Identification of conserved transcriptome features between humans and Drosophila in the aging brain utilizing machine learning on combined data from the NIH Sequence Read Archive

Abstract

Aging is universal, yet characterizing the molecular changes that occur in aging which lead to an increased risk for neurological disease remains a challenging problem. Aging affects the prefrontal cortex (PFC), which governs executive function, learning, and memory. Previous sequencing studies have demonstrated that aging alters gene expression in the PFC, however the extent to which these changes are conserved across species and are meaningful in neurodegeneration is unknown. Identifying conserved, age-related genetic and morphological changes in the brain allows application of the wealth of tools available to study underlying mechanisms in model organisms such as Drosophila melanogaster. RNA sequencing data from human PFC and fly heads were analyzed to determine conserved transcriptome signatures of age. Our analysis revealed that expression of 50 conserved genes can accurately determine age in Drosophila (R2 = 0.85) and humans (R2 = 0.46). These transcriptome signatures were also able to classify Drosophila into three age groups with a mean accuracy of 88% and classify human samples with a mean accuracy of 69%. Overall, this work identifies 50 highly conserved aging-associated genetic changes in the brain that can be further studied in model organisms and demonstrates a novel approach to uncovering genetic changes conserved across species from multi-study public databases.

Fig 1. Workflow of biological age prediction using XGBoost across species.

Depicting selection of genes for both aging involvement and conservation between Human and Drosophila.

Fig 2. Biological age prediction using XGBoost across species.

A) Sample regression analysis using all human genes to predict human age. B) Histogram of R² values for predicting human age with all available human genes with a mean R² of 0.61. C) A sample regression analysis with all Drosophila genes to predict Drosophila age. D) Histogram of R² values for predicting Drosophila age using all available Drosophila genes with a mean R² of 0.93. All results calculated using XGBoost. Histograms represent averages across 1000 bootstrapped random samplings where the regressor or classifier was trained on 75% of the samples and tested on 25% with values reported as the mean across 1000 iterations. Median AE stands for Median Average Error.

Fig 3. Feature selection using aging correlated genes across species.

A) Sample regression analysis using the top 1000 human aging correlated genes to predict human age. B) Histogram of R² values using the top 1000 human aging correlated genes to predict human age with a mean R² of 0.62. C) Sample regression analysis predicting Drosophila age using the top 1000 aging correlated Drosophila genes. D) Histogram of R² values predicting Drosophila age using the top 1000 Drosophila aging correlated genes with mean R² of 0.95. All results calculated using XGBoost. All histograms represent averages across 1000 bootstrapped random samplings where the regressor or classifier was trained on 75% of the samples and tested on 25%. All histogram values are reported as the mean across 1000 iterations. Median AE stands for median average error.

Fig 4. Intersection of homologous aging-correlated genes.

A) Sample regression analysis of the overlapping 50 aging correlated genes across humans. B) Histogram of R² values of human age prediction using the conserved 50 aging correlated genes with a mean R² of 0.46. C) Sample regression analysis of the mean R² values predicting Drosophila age using the 50 aging correlated genes. D) Histogram of R² values predicting Drosophila age using the conserved 50 aging correlated genes with a mean R² of 0.85. All results calculated using XGBoost. All histogram results represent averages across 1000 bootstrapped random samplings where the regressor or classifier was trained on 75% of the samples and tested on 25%. All histogram values are reported as the mean across 1000 iterations. Median AE stands for Median Average Error.

Fig 5

A) A heat map of gene expression in young, middle aged, and old human prefrontal cortex of the 50 conserved genes. Shift of color from blue to red indicates an increase in expression relative to all gene expression in the dataset as indicated B) Interactome representing STRING Network interactions of 50 conserved genes. Networks connected through ‘edge’ lines represented by STRING confidence score. Line thickness indicates interaction confidence scores with greater thickness indicating higher confidence interaction. The table on the right indicates KEGG pathways represented by the STRING network. Colors in nodes represent involvement in pathway, see table. Orange border surrounding a node indicates occurrence within 1000 genes with a reported association of either Parkinson’s or Alzheimer’s disease.