Drug resistance threatens many critical therapeutics through mutations in the drug target. The molecular mechanisms by which combinations of mutations, especially involving those distal from the active site, alter drug binding to confer resistance are poorly understood and thus difficult to counteract. A machine learning strategy was developed that couples parallel molecular dynamics simulations and experimental potency to identify specific conserved mechanisms underlying resistance. A series of 28 HIV-1 protease variants with 0-24 substitutions each were used as a rigorous model of this strategy. Many of the mutations were distal from the active site and the potency of variants to a drug (darunavir) varied from low picomolar to near micromolar. With features extracted from the simulations, elastic network machine learning was applied to correlate physical interactions with loss of potency and succeeded to within 1 kcal/mol of experimental affinity for both the training and test sets, outperforming MM/GBSA calculations. Feature reduction resulted in a model with 4 specific features that describe interactions critical for potency for all 28 variants. These predictive features, that specifically vary with potency, occur throughout the enzyme and would not have been identified without dynamics and machine learning. This strategy thus captures the conserved dynamic mechanisms by which complex combinations of mutations confer resistance and identifies critical features that serve as bellwethers of loss of inhibitor potency. Machine learning models leveraging molecular dynamics can thus elucidate mechanisms of drug resistance that confer loss of affinity and will serve as predictive tools in future drug design.

中文翻译：

---

通过耦合的分子动力学和机器学习来探索抵抗力和效能丧失的复杂机制。

耐药性通过药物靶点的突变威胁到许多重要的治疗方法。突变的组合，尤其是涉及活性位点远端的突变，通过组合改变了药物结合以赋予耐药性的分子机制了解甚少，因此难以抵消。开发了一种机器学习策略，该策略将并行的分子动力学模拟与实验能力相结合，以确定潜在的抗性保守机制。一系列28个HIV-1蛋白酶变异体（每个变异体0-24个被替换）用作该策略的严格模型。许多突变都远离活性位点，并且变种对药物（darunavir）的效力从低皮摩尔到接近微摩尔。利用从仿真中提取的特征，运用弹性网络机器学习将物理相互作用与效能损失相关联，并成功地在1 kcal / mol的实验亲和力内对训练和测试集进行了测试，优于MM / GBSA计算。特征减少导致模型具有4个特定特征，这些特征描述了对所有28个变体的效能至关重要的相互作用。这些预测特征随效力而异，在整个酶中都存在，如果没有动力学和机器学习就无法确定。因此，该策略捕获了保守的动态机制，通过该机制，突变的复杂组合赋予了抗性并确定了关键特征，这些关键特征是抑制剂效能丧失的领头羊。