Abstract

   Nonsyndromic cleft lip with palate (nsCLP) is a common birth defect
   disease. Current diagnostic methods comprise fetal ultrasound images,
   which are mainly limited by fetal position and technician skills. We
   aimed to identify reliable maternal serum lipid biomarkers to diagnose
   nsCLP. Eight-feature selection methods were used to assess the
   dysregulated lipids from untargeted lipidomics in a discovery cohort.
   The robust rank aggregation algorithm was applied on these selected
   lipids. The data were subsequently processed using 7 classification
   models to retrieve a panel of 35 candidate lipid biomarkers. Potential
   lipid biomarkers were evaluated using targeted lipidomics in a
   validation cohort. Seven classification models and multivariate
   analyses were constructed to identify the lipid biomarkers for nsCLP.
   The diagnostic model achieved high performance with 3 lipids in
   determining nsCLP. A panel of 3 lipid biomarkers showed great potential
   for nsCLP diagnosis. FA (20:4) and LPC (18:0) were also significantly
   downregulated in early serum samples from the nsCLP group in the
   additional validation cohort. We demonstrate the applicability and
   robustness of a machine-learning algorithm to analyze lipidomic data
   for efficient and reliable biomarker screening.

   Keywords: Development, Metabolism

   Keywords: Cholesterol, Embryonic development
     __________________________________________________________________

   [48]graphic file with name jciinsight-10-186629-g135.jpg
     __________________________________________________________________

   This study provides an approach for the early identification of foetal
   nsCLPs based on untargeted and targeted lipidomics combined with the
   machine learning algorithm.

Introduction

   Nonsyndromic cleft lip with palate (nsCLP) is a common orofacial
   abnormality with a global incidence of 1 per 700 live births ([49]1).
   Children with nsCLP are faced with hearing, eating, speech, and dental
   problems that require multiple surgeries, inducing a heavy financial
   burden and lifetime psychological problems for patients and families
   ([50]2, [51]3). The exact etiology of nsCLP involves a combination of
   genetic and environmental risk factors ([52]4). Metabolic
   abnormalities, including maternal diabetes and obesity, increase the
   risk of orofacial abnormality ([53]5, [54]6). Current clinical
   strategies to diagnose nsCLP prenatally rely on ultrasound as early as
   18–24 weeks of pregnancy ([55]7). In China, the common gestational age
   for ultrasound is 22–26 weeks, depending on the consensus of experts
   and clinical practice. Ultrasound scan is performed in large prenatal
   diagnosis centers, and the success of examination is profoundly limited
   by maternal obesity, fetal position, equipment quality, and technician
   skill ([56]2). More importantly, the palate is often not adequately
   visualized for acoustic shadowing, resulting in misdiagnoses of the
   cleft palate, and appropriate diagnoses are often not established until
   after birth ([57]8).

   Liquid biopsy-based tests have emerged as promising noninvasive
   diagnostic methods for congenital malformations. In clinical practice,
   maternal serum α-fetoprotein is used to screen for fetal neural tube
   defects. Such strategies have also been used for the prenatal diagnosis
   of fetal chromosomal aneuploidies, including detection of cell-free
   fetal DNA in maternal plasma. However, effective prenatal biomarkers
   for other congenital malformations, including nsCLPs, are still
   lacking. Prenatal diagnosis can give expecting parents time to prepare
   psychological and financial support for multidisciplinary surgical
   interventions, helping them avoid termination of the pregnancy. Human
   fetuses can heal from these interventions without visible scarring
   ([58]9), and prenatal diagnosis combined with the development of fetal
   surgery and other therapeutic strategies could provide potential
   benefits for fetuses with nsCLP in the future. Therefore, a highly
   accurate, sensitive, and noninvasive serum testing strategy is urgently
   warranted to facilitate the prenatal diagnosis of nsCLP in clinical
   settings.

   Compared with genomics, proteomics, and transcriptomics, lipidomics is
   still in its early stages. Owing to the high metabolite coverage of
   untargeted lipidomics and reliability of targeted lipidomics, the
   integration of both assays makes for a powerful disease-related lipid
   biomarker study ([59]10). Recently, with the continuous development of
   efficient, high-throughput, and high-sensitivity analytical methods,
   lipidomics has been widely used to identify biomarkers in small volumes
   of plasma or serum for disease detection, prediction, and patient
   stratification ([60]11, [61]12). Lipids contain several key components
   of the cell membrane structure, energy storage sources, and signaling
   molecules. Lipid homeostasis plays a critical role in supporting the
   physiological processes of cells and organisms. For instance, the
   dysregulation of lipid metabolism has been linked to fetal
   neurodegeneration and heart defects ([62]13, [63]14). Long-chain
   polyunsaturated fatty acids (FA) and their metabolites are involved in
   the proper patterning of the dorsoventral axis of zebrafish ([64]15).
   Our previous study found that 7 lipid biomarkers showed significant
   differences in patients following abortion and term birth after in
   vitro fertilization and embryo transfer ([65]16). Palatal tissue
   lipidomics were analyzed in all-trans retinoic acid–treated cleft
   palate mice and normal mice, and unsaturated TGs were found to have a
   profound impact on palatogenesis ([66]17). Embryonic development is
   strictly regulated by lipid metabolism, and lipid-related research is
   important for exploring biomarkers and the etiology of human nsCLP.

