To the editor

Preterm birth (PTB) is a leading cause of neonatal mortality worldwide and the second most common cause of child deaths under the age of five years [1]. Additionally, premature neonates are at risk of numerous health complications, including neurological damage in early childhood but also respiratory and gastrointestinal disorders. Existing diagnostic methods involve the collection of maternal obstetric history and cervical measurements via transvaginal ultrasound imaging conducted in the first and second trimesters of pregnancy. However, diagnoses are often inaccurate, as physician experience varies and the processes can be time-consuming.

Existing literature suggests that vaginal microbial communities could be involved in the pathophysiology of PTB delivery [2]. This microbiome is extremely important to the host tissue, as it maintains an acidic environment, inhibits the growth of pathogenic bacteria, and modulates inflammation by cross-kingdom signalling. Despite the efforts of longitudinal studies and meta-analyses, no clear distinct microbial signatures have been characterized to identify the risk of PTB [3]. We propose DeepMPTB, a vaginal microbiota-based deep neural network (DNN) for efficient PTB prediction (Fig. 1; Supplementary Material).

Fig. 1
figure 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

Full size image

A total of 234 786 trainable parameters were optimized and the optimal hyperparameter combination for the final model (Fig. 2A, Fig. S2), included 416 units (neurons) in the 1st hidden layer and a total of 3 hidden layers, with the number of units in each layer set to half that in the preceding layer (Fig. S3). To deal with class imbalance (1029 TB and 261 PTB) in our datasets, we evaluated model performance using multiple metrics (Supplementary Material). The 20 most important features contributing to these results were also determined by the SHAP explainer (Fig. 2B). Interestingly, low-abundance species were also observed to contribute to PTB classification. Moreover, these contributing features included clinical and demographic data.

Fig. 2
figure 2

Performance of DeepMPTB based on the 20% test set (239 samples). A During model evaluation, the three metadata features (age, ethnicity and trimester of sample collection) were associated with each sample. For each sample, the true and predictive phenotypes were compared to evaluate the performance of DeepMPTB. B A summary plot for the SHAP values was generated to understand the contributions of the first 20 features in this performance analysis. Features related to clinical/demographic metadata are indicated by an asterisk. Each dot represents a sample. Negative and positive SHAP values are associated to TB and PTB prediction, respectively. Low and high SHAP values are shown in blue and red, respectively. TB: Term birth; PTB: Preterm birth

Full size image

The performance of DeepMPTB was compared with that of seven state-of-the-art classification algorithms, namely, the decision tree (DT), K-nearest neighbour (KNN), random forest (RF), naïve Bayes (NB), extreme gradient boosting (XGBoost), logistic regression (LR), and support vector machine (SVM) models, which were trained and optimized based on the same input data as the DNN (Supplementary Material). DeepMPTB outperformed all other prediction models, with an AUROC score of 0.877 ± 0.11 (p < 0.05 for ANOVA test) and an accuracy of 84.10% (Table S2).

The model trained based on the third trimester data displayed the highest accuracy of 88%, suggesting that samples collected during the third trimester may lead to better prediction rates, although the models trained based on first and second trimester data obtained also very good accuracies of 71% and 83%, respectively (Fig. S4).

We also argue that input data quality has a significant impact on model performance (Fig. S5). We compared performance of three DNN trained using species coupled to their relative abundances determined with RiboTaxa [8] or two other popular metagenomics classifiers (p < 0.05 for ANOVA test) using only the biggest cohort (Supplementary Material). The DNN trained based on input data from RiboTaxa [8] showed the best performance, with an AUROC score of 0.898 ± 0.09. The DNN trained based on DeepMicrobes [9] and MetaPhlAn3 [10] data showed an AUROC score of 0.838 ± 0.14 and 0.795 ± 0.08, respectively. When only microbiome data obtained with RiboTaxa (without metadata, keeping phenotype) were used for model training, the AUROC value decreased to 0.831 ± 0.12 (p > 0.05 for Mann‒Whitney U test).

To show the generality of this model, we used a completely new set of 694 vaginal metagenomic data (430 TB and 264 PTB cases) from Baud et al. [11]. Overall, the optimized DNN successfully identified 80% of TB samples and 66% of the PTB samples. Importantly, phenotype prediction, especially in the case of PTB, is not determined by the presence of the same species or group of species (Fig. S6).

In conclusion, the present study presents a cutting-edge deep learning model to efficiently predict TB and PTB using vaginal microbiome data of pregnant women combined to clinical data. This new model based on data from 5 cohorts outperforms previously published machine learning-based model for PTB prediction [11, 12]. Continued accumulation of high-quality microbiome data and complete phenotypic data in perfectly controlled cohorts will certainly improve the individual phenotype prediction performance of deep learning models. Furthermore, including virome information, known to drive microbiota dynamics, would help to reach better performances. Finally, DNN enables to distinguish complex interindividual microbial interactions related to term and preterm deliveries, to highlight in-depth microorganisms potentially associated to phenotype. Interestingly we observed that different microbial profiles led to the same phenotype. This efficient TB and PTB predictive diagnosis should be highly helpful for clinicians in a personalized medicine context.