Use of deep neural network ensembles to identify embryonic-fetal transition markers: repression of COX7A1 in embryonic and cancer cells

Abstract

Here we present the application of deep neural network (DNN) ensembles trained on transcriptomic data to identify the novel markers associated with the mammalian embryonic-fetal transition (EFT). Molecular markers of this process could provide important insights into regulatory mechanisms of normal development, epimorphic tissue regeneration and cancer. Subsequent analysis of the most significant genes behind the DNNs classifier on an independent dataset of adult-derived and human embryonic stem cell (hESC)-derived progenitor cell lines led to the identification of COX7A1 gene as a potential EFT marker. COX7A1, encoding a cytochrome C oxidase subunit, was up-regulated in post-EFT murine and human cells including adult stem cells, but was not expressed in pre-EFT pluripotent embryonic stem cells or their in vitro-derived progeny. COX7A1 expression level was observed to be undetectable or low in multiple sarcoma and carcinoma cell lines as compared to normal controls. The knockout of the gene in mice led to a marked glycolytic shift reminiscent of the Warburg effect that occurs in cancer cells. The DNN approach facilitated the elucidation of a potentially new biomarker of cancer and pre-EFT cells, the embryo-onco phenotype, which may potentially be used as a target for controlling the embryonic-fetal transition.

Keywords: Warburg effect; cancer marker; deep neural network; embryonic-fetal transition; stem cells.

Figure 1. Predicting embryonic state through DNN ensemble

(A) Validation confusion matrix performance for DNN ensemble trained on Illumina data. (B) Validation confusion matrix performance for DNN ensemble trained on Affymetrix data. (C) Embryonic scores obtained through Affymetrix DNN ensembles for GEO next generation data set GSE62193 consisting of samples representing different stages of human photoreceptor development from ES cells.

Figure 2. Expression analysis of COX7A1, Lin28b and Rps10 transcripts in mouse development measured by RNA-seq

(A, B) Analysis of expression of key embryonic-fetal makers had been conducted in mouse to demonstrate gradient upregulation of COX7A1 along with gradient downregulation of Lin28b during mouse embryonal development as measured by NGS, where FPKM is relative RNA expression units and DPC (days post coitum) reflects embryonic stage. (C) Rps10 expression was used to ensure equal amount of RNA was used across all samples.

Figure 3. Expression analysis of COX7A1, LIN28B and RPS10 transcripts in human tissues at different stages of development along with methylation analysis of COX7A1, LIN28B and RPS10 genes in human cell lines

(A, C) Dermal fibroblasts of the upper arm from developmental stages spanning the onset of fetal development (eight weeks of gestation) through adulthood were synchronized in quiescence in vitro and RNA subjected to analysis on Illumina gene expression bead arrays. COX7A1 had been upregulated in adult stages while LIN28B displayed the opposite pattern. It should be noted that in iPSCs generated from matching adult tissues the level of expression of these genes demonstrated the reverse pattern compared to adult tissues. (B, D) Four human cell lines were used for methylation analysis by bisulfite sequencing. In two embryonic derived cell lines, 4D20.8 and 30-MV2-6, genomic DNA appears to be methylated at COX7A1 region, while in two adult derived cell lines where COX7A1 expression had been detected its genomic region appears to be relatively unmethylated. LIN28B methylation pattern seems to be unchanged in embryonic and adult derived cell lines. Blue bars represent levels of methylation, one bar for every methylated C. The height of the bar corresponds to the fraction of reads covering that C that are methylated (the highest bars = 1–meaning Cs in all reads are methylated). (E, F) RPS10 was used as a housekeeping control for methylation and expression analysis.

Figure 4. Expression analysis of COX7A1 and RPS10 transcripts in cancer, embryonic and adult cell lines along methylation analysis of COX7A1 and RPS10 genes in cancer and healthy samples

(A) Embryonic progenitors capable of osteochondral differentiation such as the line 4D20.8 showed no evidence of COX7A1 expression either in the progenitor state or in the differentiated state despite expressing high levels of osteochondral markers. Similarly, embryonic progenitor adipocytes E3 and myoblasts SK5 did not express COX7A1. In contrast, adult-derived MSCs expressed COX7A1 before and after differentiation. The same situation was observed with adult-derived preadipocytes and myoblasts. When expression levels of COX7A1 were measured in osteosarcomas, liposarcomas and rhabdomyosarcoma all lines except one showed an embryonic pattern of COX7A1 expression. (D) Several cancer cell lines demonstrated decreased level of COX7A1 expression compared to healthy tissue controls; ESCs and adult MSCs were used as internal controls for COX7A1 expression. (C, F, G) Methylation analysis of cancer samples obtained from lung, liver and oral carcinomas demonstrated statistically significant increase of methylation of COX7A1 compare to healthy controls. (B, C, E–G) RPS10 gene used as a housekeeping control for methylation and expression analysis.

Figure 5. Comparative analysis of COX7A1 expression in sarcoma cell lines vs. normal mesenchymal cell lines

Analysis by t-test demonstrated statistically significant decrease of COX7A1 expression in cancer cell lines comparing to matching controls. Normalized expression or relative expression values of COX7A1 were calculated using transcriptomic data from (A) Sarcoma Project–67 samples, (B) BioTime internal data–103 samples, (C) Fantom5 Project–71 samples, (D) GEO pooled sarcoma and mesenchymal cell lines–over 1000 samples.

Figure 6. Warburg effect in cells with COX7A1 deletion

(A, B) The figure shows that a glycolytic shift, normally observed in cancer cell lines (CRL3042, CRL3044) and E3 (hESC derived) progenitor cell line, but not adult primary preadipocyte cells, (C, D) Glycolytic shift is also observed in cells, lacking COX7A1 gene: heart cells derived from COX7A1 –/− mouse; heart cells from a COX7A1 +/+ littermate mouse, age 2 months. Just as in cancer and E3 progenitor cells, glycolysis level was higher in cells with COX7A1 deletion.