   In the age of “big” data, machine learning (ML) is an important branch
   of artificial intelligence that can learn patterns from data; trained
   models can be used to predict outputs. The application of ML approaches
   may broaden and deepen our understanding of prenatal diagnosis, thus
   allowing us to improve delivery management and safety of mothers and
   newborns ([67]18, [68]19). The combination of ML and lipidomics is an
   attractive and promising diagnostic approach. ML is utilized to
   recognize predictive patterns and actively used in lipidomics for data
   processing ([69]20). Recent studies have illustrated that ensemble ML
   algorithms exhibit good performance in disease detection. An ensemble
   ML model of 9 individual classification models was built based on the
   maternal serum metabolomic profile to detect the presence of birth
   defects ([70]21). The ensemble feature selection (ensemble-FS) strategy
   has proven to be efficient in selecting key features compared with
   typical feature selection methods ([71]22). Li et al. designed an
   online tool based on ensemble-FS ranking to discover optimal proteomic
   features ([72]23). However, no relevant studies have used the
   ensemble-FS strategy for lipidomic data analysis.

   This study aimed to identify prenatal lipid biomarkers using a 2-step
   approach. First, a broad untargeted lipidomics approach was used in a
   discovery cohort, which has the potential to identify a wide range of
   lipid features. We employed the robust rank aggregation (RRA) approach
   to rank lipid features selected by 8-feature selection methods
   ([73]24), followed by evaluating the RRA-ranked lipid features using 7
   classification models. The second step involved a target
   lipidomics-based multiple-reaction monitoring (MRM) approach, in which
   the lipid features selected via the ML models in the first step are
   screened in the validation and additional validation cohorts. In this
   study, we evaluated the diagnostic accuracy of ML models based on
   maternal serum untargeted and targeted lipid signatures for the
   detection of prenatal lipid biomarkers of fetal nsCLP.

Results

Lipid profiling based on untargeted lipidomics.

   The distribution of the relative standard deviation (RSD) of the peak
   areas for all identified lipids in the quality control (QC) samples
   indicated that 98% in both positive and negative modes of the lipid
   features had an RSD of ≤ 30% ([74]Supplemental Figure 1A; supplemental
   material available online with this article;
   [75]https://doi.org/10.1172/jci.insight.186629DS1). Results showed
   satisfactory stability and repeatability of untargeted lipidomic
   analysis. A total of 31 lipid classes and 1,302 lipid species was
   detected in 60 serum samples collected at a gestational age of 22–26
   weeks. Phosphatidylcholine (PC; n = 348), triglyceride (TG; n = 317),
   sphingomyelin (SM; n = 174), phosphatidylethanolamine (PE; n = 104),
   and lysophosphatidylcholine (LPC; n = 78) were the top 5 classes with
   the highest number of lipid species ([76]Figure 1A). Among the
   untargeted lipidomic data, 6 lipid classes were significantly
   dysregulated in the sera of pregnant women carrying fetuses with nsCLP
   compared with pregnant individuals in the control group. Acyl carnitine
   and sulfatide were significantly upregulated, while CerG2GNAc1,
   cholesterol ester, coenzyme (Co), and LPC were significantly
   downregulated in the nsCLP group ([77]Figure 1, B–G). In total, 1,302
   lipids were detected, including 19 upregulated, 162 downregulated, and
   1,121 nonsignificant lipids in the sera of pregnant women carrying
   fetuses with nsCLP compared with individuals in the control group
   ([78]Figure 1H). The orthogonal partial least squares-discriminant
   analysis (OPLS-DA) model was established through 7-fold
   cross-validation to obtain a list of features for defining the
   variation in lipid profiles between the nsCLP and control groups
   ([79]Figure 1I). Even though the permutation test of the OPLS-DA model
   showed no overfitting (R^2 intercept was 0.541, and the Q^2 intercept
   was –0.528; [80]Figure 1J), the R^2X, R^2Y, and Q^2 values were 0.395,
   0.662, and 0.311, respectively, indicating that the stability and
   reliability of the OPLS-DA model were unsatisfactory. To further
   improve diagnostic accuracy, we subsequently used multiple ML
   approaches to establish optimal prenatal lipid biomarkers for fetal
   nsCLP.

Figure 1. Serum lipid profiles identified by the untargeted lipidomics
analysis.

   [81]Figure 1
   [82]Open in a new tab

   (A) A total of 1,302 lipid species were partitioned in 31 classes.
   (B–G) Lipid intensity of the 6 differentially dysregulated lipids
   classes (acyl carnitine [AcCa], CerG2GNAc1, cholesterol ester [ChE],
   coenzyme [Co], lysophosphatidylcholine [LPC], and sulfatide [ST]) in
   the control and nsCLP groups. Data are expressed as mean ± SD. n = 30,
   each group. *P < 0.05; **P < 0.01; ****P < 0.0001 by unpaired t test or
   median with interquartile ranges by Mann-Whitney test. (H) Volcano plot
   analysis of differentially expressed lipids in maternal serum from the
   control and nsCLP group (FC ≥ 1.2 or FC ≤ 0.83 and P < 0.05). (I)
   OPLS-DA score plot based on control (green dots) and nsCLP samples
   (blue dots) (n = 30, each group). (J) Statistical validation of the
   OPLS-DA model by permutation testing (200 iterations).

ML algorithm performance in untargeted lipidomics.

