Biocomputation across distributed private datasets to enhance drug discovery

项目摘要 协作药物发现（CDD）提出开发技术，将大大简化和整合所有 利用药物发现的预测模型所需的过程。该软件将使它很容易为 没有接受过信息学专业培训的科学家创建、培训、应用、评估、共享和存档模型 以最小的努力，还利用了一个大型库的预先计算的模型，零努力。该软件将 还使在不同组织工作的科学家能够从他们的汇总数据中共同构建模型 并共享这些模型，而不共享底层的训练数据。 我们的目标是使计算模型在药物发现中的作用民主化-历史上， 仅限于计算专家-并允许模型成为发现工作流程的常规辅助工具 在学术界，基金会，政府实验室和小公司，没有资源， 今天雇用他们。在第二阶段，我们直接在CDD的网络中实现了改进的贝叶斯模型构建， 基于CDD Vault平台，可安全地托管结构-活动关系（SAR）数据;任何用户现在都可以 使用存储在她私人保险库中的实验数据轻松训练贝叶斯模型，然后将该模型应用于 预测未测试化合物的活性。在阶段2B中，我们建议使用以下内容来概括此功能 新的具体目标，这是实现广泛的科学和商业影响所必需的： 目标1：集成一套不同的计算技术（如QSAR、神经网络、支持向量 机器，随机森林，k-最近邻，以及可能的其他）到一个单一的框架，以允许直接 并排比较。 目标2：开发和验证一个通用指标，对应用于以下方面的每种方法的预测强度进行排名： 一个特定的数据集。 目标3：应用该指标从高质量的公共访问中自动生成数千个模型 结构-活性和ADME/Tox数据集，并向用户呈现关键结果。 目标4：通过聚合多个数据集并共享，开发协作构建模型的新功能 这些模型没有揭示训练集中的化合物和数据。

Project Summary Collaborative Drug Discovery (CDD) proposes to develop technology that will vastly simplify and integrate all the processes required to exploit predictive models for drug discovery. The software will make it easy for scientists without specialized training in informatics to create, train, apply, evaluate, share, and archive models with minimal effort, and also leverage a large library of pre-computed models with zero effort. The software will also enable scientists working in different organizations to collectively build models from their aggregated data and share these models, without sharing the underlying training data. Our goal is to democratize the role in drug discovery of computational models – which have historically been restricted to computational experts – and allow models to become routine aids to the discovery workflow in academia, foundations, government laboratories, and small companies that do not have the resources to employ them today. In Phase 2 we implemented modified Bayesian model building directly within CDD’s web- based CDD Vault platform, which securely hosts structure-activity relationship (SAR) data; any user can now easily train a Bayesian model with experimental data stored in her private Vault, then apply the model to predict activity for untested compounds. In Phase 2B we propose to generalize this capability with the following new Specific Aims, which are needed to achieve a widespread scientific and commercial impact: Aim 1: Integrate a suite of diverse computational techniques (such as QSAR, Neural Networks, Support Vector Machines, Random Forest, k-Nearest Neighbors, and possibly others) into a single framework, to allow direct side-by-side comparison. Aim 2: Develop and validate a universal metric that ranks the predictive strength of each method as applied to a particular dataset. Aim 3: Apply the metric to automatically generate thousands of models from high-quality, public-access structure-activity and ADME/Tox datasets and present key results to the user. Aim 4: Develop a novel capability to build models collaboratively, by aggregating multiple datasets, and share the models without revealing the compounds and data in the training sets.