FET:Medium: Drug discovery using quantum machine learning

现有的药物发现管道从最初的想法到市场批准需要10-15年，花费数十亿美元。大量的时间归因于广阔的搜索空间和缺乏有效的搜索工具，而成本主要归因于在临床试验中失败的劣质候选药物。需要高质量的搜索工具来增加进入优化的候选药物的种类和质量。虽然人工智能（AI）辅助的高性能计算可以快速筛选大量化合物，以缩小具有各种理想特性的候选药物的范围，但候选药物的很大一部分潜在空间仍未开发。此外，它是计算昂贵的和低效的采样所需的概率分布在溶液空间中的分子数量呈指数增长。量子人工智能更有表现力，即，它甚至可以用有限数量的量子比特和参数来对目标概率分布建模，以从搜索空间的未探测区域进行采样。然而，它们在药物发现中的真正潜力和应用仍然未被探索。该项目将通过创建量子机器学习（QML）模型来填补这一空白，该模型将采用嘈杂的量子计算机。如果成功，该项目将在发现应用中释放新的计算能力，例如，通过选择新的先导化合物与重要的靶蛋白来治疗疾病，如癌症，通过融合多学科。通用和可扩展的QML工具集将使量子计算能够用于其他发现应用，例如，物质发现该项目将通过解决可扩展性问题来推进量子计算和量子AI。它将为K-12教师开发量子计算和应用的综合介绍，包括专业发展研讨会和课程材料，这些材料涉及科学和工程教育的地方和国家标准。它还将开发由宾夕法尼亚州立大学量子未成年人计划支持的本科课程，以准备量子就绪的劳动力。研究人员将开发DrugVAE（一种量子变分自动编码器）来搜索和筛选配体，并开发QDock（一种量子对接引擎）来验证配体并帮助筛选。还将开发各种可扩展性、应用程序级并行化和分布式计算的培训方法。研究人员将优化和并行化，映射和调度来自DrugVAE和QDock的QML工作负载到目标量子计算机中，同时考虑性能，弹性和成本的架构和硬件限制。输出特征将根据需要提供给经典神经网络。研究人员将通过计算验证QML生成的化合物与较慢的传统对接以及实验确定的结合亲和力。这项研究将为劳动力发展和本科课程提供材料。各种任务将通过新技术协同进行，例如QML特定优化，目标特定搜索和模型参数的细化，以及基于验证结果的优化。该项目将涵盖所有级别的抽象，以满足药物发现的最终目标，例如，该奖项反映了NSF的法定使命，并通过使用基金会的智力价值和更广泛的影响审查标准进行评估，被认为值得支持。

Existing drug discovery pipelines take 10-15 years from initial idea to market approval and cost billions of dollars. Extensive time is attributed to the expansive search space and lack of efficient search tools, whereas the cost is primarily attributed to inferior quality drug candidates that fail in clinical trials. High-quality search tools are required to increase the variety and quality of drug candidates that enter optimization. While high-performance computing assisted by Artificial Intelligence (AI) can screen a large pool of chemical compounds quickly to narrow down candidates that possess various desirable properties, a very large fraction of potential space for candidate drugs still goes unexplored. Furthermore, it is computationally expensive and inefficient in sampling the desired probability distributions in solution space which grows exponentially with the number of molecules. Quantum AI is more expressive, i.e., it can model a target probability distribution even with a limited number of qubits and parameters to sample from the unexplored regions of the search space. However, their true potential and application in drug discovery remain unexplored. This project will fill this void by creating Quantum Machine Learning (QML) models that will employ noisy quantum computers. If successful, this project will unleash new computational capabilities in discovery applications, e.g., by selecting novel lead chemical compounds versus important target proteins to treat diseases, such as cancer, by converging multiple disciplines. The generic and extendible QML toolset will enable the use of quantum computing for other discovery applications, e.g., material discovery. This project will advance quantum computing and quantum AI by addressing the scalability issue. It will develop an integrated introduction to quantum computing and application for K-12 teachers, including a professional development workshop and curricular materials that address local and national-level standards in science and engineering education. It will also develop undergraduate coursework supported by Penn State Quantum Minor program to prepare a quantum-ready workforce. Researchers will develop DrugVAE (a quantum variational autoencoder) to search and screen ligands and QDock (a quantum docking engine) to validate the ligands and aid in screening. Various scalability, application-level parallelization and training approaches for distributed computing will also be developed. Researchers will optimize and parallelize, map, and schedule the QML workloads from DrugVAE and QDock into target quantum computers considering architectural and hardware constraints for performance, resilience and cost. The output features will be provided to the classical neural network as needed. Researchers will computationally validate QML-generated compounds against slower, traditional docking as well as experimentally determined binding affinities. The research will provide materials for workforce development and undergraduate curriculum. Various tasks will be synergized through novel techniques, such as QML-specific optimization, target-specific search and refinement of model parameters, and optimization based on validation results. This project will cover all levels of abstractions to meet the end goal of drug discovery, e.g., program/circuit design, optimization, circuit-to-architecture mapping, parallelization, and scheduling.This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.