   First, 8 feature evaluation methods were performed on the 181
   differentially dysregulated lipid features (fold change [FC] ≥ 1.2 or ≤
   0.83, P < 0.05). Each algorithm was used to independently evaluate all
   lipid features, resulting in a feature-ranking list. After deleting
   lipid isomers and lipids that could not be validated using targeted
   lipidomics, 103 lipid features were included in the RRA ranking fusion
   approach. Subsequently, 103 lipid features were selected using the
   aforementioned ensemble-FS method. The details are provided in the
   [83]Supplemental Data. Seven classification models were used to
   determine the diagnostic performance of the 103 lipid features.
   [84]Figure 2A shows the mean AUC values of different lipid features
   based on different classification models. The highest AUC value among
   the individual models resulting from the naive Bayes (NB)
   classification model was 0.95, when the top 35 lipid features were
   included. [85]Table 1 shows the diagnostic performances of the 35 lipid
   features for each classification model when applied to a discovery
   cohort. To further analyze the performance of the different
   classifiers, we plotted the training and testing curves for different
   classifiers using accuracy as the metric ([86]Figure 2, B–H). These
   results also demonstrated the superiority of the NB classifier. The NB
   classification model did not pose any risk of overfitting, achieving
   high classification accuracy and a balanced distribution on the
   training and testing sets. Tree-based classification models, including
   random forest (RF), Decision tree (DT), and Adaboost (ADA), tended to
   achieve 100% classification accuracy on the training set; however, it
   was difficult to achieve a classification accuracy above 0.8 on the
   testing set, indicating poor performance and a risk of overfitting.
   However, the logistic regression (LR), support vector machine (SVM),
   and K-nearest neighbor (KNN) models did not achieve satisfactory
   accuracy values, even when all samples were included as inputs. The
   correlograms were represented by the Pearson correlation coefficient.
   Among all feature correlations, 75% (449 of 595) had a correlation
   coefficient of less than 0.5, reflecting relatively weak linear
   associations among the majority of features ([87]Figure 2I).

Figure 2. Classification model performance on the discovery cohort.

   [88]Figure 2
   [89]Open in a new tab

   (A) Analysis of the mean AUC changes in 103 lipid features under
   different classifiers. (B–H) Learning curve analysis of the 7 different
   classifiers, including SVM, NB, RF, DT, LR, ADA, and KNN. (I)
   Correlation clustering heatmap for these 35 lipid features identified
   via the NB classifier.

Table 1. Diagnostic performance metrics of the classification models when
applied to the untargeted lipidomics in discovery cohort.

   [90]graphic file with name jciinsight-10-186629-g129.jpg
   [91]Open in a new tab

Multivariate analysis of candidate lipids biomarkers in the discovery cohort.

   The principal component analysis (PCA) model demonstrated a separation
   tendency of 35 lipid features between the control and nsCLP groups in
   the score plot ([92]Figure 3A). A heatmap of the 35 selected features
   demonstrated that the expression levels in the nsCLP group were
   significantly different from those in the control group ([93]Figure
   3B). The average AUC values of NB classifier were repeatedly calculated
   5 times for hierarchical cross-validation. In combination with the 35
   RRA-selected lipid features, the NB classifier achieved optimal
   performance with an AUC value, sensitivity, and specificity of 0.95,
   90%, and 87%, respectively, indicating high discrimination performance
   between the 2 groups ([94]Figure 3C). [95]Figure 3D showed 15 pathways
   enriched by 35 dysregulated lipid features, with bubble size and
   distance to origin varying with lipid counts and –log[10] (P value),
   respectively. Five pathways with a P value of less than 0.05 —
   glycerophospholipid metabolism, ferroptosis, choline metabolism in
   cancer, retrograde endocannabinoid signaling, and pathogenic E. coli
   infection — were considered the most significantly enriched pathways.
   The detailed results of pathway enrichment are summarized in
   [96]Supplemental Table 1.

Figure 3. Differentially expressed lipids and their enriched metabolism
pathways.

   [97]Figure 3
   [98]Open in a new tab

   (A) PCA score plot of 35 lipid features based on control (red dots) and
   nsCLP samples (green dots) (n = 30, each group). (B) Feature heatmap
   analysis of 35 lipid features in the control and nsCLP samples (n = 30,
   each group). (C) ROC curve analysis of the NB classifier of 35 lipid
   features. Data are presented as mean ± SEM. (D) Bubble plot shows 15
   lipid metabolism pathways enriched by 35 lipids according to lipid hits
   and –log[10] (P value).

Performance of ML algorithms in relation to targeted lipidomics.

   The expression levels of 35 lipid features were tested using targeted
   lipidomic analysis in a validation study with an independent
   population, and 16 lipids were detected as differentially expressed
   features. Results revealed that the sum of the responses for these
   lipids with a QC RSD percentage of ≤ 30% reached 80%, demonstrating
   good stability for targeted lipidomics ([99]Supplemental Figure 1B). A
   heatmap of the 16 selected features demonstrated that the expression
   levels in the nsCLP group were significantly different from those in
   the control group ([100]Figure 4A). A correlogram showed that 8 LPC
   features were associated with metabolic disorders in the nsCLP group
   ([101]Figure 4B). Two pathways (glycerophospholipid metabolism and
   choline metabolism in cancer) were significantly enriched in the bubble
   dot diagram ([102]Figure 4C). The detailed results of pathway
   enrichment are summarized in [103]Supplemental Table 2. Seven
   classification models were used to build the diagnostic performance of
   the 16 lipid features in the targeted analysis. [104]Figure 4D shows
   the AUC values for lipid features based on different classification
   models. The highest AUC value among the individual models resulting
   from the NB classification model was 0.97, when the top 3 lipid
   features (FA [20:4], LPC [18:0], and PC [16:0e/22:0]) were included.
   [105]Table 2 showed the diagnostic performance of each classification
   model when applied to the validation cohort. Consistent with our
   previous findings from the untargeted analysis, the NB classifier
   demonstrated a clear advantage among the 7 classification models
   ([106]Figure 4E). The individual training curves of the other 6
   classification models are shown in [107]Supplemental Figure 2.

