Fig. 1: Deep learning models can be used to discriminate amongst the evolutionary pressures shaping blood evolution.

a The impact of selection and genetic drift on shaping clonal dynamics. Cells accumulate somatic mutations in each division. The majority of mutations will be either neutral (blue) or mildly damaging (red). Driver mutations will increase the fitness of a cell and increase the frequency in the population (green). However, mutations are also able to rise in frequency through genetic drift. b Mutation summary statistics extracted from blood cell populations. Summary statistics fall into three categories: (1) Counts of mutations in each blood sample (overall and stratified according to mutation type (silent and missense) across variant allele frequency intervals, (2) The frequency of mutations (variant allele frequency), and (3) mutation annotation and respective ratios (proportion of missense relative to total missense sites over the proportion of silent mutation relative to total silent sites). A total of 16 summary statistics are extracted from each population. c Deep Neural Network Architecture. Each DNN was trained as a multi-task neural network and classifies a population into one of four overarching evolutionary classes and predicts four continuous parameters. Each neural network consisted of an input layer (16 units with each unit corresponding to a summary statistic), three hidden layers (512 units), and five output layers which included the classification output (four units) and four regression outputs (one unit each). d DNN Ensemble. We trained a total of ten deep neural networks (DNNs) independently, yet with identical architecture. Through employing an ensemble-based approach, we are able to obtain a distribution of predictions for each population. e Classification performance for simulated evolutionary classes. The y-axis represents the true evolutionary class, and the x-axis represents the predicted evolutionary class. Classification accuracy ranges from blue (low accuracy) to red (high accuracy). We obtain a high classification accuracy across evolutionary classes (94.8%). Positive and combination classes are predicted with 99.7% and 97.4%, respectively. We observe a reduction in accuracy in neutral (80.6%) and negative (83.4%) classes of evolution.