One potential limitation of our simulation framework was the use of of a small N approximation (simulating tumour populations with a final size of 103) to improve computational speed and efficiency. Therefore, to examine the generalizability of our deep learning framework in larger population sizes, we collected a dataset of 900 synthetic tumours simulated using an orthogonal framework called TEMULATOR. We evaluated our method using an alternative framework for two reasons. Firstly, by using an orthogonal framework, we avoid potential problems in cases where our overparametrized deep learning models learned to separate classes by learning the underlying noise distributions. Secondly, the dataset generated by Caravagna et al. 2020 simulated tumours to final population sizes of >108, which provides a strong benchmark for evaluating the small N approximation. The dataset consists of 900 tumours across variable sequencing depths. All tumours were simulated with 1 subclone that is either at detectable frequencies (~10 - 40% VAF) or non-detectable, effectively neutral frequencies (subclones less than ~10% VAF or subclones greater than 40% VAF). (Simulator TEMULATOR; Simulations from Caravagna et al. 2020).