Figure 4. Functional analysis and classification model performance on
targeted lipidomic data.

   [108]Figure 4
   [109]Open in a new tab

   (A) Feature heatmap analysis of 16 lipid features in the control and
   nsCLP samples (n = 20, each group). (B) Correlation clustering heatmap
   for these 16 lipid features identified by targeted lipidomics (n = 20,
   each group). (C) Bubble plot showing 10 lipid metabolism pathways
   enriched by 16 lipids according to lipid hits and –log[10] (P value).
   (D) Analysis of the mean AUC changes in 16 lipid features under
   different classification models. (E) Learning curve analysis of the NB
   classifier.

Table 2. Diagnostic performance metrics of the classification models when
applied to the targeted lipidomics in validation cohort.

   [110]graphic file with name jciinsight-10-186629-g130.jpg
   [111]Open in a new tab

Lipid features as prenatal biomarkers for nsCLP.

   The PCA score plot of the 3 lipid features demonstrated significant
   separation between the control and nsCLP groups, indicating a
   dramatically dysregulated lipid profile ([112]Figure 5A). Feature
   heatmap analysis showed that the 3 lipid features in the nsCLP group
   were significantly different from those in the control group
   ([113]Figure 5B). Partial dependence plots (PDP) serve as a valuable
   tool for illustrating the impact of a specific feature on model
   predictions. To verify the validity of the results, we conducted a PDP
   analysis of the features selected using the NB classifier model. The 3
   lipid features were obviously steep, indicating that the features had a
   substantial influence on prediction results ([114]Figure 5, C–E).
   Therefore, these lipids contributed significantly to the model and
   validated the results. The expression levels of 3 lipid features (FA
   [20:4], LPC [18:0], and PC [16:0e/22:0]) were significantly
   downregulated in the nsCLP group ([115]Figure 5, F–H). The mean AUC for
   the 3 lipid features was 0.97, with a sensitivity of 80% and
   specificity of 84% ([116]Figure 5I). To explore the prenatal diagnosis
   efficiency of the 3 lipid biomarkers, we collected serum samples from
   women carrying fetuses with nsCLP in the early stages of pregnancy. The
   results showed that the expression levels of FA (20:4) and LPC (18:0)
   were also significantly downregulated in sera from the pregnant women
   carrying fetuses with nsCLP with gestational ages of 16–17 weeks
   compared with that in the control group ([117]Supplemental Figure 3).

Figure 5. Three lipids biomarkers for nsCLP.

   [118]Figure 5
   [119]Open in a new tab

   (A) PCA score plot of 3 lipid features based on control (red dots) and
   nsCLP samples (green dots) (n = 20, each group). (B) Feature heatmap
   analysis of 3 lipid features (n = 20, each group). (C–E) PDP analysis
   for the 3 lipid features (FA [20:4], LPC [18:0], and PC [16:0e/22:0]).
   (F–H) Relative expression of the 3 lipid features in nsCLP compared
   with control group. Data are presented as mean ± SD. n = 20, each
   group. *P < 0.05; ****P < 0.0001, by unpaired t test. (I) ROC curve
   analysis of NB classifier of 3 lipid features. Data are presented as
   the mean ± SEM. n = 20, each group.

Discussion

   In this study, a combination of lipidomics and ML was used to screen
   serum lipid biomarkers for the diagnosis of fetal nsCLP. Three lipid
   features were ultimately identified as biomarkers, and a diagnostic
   model based on a panel of these 3 features was built and evaluated
   using a validation cohort. Our findings provide a noninvasive and
   efficient diagnostic method for fetal nsCLP as well as a study workflow
   that combines lipidomics and ML that can be used to efficiently and
   reliably screen for biomarkers. In the discovery cohort of this study,
   extensive lipid features were screened using untargeted lipidomic
   analysis. We employed 8-feature selection methods to rank all lipid
   features according to their relative contributions in discriminating
   the different groups. The RRA approach was subsequently used to
   generate a unified ranking list based on the selected features via the
   8-feature selection methods. The 7 classification models provided good
   overall accuracy in terms of the selected lipid features ranked using
   the RRA approach. We further employed an NB-based classification model
   to screen the lipidomics for the following reasons. (a) The model is
   based on the assumption of the conditional independence of features to
   perform accurate classification with a small sample size; thus, it is
   suitable for our lipidomics ([120]25). (b) The NB classifier is robust
   against noisy data: as the underlying algorithm is based on a
   probabilistic model, it is insensitive to noise or outliers in the
   input data ([121]26). (c) In the present study, the NB classifier was
   ranked as the best among the 7 classification models, with the lowest
   number of lipid biomarkers in relation to AUC, sensitivity, and
   specificity. To validate the panel of lipid biomarkers and ultimately
   apply it in clinical settings, we performed targeted lipidomic analysis
   on 35 potential lipid biomarkers in a validation cohort and identified
   16 differentially dysregulated lipid features. The 7 included
   classifications were used to analyze 16 lipid biomarkers, and the NB
   classifier was the optimal classifier in the validation cohort. Our
   results showed good diagnostic power in the validation cohort with the
   3 lipid features.

