De novo inference of thermodynamic binding energies using deep learning models of in vivo transcription factor binding

We introduce Affinity Distillation (AD), a method for extracting thermodynamic affinities de-novo from in-vivo immunoprecipitation experiments using deep learning. We show that neural networks modeling base-resolution in-vivo binding profiles of yeast and mammalian TFs can accurately predict energetic impacts of varying underlying DNA sequence on TF binding. Systematic comparisons between Affinity Distillation predictions and other predictive algorithms consistently show that Affinity Distillation more accurately predicts affinities across a wide range of TF structural classes and DNA sequences. Affinity Distillation relies on in-silico marginalization against many sequence backgrounds, resulting in a higher dynamic range and more accurate predictions than motif discovery algorithms. Moreover, we show that Affinity Distillation can learn differential paralog-specific affinities, thereby making it possible to more accurately reconstruct regulatory networks in cells.我们提出了亲和力蒸馏（Affinity Distillation，AD）方法，该方法可借助深度学习从体内免疫沉淀（in vivo immunoprecipitation）实验中从头（de novo）提取热力学亲和力。我们证实，对酵母与哺乳动物转录因子（Transcription Factors, TFs）的碱基分辨率体内结合谱进行建模的神经网络，能够精准预测不同潜在DNA序列对转录因子结合产生的能量影响。将亲和力蒸馏的预测结果与其他预测算法进行系统性对比后可发现，在涵盖多种转录因子结构类别与DNA序列的广泛范围内，亲和力蒸馏均能更准确地预测亲和力。相较于基序发现算法，亲和力蒸馏通过对大量序列背景开展计算机模拟（in silico）边缘化操作，可获得更高的动态范围与更精准的预测结果。此外，我们证实亲和力蒸馏能够学习得到差异旁系同源特异性亲和力，从而可更精准地重构细胞内的调控网络。