Evaluation and optimization of sequence-based gene regulatory deep learning models

Proformer: a hybrid macaron transformer model predicts expression values from promoter sequences

The breakthrough high-throughput measurement of the cis-regulatory activity of millions of randomly generated promoters provides an unprecedented opportunity to systematically decode the cis-regulatory logic that determines the expression values. We developed an end-to-end transformer encoder architecture named Proformer to predict the expression values from DNA sequences. Proformer used a Macaron-like Transformer encoder architecture, where two half-step feed forward (FFN) layers were placed at the beginning and the end of each encoder block, and a separable 1D convolution layer was inserted after the first FFN layer and in front of the multi-head attention layer. The sliding k-mers from one-hot encoded sequences were mapped onto a continuous embedding, combined with the learned positional embedding and strand embedding (forward strand vs. reverse complemented strand) as the sequence input. Moreover, Proformer introduced multiple expression heads with mask filling to prevent the transformer models from collapsing when training on relatively small amount of data. We empirically determined that this design had significantly better performance than the conventional design such as using the global pooling layer as the output layer for the regression task. These analyses support the notion that Proformer provides a novel method of learning and enhances our understanding of how cis-regulatory sequences determine the expression values.

Hold out the genome: a roadmap to solving the cis-regulatory code

Gene expression is regulated by transcription factors that work together to read cis-regulatory DNA sequences. The 'cis-regulatory code' - how cells interpret DNA sequences to determine when, where and how much genes should be expressed - has proven to be exceedingly complex. Recently, advances in the scale and resolution of functional genomics assays and machine learning have enabled substantial progress towards deciphering this code. However, the cis-regulatory code will probably never be solved if models are trained only on genomic sequences; regions of homology can easily lead to overestimation of predictive performance, and our genome is too short and has insufficient sequence diversity to learn all relevant parameters. Fortunately, randomly synthesized DNA sequences enable testing a far larger sequence space than exists in our genomes, and designed DNA sequences enable targeted queries to maximally improve the models. As the same biochemical principles are used to interpret DNA regardless of its source, models trained on these synthetic data can predict genomic activity, often better than genome-trained models. Here we provide an outlook on the field, and propose a roadmap towards solving the cis-regulatory code by a combination of machine learning and massively parallel assays using synthetic DNA.