   In recent years, more and more studies have applied ML algorithms to
   medical applications, taking advantage of their remarkable sensitivity
   and promising potential for diagnosis. For example, Zhou et al.
   implemented an SVM-based ML algorithm to search a panel of 11 plasma
   lipids for malignant brain gliomas ([122]27). ML approaches have also
   been used to identify and predict common birth defects, including
   congenital heart defects ([123]28). Our previous study used an SVM
   classifier to establish a prediction model based on a combination of 3
   complement proteins and α-fetoprotein, which could serve as optimal
   prenatal biomarkers for neural tube defects ([124]29). This study
   employed a 2-stage ML approach. Initially, a classification model was
   developed to identify potential lipid biomarkers from untargeted
   lipidomics. Subsequently, a diagnostic model was constructed using
   targeted lipidomics from the validation cohort to develop a robust
   diagnostic framework for fetal nsCLP detection. To enhance the
   stability and effectiveness of biomarker selection and eliminate the
   randomness and bias caused by a single-feature selection method, we
   integrated 8-feature selection algorithms and used the RRA method to
   combine the results of different algorithms to obtain optimal feature
   ranking. To address various factors that may influence ranking lists in
   different feature selection methods, including individual errors and
   method preference differences, RRA transforms the ranking lists into
   rank values. This approach effectively mitigates the impact of noise
   and outliers, thereby improving overall robustness. In addition, we
   tested 7 classification models and identified the optimal model for
   diagnosing nsCLP.

   The candidate lipid biomarkers belonged to 3 lipid classes (PC, LPC,
   and FA). PC represents a key molecular constituent of cellular
   membranes, pulmonary surfactants, circulating lipoproteins, and biliary
   secretions. The role of PC in emulsifying and decomposing lipids exerts
   a positive protective effect on the cardiovascular system ([125]30). PC
   is a source of lipid messengers and can be catalyzed by secretory
   phospholipase A2 to produce LPC. These data further explain the results
   of the downregulated levels of both PC (16:0e/22:0) and LPC (18:0) in
   the present study. LPC represents a class of oxidized low-density
   lipoprotein components whose homeostasis is maintained through
   enzymatic regulation. This lipid class modulates cellular functions via
   apoptosis and oxidative stress pathways. Specifically, the LPC (18:0)
   displays unique immunoregulatory effects characterized by
   proinflammatory cytokine suppression and antiinflammatory mediator
   enhancement ([126]31). More importantly, LPC is a potential therapeutic
   agent for sepsis ([127]32). Previous studies have reported that
   proinflammatory conditions, including maternal infections, can affect
   orofacial development ([128]33, [129]34). High levels of intracellular
   reactive oxygen species in mouse embryonic palatal cells may eventually
   result in nsCLP ([130]35). These data further support the decrease in
   PC (16:0e/22:0) and LPC (18:0), which reduced the antiinflammatory
   effects during orofacial development. To our knowledge, to date, no
   studies have associated PC and LPC with fetal orofacial development.
   Furthermore, glycerophospholipid metabolism was ranked at the top in
   both untargeted and targeted analyses and was closely related to a
   group of significant lipid features identified in this study,
   indicating that PC (16:0e/22:0) and LPC (18:0), may contribute to the
   developmental biology of lip and palate formations. The free LPC is
   reacylated by yielding FA, a source of energy, membrane components, and
   signaling molecules ([131]36). PUFAs regulate the antioxidant signaling
   pathways and modulate inflammatory processes ([132]37). Omega-3 and
   omega-6 PUFAs constitute the principal PUFA categories with substantial
   health implications. Inadequate levels of omega-3 PUFA negatively
   influence brain function, including behavior, learning, and cognition
   ([133]38). Omega-3 PUFA-enriched diets prevent phenytoin-induced cleft
   palate in mid-gestational mice ([134]39), while FA (20:4), an omega-6
   PUFA, critically regulates neurodevelopment during pregnancy and early
   childhood, demonstrating the importance of FA balance throughout
   development ([135]40).

   This study has some limitations. ML-based lipidomics analysis was built
   and validated by the discovery, validation, and additional validation
   cohorts, in which all pregnant women from a single center were
   included. The suitability of this approach for prenatal lipid diagnosis
   in other populations requires further investigation. Therefore,
   multicenter validation studies with large cohorts are required to
   confirm the clinical utility of these lipid biomarkers before clinical
   implementation. In humans, palatogenesis is initiated in the early
   sixth week of gestation and palatal fusion is completed by the twelfth
   week of gestation ([136]41). To explore early prenatal biomarkers, we
   established a prospective birth cohort study. Due to the low incidence
   of nsCLP, a small number of specimens were included. A third limitation
   was the lack of a robust relationship among the panel of 3 lipids
   involved in orafacial development. As the mechanism of abnormal lipid
   metabolism in nsCLP is very complex, whether it is caused by genetic
   factors or environmental factors needs to be further studied. Further
   studies are needed to fully evaluate the mechanisms underlying the 3
   lipid features associated with nsCLP development.

