A diagnostic model based on bioinformatics and machine learning to differentiate bipolar disorder from schizophrenia and major depressive disorder

Abstract

Bipolar disorder (BD) showed the highest suicide rate of all psychiatric disorders, and its underlying causative genes and effective treatments remain unclear. During diagnosis, BD is often confused with schizophrenia (SC) and major depressive disorder (MDD), due to which patients may receive inadequate or inappropriate treatment, which is detrimental to their prognosis. This study aims to establish a diagnostic model to distinguish BD from SC and MDD in multiple public datasets through bioinformatics and machine learning and to provide new ideas for diagnosing BD in the future. Three brain tissue datasets containing BD, SC, and MDD were chosen from the Gene Expression Omnibus database (GEO), and two peripheral blood datasets were selected for validation. Linear Models for Microarray Data (Limma) analysis was carried out to identify differentially expressed genes (DEGs). Functional enrichment analysis and machine learning were utilized to identify. Least absolute shrinkage and selection operator (LASSO) regression was employed for identifying candidate immune-associated central genes, constructing protein-protein interaction networks (PPI), building artificial neural networks (ANN) for validation, and plotting receiver operating characteristic curve (ROC curve) for differentiating BD from SC and MDD and creating immune cell infiltration to study immune cell dysregulation in the three diseases. RBM10 was obtained as a candidate gene to distinguish BD from SC. Five candidate genes (LYPD1, HMBS, HEBP2, SETD3, and ECM2) were obtained to distinguish BD from MDD. The validation was performed by ANN, and ROC curves were plotted for diagnostic value assessment. The outcomes exhibited the prediction model to have a promising diagnostic value. In the immune infiltration analysis, Naive B, Resting NK, and Activated Mast Cells were found to be substantially different between BD and SC. Naive B and Memory B cells were prominently variant between BD and MDD. In this study, RBM10 was found as a candidate gene to distinguish BD from SC; LYPD1, HMBS, HEBP2, SETD3, and ECM2 serve as five candidate genes to distinguish BD from MDD. The results obtained from the ANN network showed that these candidate genes could perfectly distinguish BD from SC and MDD (76.923% and 81.538%, respectively).

Fig. 1

Flow chart.

Fig. 2. DEGs between BD and SC.

A volcano plots of DEGs in dataset GSE92538-GPL10526; B volcano plots of DEGs in dataset GSE92538-GPL17027; C volcano plots of DEGs in dataset GSE12654; D, E UMAP plots before and after removal of batch effects; F combined volcano plots of DEGs in the dataset.

Fig. 3. DEGs between BD and MDD.

A volcano plot of DEGs in dataset GSE92538-GPL10526; B volcano plot of DEGs in dataset GSE92538-GPL17027; C volcano plot of DEGs in the combined dataset.

Fig. 4. Functional enrichment analysis of relevant candidate genes distinguishing BD from SC.

A Venn diagram of DEGs between bipolar disorder and schizophrenia in a single dataset versus DEGs in a combined dataset; B KEGG analysis of candidate genes; GO analysis of candidate genes for C cellular components (CC); D biological processes (BP); E molecular function (MF).

Fig. 5. Functional enrichment analysis of relevant candidate genes distinguishing BD from MDD.

A Venn diagram of DEGs between bipolar disorder and schizophrenia in the single dataset versus DEGs in the combined dataset; B KEGG analysis of candidate genes; GO analysis of candidate genes for C cellular components (CC); D biological processes (BP); E molecular function (MF).

Fig. 6. Candidate gene identification to distinguish BD from SC.

A–D LASSO regression candidate gene identification (GSE92538-GPL10526, GSE92538-GPL17027, GSE12654, and combined datasets, respectively); E LASSO regression candidate gene Venn diagram; F PPI network construction of candidate genes.

Fig. 7. Candidate gene identification to distinguish BD from MDD.

A–C LASSO regression candidate gene screening (GSE92538-GPL10526, GSE92538-GPL17027, and combined dataset, respectively); D LASSO regression candidate gene Venn diagram; E PPI network construction of candidate genes.

Fig. 8. Validation of the diagnostic value of candidate genes for BD and SC.

A–E ROC curves for different datasets (GSE92538-GPL10526, GSE92538-GPL17027, GSE12654, combined dataset, and GSE18312, respectively); F, G ANN validation of candidate genes.

Fig. 9. Validation of the diagnostic value of candidate genes for BD and MDD.

A–D ROC curves for different datasets (GSE92538-GPL10526, GSE92538-GPL17027, combined dataset, and GSE39653, respectively); E, F ANN validation of candidate genes.

Fig. 10. The expression profile analysis of candidate genes.

A–D differential expression profiling of candidate genes distinguishing BD and SC (GSE92538-GPL10526, GSE92538-GPL17027, GSE12654, and combined datasets, respectively); E–G differential expression profiling of candidate genes distinguishing BD and MDD (GSE92538- GPL10526, GSE92538-GPL17027, and the combined dataset).

Fig. 11. CIBERSORT analysis of 22 immune cells.

A relative percentage of 22 immune cells in each sample; B correlation among 22 immune cells; C difference in immune infiltration between BD and SC; D difference in immune infiltration between BD and MDD; E correlation among 6 genes and 4 immune cells.