Predictive modeling of big-omics data using deep-learning techniques is gaining momentum due to its role in personalized treatment. Since modeling such data is both compute and data-intensive it is nearly impossible to achieve the task without accelerators. One of the affordable toolboxes that aids in modelling of the data in health-care is Google’s Colaboratory. It is a cloud-based service that provides access to accelerators like GPUs and TPUs. In this work, we compared the performance of the accelerators hosted by Colaboratory to assess their advantages and shortcomings in the purview of clinical-science. The two models that we chose to train in this study are: a drug-response model and a cancer-subtype classifier. These predictive models help in the prognosis and diagnosis of carcinomas. We compared the performance of accelerators in terms of their training speed. We also compared the accuracy of the proposed models with similar methods. Results show that the models can be trained faster on the GPUs with smaller batches while TPUs achieve 5x-speedup for larger batches. The merits and pitfalls of the Colaboratory are also discussed at the end which may help the potential users. We conclude that performance of the accelerators with free/affordable options is sufficient for small research groups while they are insufficient for demanding problems that require ample processing power and memory.

Deep Learning, Multi-omics data, GPU, TPU, AutoEncoder, Convolution Neural Network