   In conclusion, this study provides an approach for the identification
   of fetal nsCLPs based on untargeted and targeted lipidomics combined
   with the ML algorithm. These lipid biomarkers and NB-based ML algorithm
   exhibited strong accuracy in identifying fetal nsCLP, highlighting
   their value in prenatal diagnosis. The advantages of this diagnostic
   strategy include minimally invasive sampling, rapid analytical
   efficiency, and reliable accuracy. Furthermore, our study demonstrated
   the feasibility, strengths, and promising potential of a strategy
   combining lipidomics and ML algorithms for identifying disease-related
   biomarkers.

Methods

Sex as a biological variable

   Sex was not considered as a biological variable. For discovery,
   validation and additional validation cohorts, the serum samples from
   pregnant women were tested.

Study population and sample collection

   Fifty-three pregnant women carrying fetuses with nsCLP and 53 pregnant
   women carrying normal fetuses for discovery, validation, and additional
   validation cohorts were obtained from Shengjing Hospital of China
   Medical University. The discovery cohort comprised 30 pregnant women
   carrying fetuses with nsCLP and 30 pregnant women carrying normal
   fetuses. In the discovery cohort, lipids were profiled using untargeted
   lipidomics. The validation cohort included 20 pregnant women with nsCLP
   and 20 pregnant women with normal fetuses. For lipid biomarker
   validation, targeted lipidomics were performed in the validation
   cohort. Serum samples in the discovery and validation cohort were
   obtained at 22–26 weeks of pregnancy. The additional validation cohort
   included 3 pregnant women with fetuses with nsCLP and 3 pregnant women
   with normal fetuses. Serum samples in the additional validation cohort
   were obtained at 16–17 weeks of pregnancy. Serum samples for lipidomic
   analysis were obtained from routine venipuncture after ultrasound
   examinations in an ongoing birth cohort study, which was reported in a
   previous study ([137]42). Blood was centrifuged within 30 minutes of
   collection, and the isolated sera were stored at –80°C until analysis.
   The study workflow is presented in Figure 6. All participants included
   in this study were matched for gestational age, maternal age, and BMI
   ([138]Table 3). There was no statistical difference in the sex of the
   fetuses ([139]Supplemental Table 3). Gynecologists and fetal
   ultrasonologists at our hospital performed evaluations to ensure
   accuracy of the final diagnosis. Women with multiple pregnancies,
   pregnancy-related complications, or metabolic diseases were excluded.

Table 3. Clinical features of pregnant women included in lipidomic analysis.

   [140]graphic file with name jciinsight-10-186629-g131.jpg
   [141]Open in a new tab

Untargeted lipidomic analysis

Lipid sample preparation.

   For untargeted lipidomic analysis, 100 μL aliquots of each pregnant
   woman’s serum sample were mixed with 240 μL precooled methanol, 800 μL
   methyl tert-butyl ether (MTBE), and 200 μL water; incubated for 20
   minutes; and centrifuged for 15 minutes at 8,000g. The the upper lipid
   layer was obtained and then dried by nitrogen gas and mixed with 240 μL
   isopropanol, followed by centrifugation for 10 minutes at 8,000g under
   the optimized column temperature at 10°C. For untargeted and targeted
   lipidomics, equivalent aliquots of each serum sample from the control
   and nsCLP groups were mixed as QC samples and used to monitor the
   repeatability of sample preparation and stability of the lipidomic
   system throughout the experiment.

Chromatographic and mass spectrometric conditions.

   Untargeted lipidomic analysis was performed using an
   ultra-high-performance liquid chromatography system (UHPLC; Nexera
   LC-30A) equipped with a mass spectrometer (Q Exactive Plus; Thermo
   Scientific) at Shanghai Applied Protein Technology Co. Ltd.
   Reverse-phase chromatography was selected for LC separation using a CSH
   C18 column (2.1 mm × 100 mm, 1.7 μm, Waters). The column temperature
   was set at 45°C, and flow rate was set at 300 μL/min. The aqueous
   components comprised 10 mM ammonium acetate in acetonitrile/water (6:4,
   v/v) as the mobile phase A and 10 mM ammonium acetate in
   acetonitrile/isopropanol (1:9, v/v) as the mobile phase B. The gradient
   elution program was as follows: from 0 to 2 minutes, B was maintained
   at 30%; from 2.1 to 25 minutes, B linearly changed from 30% to 100%;
   and from 25.1 to 30 minutes, B was maintained at 30%. During the entire
   analysis process, the samples were placed in an autosampler at 10°C. In
   order to avoid the influence caused by fluctuations in the instrument’s
   detection signals, continuous analysis of the samples was carried out
   in a random order. In the sample queue, a QC sample was placed after
   every 8 experimental samples, which was used to monitor and evaluate
   the stability of the system as well as the reliability of the
   experimental data. Lipid species were measured by electrospray
   ionization mass spectrometry in the positive and negative ion modes. If
   a lipid is identified in both modes, the mode with high signal strength
   and stable ionization efficiency was selected for quantification.
   Experiments in the positive ion mode were performed as follows: heater
   temperature, 300°C; sheath gas glow rate, 45 AU; aux gas flow rate, 15
   AU; sweep gas flow rate, 1 AU; spray voltage, 3.0 KV; capillary temp,
   350°C; and S-Lens RF level, 50%. The MS scan ranged from 200 to 1,800.
   The negative ion mode was performed using the following parameters:
   heater temperature, 300°C; sheath gas flow rate, 45 AU; aux gas flow
   rate, 15 AU; sweep gas flow rate, 1 AU; spray voltage, 2.5 KV;
   capillary temp, 350°C; and S-Lens RF level, 60%. The MS scan ranged
   from 250 to 1,800. Peak matching, retention time alignment, and peak
   area extraction were performed using the LipidSearch software version
   4.1 (Thermo Scientific). The major parameters were set as follows:
   precursor tolerance, 5 ppm; product tolerance, 5 ppm; and product ion
   threshold, 5%. LipidSearch contains data on more than 30 lipid classes,
   with information on more than 1,700,000 ion fragments. For the data
   extracted by LipidSearch, missing values for more than 50% of lipid
   molecules were deleted.

