomial logistic regression analysis of factors related taxon in glaucoma group compared with control subjects

	coefficient	SD	Odd ratio	P-value	coefficient	SD	Odd ratio	P-value*
Taxa	Age < 60				Age ≥ 60
Lactococcus	−210.481	55.386	3.88e^−92	0.0001	−172.977	45.470	7.53e^−76	0.0001
Endobacter	0.223	1.750	1.25e^+00	0.8984	−27.032	15.330	1.82e^−12	0.0778
Rikenella	89.032	237.423	4.63e^+38	0.7077	−80.714	338.227	8.83e^−36	0.8114
Candidatus Pelagibacter	−560.729	205.488	3.01e^−244	0.0064	−921.479	259.819	0.00e^+00	0.0004
Atopobium	−3.561	1.322	0.02842	0.0071	−4.122	1.244	0.01620	0.0009
	Baseline IOP < 20 mmHg				Baseline IOP ≥ 20 mmHg
Lactococcus	−177.477	45.896	8.37e^−78	0.0001	−191.253	52.588	8.70e^−84	0.0003
Endobacter	−13.841	12.697	9.75e^−07	0.2757	0.094	1.730	1.10e^+00	0.9567
Rikenella	14.974	47.738	3.18e^+06	0.7538	33.475	46.674	3.45e^+14	0.4732
Candidatus Pelagibacter	−509.789	171.543	3.99e^−222	0.0030	−319.291	145.031	2.16e^−139	0.0277
Atopobium	−4.134	1.252	1.60e^−02	0.0010	−3.561	1.310	2.84e^−02	0.0065
	RNFL thickness ≥ 80 μm				RNFL thickness < 80 μm
Lactococcus	−207.229	53.820	1.00e^−90	0.0001	−173.639	45.928	3.89e^−76	0.0002
Endobacter	−13.957	15.484	8.68e^−07	0.3674	−0.536	1.711	5.85e^−01	0.7542
Rikenella	−38.900	295.676	1.28e^−17	0.8953	61.128	142.806	3.53e^+26	0.6686
Candidatus Pelagibacter	−679.649	250.852	6.80e^−296	0.0067	−782.464	226.057	0.00e^+00	0.0005
Atopobium	−4.731	1.622	8.81e^−03	0.0035	−3.545	1.118	2.89e^−02	0.0015
									coefficient	SD	Odd ratio	P-value
	Early glaucoma				Moderate glaucoma				Advanced glaucoma
Lactococcus	−155.716	46.085	2.36e^−68	0.0007	−231.783	65.006	2.18e^−101	0.0004	−218.896	60.662	8.60e^−96	0.0003
Endobacter	−25.875	17.045	5.79e^−12	0.1290	−6.534	10.847	1.45e^−03	0.5469	0.488	1.756	1.63e^+00	0.7813
Rikenella	−35.096	155.035	5.73e^−16	0.8209	32.121	63.202	8.91e^+13	0.6113	43.848	62.576	1.10e^+19	0.4835
Candidatus Pelagibacter	−1695.298	445.654	0.00e^+00	0.0001	−1370.249	457.311	0	0.0027	−1342.329	430.780	0	0.0018
Atopobium	−5.753	1.710	3.17e^−03	0.0008	−15.354	5.126	2.15e^−07	0.0027	−2.135	0.941	1.18e^−01	0.0232

P-value* was calculated using multinomial logistic regression analysis

IOP: intraocular pressure, RNFL: retinal nerve fiber layer, SD: standard deviation

Table 4. Association rules generated by association rule mining for glaucoma using oral microbiome data

Glaucoma rules
Rule 1	Atopobium = [1] Candidatus Pelagibacter = [1] Endobacter = [1] Lactococcus = [1] → Glaucoma 0.96
Rule 2	Lactococcus = [1] → Glaucoma 0.9
Normal rules
Rule 1	Candidatus Pelagibacter = [2] Lactococcus = [2] → Normal 0.92
Rule 2	Lactococcus = [2] → Normal 0.85
Rule 3	Candidatus Pelagibacter = [2] → Normal 0.7
Rule 4	Atopobium = [2] → Normal 0.65
Rule explanations (Rule 1)
Original rule	Atopobium = [1] Candidatus Pelagibacter = [1] Endobacter = [1] Lactococcus = [1] → Glaucoma 0.96
Rule explanations	If the compositions of Atopobium ≤ 0.347 Candidatus Pelagibacter ≤ 0.001 Endobacter ≤ 0.015 Lactococcus ≤ 0.013; the sample is Glaucoma with 96.0% accuracy
Applied category for rule generation
Category	Atopobium	Candidatus Pelagibacter	Endobacter	Lactococcus	Rikenella
[1]	≤ 0.347	≤ 0.001	≤ 0.015	≤ 0.013	≤ 0.001
[2]	0.347–2.908	0.001–0.014	0.015–1.106	0.013–0.889	0.001–0.159

CPAR (classification based on predictive Association Rules) algorithm used [24].

