Regularization Improves the Robustness of Learned Sequence-to-Expression Models
Understanding of the gene regulatory activity of enhancers is a major problem in regulatory biology. The nascent field of sequence-to-expression modelling seeks to create quantitative models of gene expression based on regulatory DNA (cis) and cellular environmental (trans) contexts. All quantitative models are defined partially by numerical parameters, and it is common to fit these parameters to data provided by existing experimental results. However, the relative paucity of experimental data appropriate for this task, and lacunae in our knowledge of all components of the systems, results in problems often being under-specified, which in turn may lead to a situation where wildly different model parameterizations perform similarly well on training data. It may also lead to models being fit to the idiosyncrasies of the training data, without representing the more general process (overfitting).\n\nIn other contexts where parameter-fitting is performed, it is common to apply regularization to reduce overfitting. We systematically evaluated the efficacy of three types of regularization in improving the generalizability of trained sequence-to-expression models. The evaluation was performed in two types of cross-validation experiments: one training on D. melanogaster data and predicting on orthologous enhancers from related species, and the other cross-validating between four D. melanogaster neurogenic ectoderm enhancers, which are thought to be under control of the same transcription factors. We show that training with a combination of noise-injection, L1, and L2 regularization can drastically reduce overfitting and improve the generalizability of learned sequence-to-expression models. These results suggest that it may be possible to mitigate the tendency of sequence-to-expression models to overfit available data, thus improving predictive power and potentially resulting in models that provide better insight into underlying biological processes.