Prediction of fine-tuned promoter activity from DNA sequence

The quantitative prediction of transcriptional activity of genes using promoter sequence is fundamental to the engineering of biological systems for industrial purposes and understanding the natural variation in gene expression. To catalyze the development of new algorithms for this purpose, the Dialogue on Reverse Engineering Assessment and Methods (DREAM) organized a community challenge seeking predictive models of promoter activity given normalized promoter activity data for 90 ribosomal protein promoters driving expression of a fluorescent reporter gene. By developing an unbiased modeling approach that performs an iterative search for predictive DNA sequence features using the frequencies of various k-mers, inferred DNA mechanical properties and spatial positions of promoter sequences, we achieved the best performer status in this challenge. The specific predictive features used in the model included the frequency of the nucleotide G, the length of polymeric tracts of T and TA, the frequencies of 6 distinct trinucleotides and 12 tetranucleotides, and the predicted protein deformability of the DNA sequence. Our method accurately predicted the activity of 20 natural variants of ribosomal protein promoters (Spearman correlation r = 0.73) as compared to 33 laboratory-mutated variants of the promoters (r = 0.57) in a test set that was hidden from participants. Notably, our model differed substantially from the rest in 2 main ways: i) it did not explicitly utilize transcription factor binding information implying that subtle DNA sequence features are highly associated with gene expression, and ii) it was entirely based on features extracted exclusively from the 100 bp region upstream from the translational start site demonstrating that this region encodes much of the overall promoter activity. The findings from this study have important implications for the engineering of predictable gene expression systems and the evolution of gene expression in naturally occurring biological systems. Author Summary Gene expression is the first step at which information encoded in DNA is transcribed into RNA. Predicting gene expression from DNA sequence can provide insights into the natural variation of gene expression underlying various phenotypes and direct the engineering of genes of desired activity, for example in industrial processes. While several studies show that gene expression is influenced by DNA sequence. its quantitative prediction from DNA sequence alone remains a challenging problem. Unfortunately, studies aimed at developing quantitative models for gene expression prediction are not directly comparable because most have used distinct data sets for training and evaluation. and many of the methods have not been independently verified. Open innovation challenges in which a problem is posed to a wide community provide a framework for independent verification of the performance of various computational methods using the same benchmark data sets and statistical procedures. Here. we describe the best performing computational model amongst those of 20 other teams in the DREAM6 Gene Expression Prediction challenge. We show that a highly predictive gene expression model can be obtained by an unbiased. data-driven approach that makes little assumption on the role of known mechanisms for gene regulation.