Table 5. Multinomial logistic regression analysis of factors related taxon in primary open-angle glaucoma

	coefficient	SD	Odd ratio	P-value	coefficient	SD	Odd ratio	P-value*
	Age < 60				Age ≥ 60
Lactococcus	−175.539	65.322	5.81e^−77	0.0072	−157.432	45.664	4.25e^−69	0.0006
Endobacter	−10.362	14.712	3.16e^−05	0.4813	−57.292	24.940	1.31e^−25	0.0216
Rikenella	95.277	195.573	2.39e^+41	0.6261	−97.172	373.910	6.29e^−43	0.7950
Candidatus Pelagibacter	−328.189	215.986	2.95e^−143	0.1286	−637.095	220.119	2.06e^−277	0.0038
Atopobium	−3.874	2.006	0.0208	0.0534	−3.826	1.297	0.02179	0.0032
	Baseline IOP < 20 mmHg				Baseline IOP ≥ 20 mmHg
Lactococcus	−159.538	46.999	5.17e^−70	0.0007	−163.518	56.609	9.66e^−72	0.0039
Endobacter	−27.835	15.960	8.15e^−13	0.0812	−52.336	33.014	1.87e^−23	0.1129
Rikenella	18.174	46.865	7.81e^+07	0.6982	39.064	45.859	9.23e^+16	0.3943
Candidatus Pelagibacter	−602.133	230.776	3.14e^−262	0.0091	−435.431	232.158	7.85e^−190	0.0607
Atopobium	−5.446	1.773	4.31e^−03	0.0021	−2.531	1.203	7.96e^−02	0.0353
	RNFL thickness ≥ 80 μm				RNFL thickness < 80 μm
Lactococcus	−204.775	58.817	1.17e^−89	0.0005	−140.536	46.567	9.24e^−62	0.0025
Endobacter	−44.285	25.725	5.85e^−20	0.0852	−28.076	16.803	6.41e^−13	0.0947
Rikenella	−41.052	327.684	1.48e^−18	0.9003	64.161	125.357	7.32e^+27	0.6088
Candidatus Pelagibacter	−323.115	164.296	4.71e^−141	0.0492	−401.957	172.983	2.71e^−175	0.0201
Atopobium	−4.243	1.712	1.44e^−02	0.0132	−3.599	1.341	2.74e^−02	0.0073
									coefficient	SD	Odd ratio	P-value
	Early glaucoma				Moderate glaucoma				Advanced glaucoma
Lactococcus	−125.291	45.677	3.86e^−55	0.0061	−229.962	79.424	1.35e^−100	0.0038	−216.259	72.337	1.20e^−94	0.0028
Endobacter	−31.014	18.134	3.39e^−14	0.0872	−29.294	25.434	1.89e^−13	0.2494	−45.306	32.995	2.11e^−20	0.1697
Rikenella	−28.362	105.738	4.82e^−13	0.7885	29.192	42.020	4.76e^+12	0.4872	39.907	41.176	2.14e^+17	0.3325
Candidatus Pelagibacter	−1122.831	394.907	0.00e^+00	0.0045	−854.987	435.199	0	0.0495	−777.340	372.864	0	0.0371
Atopobium	−5.434	1.875	4.36e^−03	0.0038	−19.236	7.884	4.42e^−09	0.0147	−1.926	1.071	1.46e^−01	0.0722

P-value* was calculated using multinomial logistic regression analysis

IOP: intraocular pressure, RNFL: retinal nerve fiber layer, SD: standard deviation

Table 6. Association rules generated by association rule mining for primary open-angle glaucoma

POAG rules
Rule 1	Atopobium = [1] Candidatus Pelagibacter = [1] Endobacter = [1] Lactococcus = [1] → POAG 0.93
Normal rules
Rule 1	Candidatus Pelagibacter = [2] → Normal 0.93
Rule 2	Lactococcus = [2] → Normal 0.8
Rule 3	Atopobium = [2] → Normal 0.73
Rule 4	Endobacter = [2] → Normal 0.53
Rule explanation (Rule 1)
Original rule	Atopobium = [1] Candidatus Pelagibacter = [1] Endobacter = [1] Lactococcus = [1] → POAG 0.93
Rule explanations	If the compositions of Atopobium ≤ 0.433 Candidatus Pelagibacter ≤ 0.001 Endobacter ≤ 0.015 Lactococcus ≤ 0.015; the sample is POAG with 93.0% accuracy
Applied category for rule generation
Category	Atopobium	Candidatus Pelagibacter	Endobacter	Lactococcus	Rikenella
[1]	≤ 0.433	≤ 0.001	≤ 0.015	≤ 0.015	≤ 0.001
[2]	0.433–2.908	0.001–0.014	0.015–0.475	0.015 − 0.889	0.001–0.159

CPAR (classification based on predictive Association Rules) algorithm used [24].

Quigley HA. Glaucoma. Lancet. 2011;377(9774):1367–14.

 PubMed Web of Science ®Google Scholar

Fechtner RD, Khouri AS. Evolving global risk assessment of ocular hypertension to glaucoma. Curr Opin Ophthalmol. 2007;18(2):104–109.

 PubMed Web of Science ®Google Scholar

Yin X, Cpar: HJ Classification based on predictive association rules. Proceedings of the 2003 SIAM International Conference on Data Mining: San Francisco, CA: SIAM; 2003. p. 331–335.

 Google Scholar