Targeted lipidomic analysis

Lipid sample preparation.

   For targeted lipidomic analysis, 50 μL aliquots of each pregnant
   woman’s serum sample were mixed with 200 μL methanol, 20 μL internal
   lipid standards, and 800 μL MTBE. The mixture was adequately vortexed,
   sonicated for 20 minutes at 4°C, and subsequently kept for 30 minutes
   in room temperature. Subsequently, 200 μL MS-grade water was added.
   After vortex mixing for 20 seconds, the mixture was vortexed and
   centrifuged at 18,800g at 4°C for 15 minutes. The upper organic solvent
   layer was obtained and dried under the nitrogen atmosphere. The residue
   was reconstituted with 200 μL of 90% isopropanol/acetonitrile mixture
   and centrifuged at 18,800g at 4°C for 15 minutes. The supernatant was
   transferred to an autosampler vial and used for subsequent detection.

Chromatographic and mass spectrometric conditions.

   UHPLC coupled with triple-quadrupole linear ion-trap tandem mass
   spectrometry was used for targeted lipidomic analyses. Analysis was
   performed using a UHPLC system (Nexera LC-30A, Shimadzu) and separated
   on HILIC (Phenomenex, Luna NH2, 2.0 mm × 100 mm, 3 μm) and C18
   (Phenomenex, Kinetex C18, 2.1 mm × 100 mm, 2.6 μm) columns. The HILIC
   NH[2] column temperature was maintained at 40°C, and flow rate was set
   at 0.4 mL/min. The aqueous components comprised 10 mM ammonium acetate
   in methanol/acetonitrile (5:5, v/v) as the mobile phase A and 10 mM
   ammonium acetate in acetonitrile/water (5:5, v/v) as the mobile phase
   B. The gradient program was performed using the following parameters:
   0–3 minutes, 3% B; 3–13 minutes, 3% B to 100% B; 13–17 minutes, 100% B;
   and 17–22 minutes, 3% B. The C18 column temperature was set at 40°C and
   at a flow rate of 0.35 mL/min. The aqueous component comprised 5 mM
   ammonium acetate in acetonitrile/water (7:3, v/v) as mobile phase A and
   IPA solution (isopropanol + 5 mM ammonium acetate) as mobile phase B.
   The gradient duration program was as follows: from 0 to 5 minutes, 20%
   B was changed to 60% B; from 5 to 13 minutes, 60% B was changed to 100%
   B; and from 13 to 17 minutes, B was maintained at 20%. The samples were
   placed at 10°C during the whole analysis process. 6500+QTRAP (AB SCIEX)
   was used in positive and negative switch modes. ESI was performed using
   the following parameters: source temperature, 400°C; ion source gas 1,
   50; ion source gas 2, 55; curtain gas, 35; and ionSpray voltage, +5,500
   V or –4,500 V in positive or negative modes, respectively. The MRM
   method was used for mass spectrometry quantitative data acquisition.
   Sciex OS software was used for the peak extraction of raw MRM data to
   obtain each lipid and ratio of peak area to internal standard peak
   area. The content of each lipid was subsequently calculated.

ML

Feature preliminary filtering method.

   We applied the 3-sigma principle to detect outliers in all data. For
   all identified outliers, we used the KNN algorithm with a neighbor
   count of 5 to fill in the missing values and outliers. Moreover, we
   performed a log[2] transformation on all data to reduce the scale
   disparities among features and to help achieve a more stable and
   comparable distribution for subsequent statistical analyses and ML
   modeling. Thereafter, we performed an initial statistical filtering
   step to focus on differentially expressed lipids. Specifically, we
   applied a significance level of P < 0.05 and FC thresholds (FC ≥ 1.2 or
   ≤ 0.83) based on commonly accepted standards and empirical evidence
   ([142]43, [143]44). This approach helps reduce noise, lower
   computational complexity, and ensure that subsequent ML analyses
   prioritize lipid features with greater biological relevance. These
   criteria were used to generate a subset of features that satisfied the
   filtering conditions.

Ensemble feature selection.

