Variable deep learning training horizons reveal the temporal complexity of biological systems.

Abstract

The increasing quantity of time-series images presents new opportunities for extracting biological insights from data. Here, we introduce a deep learning framework with a variable input sequence length to predict cell and colony morphologies. We apply this framework to in silico and in vitro microscopy datasets, evaluating the impact of temporal data on performance. We find that while performance increases monotonically with increasing in silico training data, performance is varied in the in vitro case studies. The varying results reflect the intrinsic challenges stochastic, complex biological systems pose to data-driven modeling, and offer a new method through which we can identify biological transition points using temporal dynamics.