A comprehensive fitness landscape model reveals the evolutionary history and future evolvability of eukaryoticcis-regulatory DNA sequences

Abstract

Mutations in non-coding cis-regulatory DNA sequences can alter gene expression, organismal phenotype, and fitness. Fitness landscapes, which map DNA sequence to organismal fitness, are a long-standing goal in biology, but have remained elusive because it is challenging to generalize accurately to the vast space of possible sequences using models built on measurements from a limited number of endogenous regulatory sequences. Here, we construct a sequence-to-expression model for such a landscape and use it to decipher principles of cis-regulatory evolution. Using tens of millions of randomly sampled promoter DNA sequences and their measured expression levels in the yeast Sacccharomyces cerevisiae, we construct a deep transformer neural network model that generalizes with exceptional accuracy, and enables sequence design for gene expression engineering. Using our model, we predict and experimentally validate expression divergence under random genetic drift and strong selection weak mutation regimes, show that conflicting expression objectives in different environments constrain expression adaptation, and find that stabilizing selection on gene expression leads to the moderation of regulatory complexity. We present an approach for detecting selective constraint on gene expression using our model and natural sequence variation, and validate it using observed cis-regulatory diversity across 1,011 yeast strains, cross-species RNA-seq from three different clades, and measured expression-to-fitness curves. Finally, we develop a characterization of regulatory evolvability, use it to visualize fitness landscapes in two dimensions, discover evolvability archetypes, quantify the mutational robustness of individual sequences and highlight the mutational robustness of extant natural regulatory sequence populations. Our work provides a general framework that addresses key questions in the evolution of cis-regulatory sequences.

Discovering molecular features of intrinsically disordered regions by using evolution for contrastive learning

“…a pre-specified loss function) (LeCun, Bengio and Hinton, 2015). Current approaches to infer sequence function using neural networks employ regression tasks, where models learn to predict expert annotations or largescale measurements (Alipanahi et al, 2015;Avsec et al, 2021;Dhaval Vaishnav et al, 2021). For example, training genomic sequence models on labels representing the presence or absence of transcription factor binding leads to the model learning features that directly correspond with the consensus motifs for these transcription factors (Koo and Eddy, 2019).…”

Reference-based cell type matching of spatial transcriptomics data

“…The segmentation step produced the cell-by-gene matrices that were used to assign the spatially-resolved cell types to scRNA-seq reference cell types using cell type matching algorithms. Teams of the SpaceTx Consortium explored six computational algorithms (map.cells* [1], mfishtools [24], fitness landscape model (FLM) [25], FR-Match [26, 27], Tangram [28], and pciSeq [29]), which produced individual cell type assignments with various probabilistic assignment scores. To arrive at consensus cell type assignments, two meta-analysis strategies were developed to combine the individual assignments more qualitatively (Negative Weighting Combining Strategy, hereinafter NWCS), or more quantitatively (Geometric Mean Combining Strategy, hereinafter GMCS) ( Methods ).…”

“…Six cell type matching algorithms (map.cells* [1], mfishtools [24], fitness landscape model (FLM) [25], FR-Match [26, 27], Tangram [28], and pciSeq [29]) were applied to assign reference scRNA-seq cell types to each segmented cell with an associated confidence score (a.k.a. probabilistic assignment) based on the cell-by-gene count matrix ( Method ) produced for the MERFISH data.…”

“…We evaluated six computational cell type matching algorithms, namely map.cells* [1], mfishtools [24], fitness landscape model (FLM) [25], FR-Match [26, 27], Tangram [28], and pciSeq [29].…”

Genomics enters the deep learning era

“… Vaishnav et al (2021) trained a deep transformer network to predict the gene expression level associated with 20 million randomly sampled 80-bp long DNA sequences introduced in a Saccharomyces cervisiae promoter region. They assessed the effect of all single mutations in promoter regions and discovered four evolvability archetypes: robust promoters on which mutations have little effect, plastic promoters on which every mutations have a small effect and minimal or maximal promoters on which only some mutations can dramatically decrease or increase the associated expression level.…”