An informatics bridge over the valley of death for cancer Phase I trials of drug-combination therapies

一座跨越死亡之谷的信息学桥梁，用于药物联合癌症治疗的第一阶段试验 摘要 第一阶段研究通常集中于药物毒性和药代动力学，大多数(58%)药物用于治疗癌症 治疗方法未能通过这些初步试验。因此，第一阶段研究是药物治疗过程中最大的死亡谷。 发展。与单一药物I期研究的设计不同，药物联合研究的设计需要 事先了解其中一种药物是否会改变另一种药物的暴露情况，这两种药物都有毒性，而且每种药物 药物有既定的最大耐受量。尽管有丰富的毒性和PK数据可供公开使用 域源，数据没有被集成，并且没有单个数据库集成关于毒性和 PK。此外，文献中也有关于单一药物和药物组合的MTD和DLT的数据，但 没有出现在任何数据库中。我们相信，一座横跨第一阶段死亡之谷的桥梁可以建成 利用信息学和药物计量学方法进行癌症多药研究和开发 单一药物的丰富毒性和PK数据的优势。在这笔赠款中，我们建议一项 整合毒性和PK数据的药物相互作用翻译知识库(TDCKB)。目标1将会发展 新的主动学习方法来挖掘关于药物相互作用的毒性和PK证据 文学。主动学习方法将采用几项创新，包括随机负抽样， 基于PubMed查询的分层主动学习和嵌入的深度学习。决赛 主动学习方法是通过对这些创新成分的彻底整合而得到优化的。目标2将会发展 癌症研究的翻译药物相互作用知识库(TDCKB)。TDCKB将整合毒性 以及来自各种数据源的单一药物和药物组合的PK证据。DDI的证据将 被归类为毒性或PK，证据的强度将被批注。综合证据， 例如两种药物之间重叠毒性和预测的药物相互作用，将有助于I期药物 联合试验设计。在数据管理和TDCKB期间，将仔细进行质量控制 软件开发。计划让TDCKB用户和ITCR社区参与进来。

An informatics bridge over the valley of death for Phase I trials of drug-combination cancer therapies Summary Phase I studies usually focus on drug toxicity and pharmacokinetics, and most (58%) drugs intended as cancer therapies fail these initial trials. Thus, Phase I studies represent the largest valley of death in the course of drug development. Unlike the design of a single-drug Phase I study, the design of a drug-combination study requires prior knowledge of whether either drug changes the other’s drug exposure, the drugs share toxicities, and each drug has an established maximum tolerable dose. Although abundant toxicity and PK data are available in public domain sources, the data are not integrated, and no single database integrates data regarding both toxicity and PK. In addition, data regarding single-drug and drug-combination MTD and DLT are present in the literature but absent from any database. We are confident that a bridge can be built across the Phase I valley of death for cancer multi-drug research and development utilizing an informatics and pharmacometrics approach to take advantage of the abundant toxicity and PK data available for single drugs. In this grant, we propose a translational drug-interaction knowledgebase (TDCKB) that integrates toxicity and PK data. Aim 1 will develop novel active-learning approaches to mine evidence of toxicity and PK regarding drug interactions from the literature. The active learning methodology will employ several innovations, including random negative sampling, stratified active learning by prescreening based on PubMed query, and deep learning with embedding. The final active-learning method is optimized by a thorough integration of these innovative components. Aim 2 will develop a translational drug-interaction knowledgebase (TDCKB) for cancer research. The TDCKB will integrate toxicity and PK evidence for single drugs and drug combinations from various data sources. The evidence of DDI will be classified as either toxicity or PK, and the strength of the evidence will be annotated. Synthesized evidences, such as overlapping toxicity and predicted drug interactions between two drugs, will assist in Phase I drug combination trial design. Quality control will be conducted carefully during both data curation and TDCKB software development. Engagement of TDCKB users and the ITCR community is planned.