Local DNA shape is a general principle of transcription factor binding specificity inArabidopsis thaliana

Abstract

A genome encodes two types of information, the "what can be made" and the "when and where". The "what" are mostly proteins which perform the majority of functions within living organisms and the "when and where" is the regulatory information that encodes when and where proteins are made. Currently, it is possible to efficiently predict the majority of the protein content of a genome but nearly impossible to predict the transcriptional regulation. This regulation is based upon the interaction between transcription factors and genomic sequences at the site of binding motifs1,2,3. Information contained within the motif is necessary to predict transcription factor binding, however, it is not sufficient4. Peaks detected in amplified DNA affinity purification sequencing (ampDAP-seq) and the motifs derived from them only partially overlap in the genome3 indicating that the sequence holds information beyond the binding motif. Here we show a random forest machine learning approach which incorporates the 3D-shape improved the area under the precision recall curve for binding prediction for all 216 tested Arabidopsis thaliana transcription factors. The method resolved differential binding of transcription factor family members which share the same binding motif. The models correctly predicted the binding behavior of novel, not-in-genome motif sequences. Understanding transcription factor binding as a combination of motif sequence and motif shape brings us closer to predicting gene expression from promoter sequence.

Toward learning the principles of plant gene regulation

“…Using shallow ML models, it was recently shown in multiple eukaryotic organisms, including Arabidopsis, that most TFs likely combine both types of readout to recognize their binding sites, as the integration of these features improves binding site predictions [12]. In line with this, local DNA structural properties can be more highly conserved than nucleotide sequence and were similarly found to contribute to TF cooperativity, a mechanism where TFs bind DNA cooperatively and thereby strengthen their affinity [5].…”

“…The specificities of protein-DNA interactions are defined both by direct protein-DNA readout, facilitated by the major groove of the DNA helix, and indirect readout, mediated through DNA backbone and minor groove contacts. The latter comprises 'weak' protein-DNA interactions, with base pairs that are not directly contacted by the protein but defined by DNA conformational and physicochemical (shape) properties at or around the specific binding sites [5,12].…”

“…Since DNNs automatically learn predictive motif representations [4], cooperative binding interactions [2,5] and genotypic variation effects [3], they represent a powerful approach to uncover the detailed cis-regulatory grammar of genomic sequences (see Figure 1C in main text). Shallow ML approaches remain a highly useful alternative for assessing the performance of numerically encoded DNA sequence features, such as DNA structural (shape) properties [12]. With deep models: interpret the internal hidden data representations x', x".…”

Synergism and antagonism in plant acclimation to abiotic stress combinations

“…The accessibility of the DNA, the presence of motifs and the matching activated transcription factors determine the activity of genes (Beer and Tavazoie 2004). Machine learning for improved pattern recognition based on large transcriptome data sets allowed for defining the local DNA landscape determining the binding sites of transcription factors (Sielemann et al 2021). In addition, DNAbinding proteins often assemble in combinatorial manner as homo-or heteromers, activate or repress transcription and are focal points of supramolecular assemblies with properties of signal integration modules as discussed above.…”

Identifying transcription factor-DNA interactions using machine learning

“…A major limitation is that this approach only finds sequences with a pattern match (Weirauch, et al, 2014). This leads to a high number of false positives, which misleads the characterization of true TF bound regions (Sielemann, et al, 2021). Moreover, motif based method searches can miss functional TF-DNA interactions for a number of reasons; 1) Some TF binding sequences do not have motifs (Inukai, et al, 2017), 2) Some TF binding sequences have multiple binding motifs (Nakagawa, et al, 2013), and 3) Recognition by TFs is also dependent upon sequences surrounding the motifs, so it is not enough to only identify the motifs alone (Inukai, et al, 2017).…”