De novodistillation of thermodynamic affinity from deep learning regulatory sequence models ofin vivoprotein-DNA binding

Abstract

Transcription factors (TF) are proteins that bind DNA in a sequence-specific manner to regulate gene transcription. Despite their unique intrinsic sequence preferences, in vivo genomic occupancy profiles of TFs differ across cellular contexts. Hence, deciphering the sequence determinants of TF binding, both intrinsic and context-specific, is essential to understand gene regulation and the impact of regulatory, non-coding genetic variation. Biophysical models trained on in vitro TF binding assays can estimate intrinsic affinity landscapes and predict occupancy based on TF concentration and affinity. However, these models cannot adequately explain context-specific, in vivo binding profiles. Conversely, deep learning models, trained on in vivo TF binding assays, effectively predict and explain genomic occupancy profiles as a function of complex regulatory sequence syntax, albeit without a clear biophysical interpretation. To reconcile these complementary models of in vitro and in vivo TF binding, we developed Affinity Distillation (AD), a method that extracts thermodynamic affinities de-novo from deep learning models of TF chromatin immunoprecipitation (ChIP) experiments by marginalizing away the influence of genomic sequence context. Applied to neural networks modeling diverse classes of yeast and mammalian TFs, AD predicts energetic impacts of sequence variation within and surrounding motifs on TF binding as measured by diverse in vitro assays with superior dynamic range and accuracy compared to motif-based methods. Furthermore, AD can accurately discern affinities of TF paralogs. Our results highlight thermodynamic affinity as a key determinant of in vivo binding, suggest that deep learning models of in vivo binding implicitly learn high-resolution affinity landscapes, and show that these affinities can be successfully distilled using AD. This new biophysical interpretation of deep learning models enables high-throughput in silico experiments to explore the influence of sequence context and variation on both intrinsic affinity and in vivo occupancy.

Short tandem repeats bind transcription factors to tune eukaryotic gene expression

“…5A), BPNet accurately predicted log-transformed read counts for held-out data ( R 2 = 0.52), with binding profiles that reproduced those observed experimentally (Fig. 5A) ( 86 ). Returned contribution weight matrices (CWMs), which identify short subsequences most predictive of TF binding, revealed E-box–like motifs (CACGTG) that sometimes included a flanking preference for CG dinucleotides, consistent with in vitro preferences (Figs.…”

“…This case study of STRs further underscores the limitations of motif-based models in predicting TF occupancy from sequence ( 37 , 86 , 127 – 130 ), because STRs composed of overlapping instances of even low-affinity sites bearing little resemblance to the known motif can substantially alter binding. Binding of the same TF to dissimilar motifs has previously been reported and attributed to alternate binding modes driven by either entropic or enthalpic effects ( 131 – 133 ).…”

“…Directly quantifying the impacts of STRs on TF binding in cells is technically challenging, because the lower-affinity binding expected for STRs is unlikely to yield distinct peaks within chromatin immunoprecipitation data. To sensitively quantify effects of STRs in vivo, we trained the BPNet ( 37 ) neural network (NN) on in vivo chromatin immunoprecipitation sequencing (ChIP-seq) data with 5-fold cross-validation to predict TF binding profiles from DNA sequence with nucleotide resolution and then applied Affinity Distillation (AD) ( 86 ) to predict log-transformed mean read counts [Δlog(counts)], which were previously shown to correlate with measured thermodynamic energies (ΔΔGs). If STRs alter gene expression in vivo by changing TF occupancies, then we would expect BPNet to learn that they affect TF binding and AD to predict sequence-dependent read count changes that mirror ΔΔGs measured in vitro.…”

“…All NN models were implemented and trained in Keras (v.2.2.4; TensorFlow backend v.1.14) ( 167 , 168 ) using the Adam optimizer ( 169 ). AD scores [∆log(counts)] were calculated by inserting a given sequence at the center of 100 different background sequences and computing the mean of the differences between the log(count) predictions for query sequence and background sequence alone, as described in ( 86 ).…”

Transcription factor stoichiometry, motif affinity and syntax regulate single-cell chromatin dynamics during fibroblast reprogramming to pluripotency

“…The models learn hierarchical layers of de novo sequence pattern detectors that can encode sequence motifs and their higher-order syntax. Interpretation of these models has revealed novel insights into the cisregulatory code of TF binding including sequence preferences and affinity landscapes of individual TFs (Avsec, Weilert, et al, 2021;Alexandari et al, 2023), soft motif syntax mediated TF cooperativity (Avsec, Weilert, et al, 2021;de Almeida et al, 2022), and effects of sequence variation and repeats (Horton et al, 2023;Alexandari et al, 2023;Avsec, Agarwal, et al, 2021;Chen et al, 2022). While neural networks have been used to dissect the sequence basis of chromatin accessibility of diverse cell types (Maslova et al, 2020;Kim et al, 2021;Janssens et al, 2022;Ameen et al, 2022), they have yet to be used to decipher cis-regulatory drivers of quantitative chromatin dynamics from single-cell profiling across continuous cell state transitions during reprogramming.…”