Machine Learning Reveals Lipidome Remodeling Dynamics in a Mouse Model of Ovarian Cancer

Abstract

Ovarian cancer (OC) is one of the deadliest cancers affecting the female reproductive system. It may present little or no symptoms at the early stages, and typically unspecific symptoms at later stages. High-grade serous ovarian cancer (HGSC) is the subtype responsible for most ovarian cancer deaths. However, very little is known about the metabolic course of this disease, particularly in its early stages. In this longitudinal study, we examined the temporal course of serum lipidome changes using a robust HGSC mouse model and machine learning data analysis. Early progression of HGSC was marked by increased levels of phosphatidylcholines and phosphatidylethanolamines. In contrast, later stages featured more diverse lipids alterations, including fatty acids and their derivatives, triglycerides, ceramides, hexosylceramides, sphingomyelins, lysophosphatidylcholines, and phosphatidylinositols. These alterations underscored unique perturbations in cell membrane stability, proliferation, and survival during cancer development and progression, offering potential targets for early detection and prognosis of human ovarian cancer.

Teaser: Time-resolved lipidome remodeling in an ovarian cancer model is studied through lipidomics and machine learning.

Figure 1.. Blood sampling scheme, study design, and analysis plan.

(a) Blood samples were collected every two weeks, starting at the two-month mark. Lipidomics experiments were conducted using ultra-high performance liquid chromatography mass spectrometry (UHPLC-MS). (b) Conversion of the mice age in weeks to percentage lifetime makes lipidomic comparisons effective. (c) Computational analysis plan.

Fig 2.. Global lipidomic changes observed upon HGSC progression.

(a) Overall fold changes for all annotated features, for all time points combined. (b) Fold changes for 87 significant lipid features (Welch’s T-test, Benjamini-Hochberg correction q-value < 0.05) for all time points combined. (c) The number of significant lipidomic features (Welch’s T-test p-value < 0.05) for each lifetime stage. (d) Upset plot showing overlapping significant lipids in various lifetime stages. Sets containing lipid features present in at least three lifetime stages are colored brown.

Fig 3.. Lipidome changes in response to ovarian cancer progression.

(a) Hierarchical clustering analysis shows the grouping of lipidome trajectories into four types of clusters. (b) Longitudinal lipid changes for the selected clusters indicating fold changes. (c) Network graph for the clusters shown in (a). Nodes represent lipids, while the links connect nodes with a high Pearson’s correlation (r ≥0.5).

Fig 4.. Discriminating DKO from DKO control mice via machine learning.

(a) Machine learning pipeline. The pipeline starts with a t-test filtering method for each of the five ML tasks: lipid features with less than 0.05 p-value (Welch’s t-test p < 0.05) were selected. Next, one of two lipid features with a high Pearson’s correlation score (r > 0.8) was removed from the dataset to avoid unnecessary redundancies. Finally, lipid features with a Gini index greater or equal to the mean of all Gini indices were selected for training and testing purposes. ROC-AUC test set for DKO classification for (b), lifetime stage I (c), lifetime stage II (d), lifetime stage III (e), lifetime stage IV (f), and lifetime stage V. (g) The best ROC-AUC scores for each lifetime stage. TPR: True positive rate, FPR: False positive rate, k-NN: k-Nearest Neighbors, RF: Random Forests, SVM: Support Vector Machine, Voting: Voting Ensemble Classifier.

Fig 5.. Discriminant lipids for each of the five lifetime stages.

(a) Lifetime stage I: 0–30% lifetime. (b) Lifetime stage II: 30–45% lifetime. (c) lifetime stage III: 45–60% lifetime (d) Lifetime stage IV: 60–75% lifetime. (e) Lifetime stage V: 75–100% lifetime. (f) frequency of lipid classes, groups, and categories in the discriminant lipid panels. TG: Triacylglycerols, FA: Fatty acids, HexCer: Hexosylceramides, LPC: Lysophosphatidylcholines, LPE: Lysophosphatidylethanolamines, PC: Phosphatidylcholines, PC-O: Ether phosphatidylcholines, PE: Phosphatidylethanolamines, PE-O: Ether phosphatidylethanolamines, PI: Phosphatidylinositols, Cer: Ceramides, and SM: Sphingomyelins

Fig 6.. Prognostic circulating lipid candidates.

Volcano plots comparing DKO lifetime stage I with (a) DKO lifetime stage II, (b) DKO lifetime stage III, (c) DKO lifetime stage IV and (d) DKO lifetime stage V. P- values for volcano plot analysis were calculated using Welch’s T-test. (e) Upset plot showing the intersection of the various groups of significant lipids selected from volcano plots. Lipids present in at least three sets were colored brown. Kaplan-Meier survival curves for (f) PC(39:4), (g) PC(37:2) and (h) PC(40:7). P-values were computed with the Log rank test. (i) Selected prognostic circulating lipids. PC: Phosphatidylcholines FC: Fold changes. ΔRMST: differences in restricted mean survival times.

Fig 7.. Schematic of metabolic pathways showing key metabolic alterations in the DKO mice lipidome.

Lipid classes detected in the study are indicated as bolded blue text, while unbolded blue text signifies other metabolites in the metabolic pathway. Red text indicates enzymes known to be overly expressed in ovarian cancer cells or other related cancer, with the relevant references. For each detected lipid class presented, information about the cluster they belong to in the temporal trend analyses is provided, in addition to the breakdown information on discriminant lipids selected by ML algorithms. A red circle represents the cumulative change in detected lipid classes (increase in DKO mice), a green circle (decrease in DKO mice), or a white circle (no cumulative change). Cumulative changes are computed by counting the number of both increased and decreased levels among the selected discriminant lipid in all lifetime stages. Pathway information was derived from existing literature. Abbreviations: G3P: Glycerol-3-phosphate, PA: Phosphatidic acid, DG: Diacylglycerols, TG: Triacylglycerols, PC: Phosphatidylcholines, PC O-: Ether phosphatidylcholines, PE: Phosphatidylethanolamines, PE O-: Ether phosphatidylethanolamines, LPE: Lysophosphatidylethanolamines, LPC: Lysophosphatidylcholines, PI: Phosphatidyl inositol, HMG CoA: 3-hydroxy-3-methylglutaryl coenzyme A, MUFA: mono-unsaturated fatty acids, PUFA: Poly-unsaturated fatty acids, SM: Sphingomyelin, Cer: Ceramide, HexCer: Hexosylceramide, CK: Choline kinase, ACC: acetyl-CoA carboxylase, ACL: ATP-citrate lyase, FAS: Fatty acid synthase, SCD1: Stearoyl-CoA desaturase-1, UGCG: uridine diphosphate-glucose ceramide glucosyltransferase.