基于人工智能与计算化学的多维度靶向药物虚拟筛选

A Multi-Dimensional Virtual Screening for Targeted Drugs Based on Artificial Intelligence and Computational Chemistry

摘要: 现代药物的研发存在着周期长、成本大、传统的高通量筛选成功率低等诸多问题。为了快速找到高效低毒靶向药物(抗肿瘤微管抑制剂)的候选物，本文从人工智能、计算化学两方面来综合论述虚拟筛选的方法。该计算管线以生物活性数据库为基础，用机器学习分类预测模型建立基于配体的预测体系，并采用SHAP归因分析破解算法“黑箱”方式来解释结果，给药物微观构效关系赋予了直观的物理指引。其次使用核心药效团模型和SMARTS规则对千万级的商业化化合物库做骨架硬编码筛选，可以快速缩减化学空间。接着利用基于受体的半柔性三维分子对接技术来确定候选小分子和受体蛋白空腔的构象契合程度以及结合自由能。最终使用无监督结构聚类技术来规避结构风险，同时配合ADMET早期预测，从中筛选出有较高靶点亲和力以及较好的临床转化前景的先导化合物。这样一种漏斗式高效的集成计算管线，很好地克服了传统研发的盲目性，给精准药物设计提供了一个可靠的数字化闭环方案。

Abstract: Current problems in the development of new drugs include a long cycle, high cost and a low success rate for traditional high-throughput screening. To accelerate the discovery of candidates for high-efficiency and low-toxicity targeted drugs (anti-tumor microtubule inhibitors), this paper introduces a multi-dimensional virtual screening strategy based on artificial intelligence and computational chemistry. First, the computational pipeline builds ligand-based machine learning classification models using bioactivity databases, and then applies SHAP attribution analysis to interpret the “black box” of the algorithm and provide intuitive physical guidance for microscopic structure-activity relationships. Second, it uses core pharmacophore models and SMARTS rules to perform hard-coded scaffold filtering on commercial compound libraries containing tens of millions of compounds and achieve a rapid reduction of chemical space. Next, receptor-based semi-flexible three-dimensional molecular docking technology is employed to determine the degree of conformational fit between candidate small molecules and the receptor protein cavity and calculate the binding free energy. Finally, unsupervised structural clustering technology is employed to mitigate structural risks; in conjunction with early ADMET prediction, high-affinity target-binding lead compounds with good clinical translation potential are carefully selected. This high-efficiency funnel-shaped integrated computational pipeline can effectively solve the problem of blindness in traditional R&D and provide a reliable digital closed-loop solution for precision drug design.