Fig. 1

Overview of model training and phenotype prediction. For model training (step 1), the shotgun metagenomics sequences of 1290 vaginal samples from 561 pregnant women were retrieved from public databases in the form of fastq files (Table S1) [3,4,5,6,7]. The RiboTaxa pipeline [8] was used to obtain taxonomic profiles from the metagenomics datasets using the SILVA SSU 138.1 NR99 database. Vaginal microbiota profiles differed greatly (Welch’s t-test, p < 0.05) within individual cohorts, illustrating the heterogeneity of the vaginal population. No significant difference in the α-diversity measure was found between the TB or PTB groups. All the output taxonomy tables were grouped into a single table containing all the bacterial and eukaryotic species-level profiles of 1290 samples. In addition, the clinical data of each sample were considered. The normalized species abundances (Fig. S1) and vectorized clinical data were used to train and optimize the neural network. Features contributing to explaining the model were extracted and visualized using SHAP. To predict the phenotype based on new unknown vaginal microbiota samples (step 2), a list of features with important biomarkers contributing to the prediction was output