Improving drug design to eliminate side effects: From computational to animal models

摘要 不必要的副作用是临床试验失败的主要原因。对于核受体（NR），一个突出的 治疗靶点家族代表了一个400亿美元/年的行业，副作用通常是通过药物引起的， 不能实现组织选择性。作为转录因子，核受体与辅因子蛋白相互作用 (i.e.,辅助调节因子），其通过染色质重塑和组蛋白修饰来调节基因表达， 因此在决定核受体的转录结果中起关键作用。核受体药物的活性是 由辅助调节因子的差异表达控制。在一个独特但知之甚少的机制中， 药物诱导构象变化，导致受体选择性结合辅调节因子。不幸的是， 非选择性NR药物驱动与预期和非预期共调节因子的关联。这种滥交使得 NR药物，以实现理想的治疗结果和意外的副作用。因为我们不 了解配体如何驱动NR选择性地与辅助调节因子结合，我们无法设计组织- 选择性药物，避免了不希望的辅调节因子结合的陷阱。该提案提出了一个创新的 了解和预测NR配体中选择性辅调节因子募集的方法。首先，通过结合祖先 蛋白质重建与计算研究和高通量NR-共调节基因分析，我们将确定 NR-药物复合物区别性地与辅助调节剂结合的生物物理基础。接下来， 机器学习模型来预测最有可能被各种药物招募的NR共调节因子。这些 模型将被开发成一个基于网络的生物信息学工具，允许筛选新的配体结构 基于其预测的共调节子关联。由于共调节因子的募集在不同的细胞中通知NR活性， 组织，该工具将代表一种强有力的方法来预测组织特异性，在体内活性的配体具有很大的 改变计算机药物设计的潜力。最后，我们将联合收割机结合我们对协同调节因子的机械理解， 招募和我们的计算预测工具走向一个独特的应用：药物的设计，以治疗非- 小鼠酒精性脂肪肝总之，这些研究将发展一个基本的， 并将这一知识与药物设计的强大应用联系起来。

Abstract Unwanted side effects represent a leading cause of clinical trial failures. For nuclear receptors (NRs), a prominent family of therapeutic targets representing a $40 billion/year industry, side effects often arise via drugs that are incapable of achieving tissue selectivity. As transcription factors, nuclear receptors interact with cofactor proteins (i.e., coregulators) which modulate gene expression through chromatin remodeling and histone modifications, thus playing key roles in deciding the transcriptional outcomes of NRs. The activity of nuclear receptor drugs is controlled by the differential expression of coregulators. In a unique but poorly understood mechanism, these drugs induce conformational changes that cause the receptor to selectively bind coregulators. Unfortunately, nonselective NR drugs drive association with both intended and unintended coregulators. This promiscuity allows NR drugs to achieve both desirable therapeutic outcomes and unintended side effects. Because we do not understand how ligands drive NRs to selectively associate with coregulators, we are unable to design tissue- selective drugs that avoid the pitfall of undesirable coregulator binding. This proposal presents an innovative approach to understand and predict selective coregulator recruitment in NR ligands. First, by combining ancestral protein reconstruction with computational studies and high throughput NR-coregulator profiling, we will determine the biophysical basis by which NR-drug complexes discriminately associate with coregulators. Next, we will train machine learning models to predict the NR coregulators most likely to be recruited by various drugs. These models will be developed into a web-based bioinformatics tool that allows new ligand structures to be screened on the basis of their predicted coregulator association. As coregulator recruitment informs NR activity in different tissues, this tool will represent a powerful approach to predict tissue-specific, in vivo activity of ligands with great potential to transform in silico drug design. Finally, we will combine our mechanistic understanding of coregulator recruitment and our computational prediction tool towards a unique application: the design of drugs to treat non- alcoholic fatty liver disease in mice. In sum, these studies will develop a fundamental, molecular understanding of coregulator selectivity and bridge this knowledge to the powerful application of drug design.