Motivation: Wide adoption of Massively Parallel Reporter Assays, MPRAs, provides a rich ground for the application of deep learning to model gene regulatory regions and predict their activity from a DNA sequence. Yet, the development of new and improvement of existing models is hindered by the lack of a conveniently standardized dataset encompassing the data from multiple species and alternative MPRA designs or a repertoire of uniformly tested baseline models.
Results: Here we present MPRA-MNIST, a curated benchmarking dataset collection, encompassing results of 14 MPRA studies from human, yeast, fly, and bacteria. The underlying data were uniformly processed and converted to a standardized format with predefined and data-leakage–free ""training-validation-test"" splits, enabling rigorous and reproducible comparison of modelling approaches.
Each dataset is equipped with a vignette illustrating the basic model training from scratch to establish the baseline performance. Technically, MPRA-MNIST is a lightweight Python package providing dataset loaders, sequence transformations, and common deep learning utilities, while its modular structure allows for including extra user-defined datasets and models. We illustrate the practical usage of the framework with comparative benchmarking of several deep learning architectures across all datasets, which also provides standardized evaluation baselines. All in all, MPRA-MNIST establishes a convenient toolbox and educational resource for further development of reproducible MPRA-based deep learning models in regulatory genomics.

Co-authors: Nikita Penzin, Eva Zubova, Arsenii Zinkevich, Ivan Kulakovsky