   To avoid the overfitting and variability in credibility caused by the
   use of a single method, we used the ensemble-FS method. Specifically,
   datasets were represented using the following formula: D[M × N] = X[1],
   X[2]... X[i]..., X[M]^, X[i] = (x[i]^(1), x[i]^(2)..., x[i] ^(j)...,
   x[i] ^(N))^T, where M denotes the number of samples and N denotes the
   number of features, X[i] represents the expression values of all
   features for the i-th sample, and x[i]^(1) represents the expression
   value matrix of the j-th feature in the i-th sample. The objective of
   the feature selection method was to construct a feature evaluation
   model, represented as EV(*), through which the data are evaluated to
   produce a subspace X[i]^* = (x[i]^(1), x[i] ^([144]2)..., x[i]^(s))^T
   (S < N) of the original features. The asterisk in EV(*) denotes a
   placeholder, indicating that this evaluation function, EV, can operate
   on any feature or feature subset within our dataset. The letter S
   represents the number of selected features after dimensionality
   reduction, which is smaller than the original feature dimension N. We
   employed 8 ML algorithms that can output feature rankings to form a
   multidimensional feature evaluator. Each algorithm was used to
   independently evaluate all lipid features, resulting in a
   feature-ranking list. These methods include L1 regularization (Lasso),
   L2 regularization (Ridge), RF, linear regression (LR), correlation, DT,
   χ^2 test, and maximal information coefficient. To ensure that different
   feature selection algorithms obtain a reasonable number of features,
   the parameter settings were as follows. For the Lasso and Ridge
   methods, the penalty coefficient was set to 0.01, and the stopping
   criterion was set to 0.001. For the DT and RF methods, the Gini
   coefficient was used as the criterion for node splitting. The maximum
   depth was set to none, min_samples_split was set to 2, and
   min_samples_leaf was set to 1. The other 4 methods used the default
   parameters in sklearn 1.2.2.

   The RRA approach was used to combine the evaluation results from
   multiple feature evaluators and generate a unified ranking list. The
   features are ranked by scores, resulting in a final feature ranking. In
   our previous study, we combined this method with a genetic algorithm
   for feature-selection tasks using microarray data ([145]45). We
   subsequently input the corresponding feature-ranking results into the
   RRA algorithm to obtain a unified feature ranking. This method detected
   lipid features that were ranked higher than expected under the null
   hypothesis of irrelevance and assigned a significance score to each
   feature.

Classification model

   To identify the optimal ML model for distinguishing between nsCLP and
   control samples using lipid features, 7 different classifier models
   were constructed as follows: SVM, NB, RF, DT, LR, ADA, and KNN. All the
   classification models also used the default parameters in sklearn
   1.2.2. In both untargeted and targeted lipidomics, the samples were
   stratified into the training and testing sets in a 4:1 manner, and the
   classifier models were subsequently trained using a 5-fold
   cross-validation method.

Statistics

   The ML experiments were conducted on a Windows 11 system equipped with
   an Intel Core i7-12700H CPU, 32 GB of RAM, and a GTX 1060 GPU. For all
   ML experiments, to ensure the reliability of results, a 5-fold
   cross-validation method was employed, and the average value of the
   results was considered as the final result. The implementation of the
   ensemble-FS was based on scikit-learn version 1.1.2. The OPLS-DA model
   and permutation test of the OPLS-DA model were performed using SIMCA-P
   software (version 14.1; Umetrics). Pathway enrichment analysis was
   conducted using LIPEA ([146]https://hyperlipea.org/home) with the Kyoto
   Encyclopedia of Genes and Genomes database
   ([147]https://www.genome.jp/kegg/), and results are represented by a
   bubble plot ([148]Figure 3D and [149]Figure 4C). PCA and receiver
   operating characteristic (ROC) curves were used to assess the
   predictive performance of potential lipid biomarkers using Python 3.9.
   Two groups of data with Gaussian distributions were compared using
   2-tailed Student’s t tests. Otherwise, nonparametric Mann-Whitney tests
   were used. A P value of less than 0.05 was considered significant.
   Prism Graphpad software (version 7.0) and SPSS 22.0 were used for
   statistical analysis.

Study approval

   All clinical studies were conducted according to the Declaration of
   Helsinki principles. Written informed consent was obtained from all
   pregnant women, and the study was approved by the Ethics Committee of
   the Shengjing Hospital (approval 2017PS264K).

Data availability

   Details regarding lipid features at every step of the ML approach are
   available in the [150]Supplemental Data. Values for all data points
   shown in graphs are reported in the [151]Supporting Data Values file.
   The raw data supporting the conclusions of this article will be made
   available from the corresponding author upon request.

Author contributions

   SJ, WX, W Li, and ZY designed the study. SJ and CY contributed to data
   analysis. YD prepared serum samples. WX and W Li performed ML
   experiments. SJ and WX wrote the first draft of the paper. W Luo, HG,
   XW, WM, DL, SC, and YB contributed to lipidomic analysis. The
   authorship order among co–first authors (SJ first and WX second) was
   based on their contributions to drafting the manuscript. All authors
   read and approved the final manuscript.

Supplementary Material

   Supplemental data
   [152]jciinsight-10-186629-s132.pdf^ (1.2MB, pdf)
   Supporting data values
   [153]jciinsight-10-186629-s133.xlsx^ (173KB, xlsx)

Figure 6. Study design using untargeted and targeted lipidomics analysis
combined with machine learning.

   [154]Figure 6
   [155]Open in a new tab

   Untargeted lipidomics: LC-MS–based untargeted lipidomics were initially
   used for serum lipid profiling in a discovery cohort, and the results
   were processed using ensemble feature selection methods and 7
   classification models to screen for potential lipid biomarkers. The top
   35 lipid features were selected as potential biomarkers for further
   validation using targeted lipidomics. NB classifier: An optimal
   classification model was constructed using untargeted and targeted
   lipidomic data. Targeted lipidomics: Absolute quantitative
   lipidomics-based MRM was used for serum lipid profiling in the
   validation and additional validation cohort, and the results were
   processed using multivariate analysis. (a) Fifteen lipid features were
   excluded as they could not be validated in the future targeted
   lipidomics. (b) Forty-four lipid features were excluded from isomers.
   (c) These 2 lipid features were excluded because the expression trend
   was opposite in untargeted lipidomics.

Acknowledgments