Model-based analysis of polymorphisms in an enhancer reveals cis-regulatory mechanisms

Abstract

It is challenging to predict the impact of small genetic changes such as single nucleotide polymorphisms on gene expression, since mechanisms involved in gene regulation and their cis-regulatory encoding are not well-understood. Recent studies have attempted to predict the functional impact of non-coding variants based on available knowledge of cis-regulatory encoding, e.g., transcription factor (TF) motifs. In this work, we explore the relationship between regulatory variants and cis-regulatory encoding from the opposite angle, using the former to inform the latter. We employ sequence-to-expression modeling to resolve ambiguities regarding gene regulatory mechanisms using information about effects of single nucleotide variations in an enhancer. We demonstrate our methodology using a well-studied enhancer of the developmental gene intermediate neuroblasts defective (ind) in D. melanogaster. We first trained the thermodynamics-based model GEMSTAT to relate the neuroectodermal expression pattern of ind to its enhancer's sequence, and constructed an ensemble of models that represent different parameter settings consistent with available data for this gene. We then predicted the effects of every possible single nucleotide variation within this enhancer, and compared these to SNP data recorded in the Drosophila Genome Reference Panel. We chose specific SNPs for which different models in the ensemble made conflicting predictions, and tested their effect in vivo. These experiments narrowed in on one mechanistic model as capable of explaining the observed effects. We further confirmed the generalizability of this model to orthologous enhancers and other related developmental enhancers. In conclusion, mechanistic models of cis-regulatory function not only help make specific predictions of variant impact, they may also be learned more accurately using data on variants.A central issue in analyzing variations in the non-coding genome is to interpret their functional impact, and their connections to phenotype differences and disease etiology. Machine learning methods based on statistical modeling have been developed to associate genetic variants to expression changes. However, associations predicted by these models may not be functionally relevant, despite being statisticaly significant. We describe how mathematical modeling of gene expression can be employed to systematically study the non-coding sequence and its relationship to gene expression. We demonstrate our method in a well studied developmental enhancer of the fruitfly. We establish the efficacy of mathematical models in combination with the polymorphism data to reveal new mechanistic insights.

Deciphering enhancer sequence using thermodynamics-based models and convolutional neural networks

“…Although the mechanisms that we find are specific to the regulation of rhomboid and other neuroectodermal enhancers, and might not necessarily generalize to the regulation of other genes, our work gives a recipe for understanding the regulatory mechanisms in a data-driven or model-driven manner. In addition to their explanatory role, the models tested here can also be useful for predicting the expression driven by unseen sequences and cellular contexts; this ability has several applications in down-stream analysis such as predicting the effect of a particular TF’s knockout, site mutagenesis, or effects of single nucleotide polymorphisms on the expression ( 68 , 69 ). We also showed that convolutional neural networks can be a reliable expression prediction tool capable of learning non-linear regulatory mechanisms from modest-sized training data.…”

Deciphering enhancer sequence using thermodynamics-based models and convolutional neural networks

“…Although the mechanisms that we find are specific to the regulation of rhomboid and other neuroectodermal enhancers, and might not necessarily generalize to the regulation of other genes, our work gives a recipe for understanding the regulatory mechanisms in a data-driven or model-driven manner. In addition to their explanatory role, the models tested here can also be useful for predicting the expression driven by unseen sequences and cellular contexts; this ability has several applications in down-stream analysis such as predicting the effect of a particular TF’s knockout, site mutagenesis, or effects of single nucleotide polymorphisms on the expression (64, 65). We also showed that convolutional neural networks can be a reliable expression prediction tool capable of learning non-linear regulatory mechanisms from modest-sized training data.…”