Table 1. Average prediction accuracies on testing data from 12 random training-validation-testing splits, by using different methods for categorical covariates (T2D study).

Acronyms: Logistic - logistic regression with elastic net penalty using original data; SVM - support vector machine classifier using original data; RF - random forest classifier using original data; MLP multi-layer perceptron using original data; MB-simCLR - logistic regression model with elastic net penalty using microbiome embeddings learned from unsupervised contrastive learning; MB-SupCon + Logistic - logistic regression model with elastic net penalty using microbiome embeddings learned from supervised contrastive learning. MB-SupCon + SVM: support vector machine classifier using microbiome embeddings learned from supervised contrastive learning; MB-SupCon + RF: random forest classifier using microbiome embeddings learned from supervised contrastive learning; MB-SupCon + MLP: multi-layer perceptron using microbiome embeddings learned from supervised contrastive learning; Avg. Acc. based on MB-SupCon: average accuracies among MB-SupCon + Logistic, MB-SupCon + SVM, MB-SupCon + RF and MB-SupCon + MLP.

Prediction Task	Logistic	SVM	RF	MLP	MB-simCLR
Insulin resistance	76.69%	79.46%	83.93%	83.73%	65.67%
Sex	65.61%	69.02%	80.38%	78.94%	59.85%
Race	72.99%	72.17%	77.90%	75.60%	68.38%
Prediction Task	MB-SupCon + Logistic	MB-SupCon + SVM	MB-SupCon + RF	MB-SupCon + MLP	Avg. Acc. based on MB-SupCon
Insulin resistance	84.42%	85.12%	84.62%	84.33%	84.62%
Sex	78.94%	79.02%	79.24%	78.71%	78.98%
Race	80.73%	80.36%	79.91%	79.17%	80.04%