Centralized assay datasets for modelling support of small drug discovery organizations

项目摘要 Collaborations Pharmaceuticals，Inc.是在确定了对软件的需求，以协助学者， 小公司在策划他们的数据和发现新的命中或铅优化。近两年 人工智能（AI）的持续重要性从这些数字的爆炸性增长中显而易见。 公司和越来越多的数百万美元的交易与制药使用机器学习（ML）， 协助药物发现。这些公司非常关注药物发现建模方面， 在体外和体内数据ADME/Tox数据的质量管理方面，仍然存在未满足的需求和瓶颈 以及验证技术的前瞻性测试。在第一阶段，我们开发了一个 中央数据库软件，并使用它与各种各样的结构活动数据来源，无论是公共和私人， 格式化和非格式化，约14名合作者致力于被忽视的，罕见或常见的疾病目标， 并将其用于我们的内部药物发现项目。在第一阶段，我们还创建了错误检查和更正 软件我们还使用收集和清理的数据集构建和验证了贝叶斯模型。而且， 此外，我们还开发了新的数据可视化工具。该软件可用于创建这些选择 用于根据需要与合作者共享的模型，以及用于对新分子进行评分和可视化多个分子的模型。 以各种格式输出。在第二阶段，我们已经将Assay CentralKit开发成一个易于使用的生产工具， 以行业标准技术为基础，提供了模型和模型信息的图形显示 适用性重要的是，我们发现客户希望我们为他们提供结果！我们开发 我们的按服务收费咨询服务模式使用分析中心来解决他们的问题， 我们的收入逐年增加。在第二阶段，我们评估了额外的ML算法和分子描述符 通过手动管理的数据集以及来自5000多个自动管理的数据集的比较算法， 化学这说明了访问多个算法的效用以及贝叶斯算法是如何 通常与这些其他ML算法相当。这也促使我们开发新的软件来整合 这些算法。我们还探索了寻找罕见疾病数据集并应用我们的数据管理和ML 接近他们。有了这些和额外的合作，以及对阿尔茨海默病的内部项目 （通过NIH NIGMS补充）我们已经能够将已经批准的药物重新用于几个目标 对于这种疾病和其他疾病。对于多个项目，我们已经进行了几轮模型构建和反馈， 将数据返回到模型中，以改进预测。最后，我们开发了原型工具， 我们开发自动化分子设计，评估其合成能力和进行逆合成分析。 这些共同的努力极大地增加了我们能够工作的项目数量（最终 发布以提高我们的知名度），创建了新的衍生产品作为模型集合（MegaTransmittance，MegaToxicity 和MegaPredictands）、分子相关的知识产权以及创造的就业机会。在IIB阶段，我们现在建议重点关注 采取措施，帮助这些技术的商业化和进一步发展。我们已经确定， 开发用于处理非结构化数据库中复杂生物数据的自动管理软件将是一个 竞争优势我们还认识到，对于许多疾病， 完整的目标集合，使我们能够了解分子如何干扰生物学 这可以应用于复杂的疾病和“不良后果途径”， 毒理学我们还提出了整合最先进的多目标生成模型的分子设计 我们的分析中心计算软件，以补充我们的模拟生成和逆合成 在第二阶段创建的工具，并有助于分子优化。我们将使用一些热门的 在II期鉴定的分子用于不同的靶点，包括人乙酰胆碱酯酶。分析中心将 然后拥有从数据管理到分子设计和逆合成的全套集成功能 分析，并将使我们能够吸引更大的交易与公司。

Project Summary Collaborations Pharmaceuticals, Inc. was formed after identifying a need for software to assist academics and smaller companies in curating their data and discovery of new hits or lead optimisation. In the past two years the continued importance of artificial intelligence (AI) is apparent from the explosive growth in number of these companies and the increasing number of multi-million dollar deals with pharma using Machine Learning (ML) to assist in drug discovery. There is a heavy focus by these companies on the drug discovery modeling aspect but there is a continued unmet need and bottleneck in the curation of quality in vitro and in vivo data ADME/Tox data for ML as well as prospective testing to validate the technologies. In Phase I, we developed a prototype of Assay CentralÒ software and used this with a wide variety of structure activity data from sources both public and private, formatted and unformatted, with ~14 collaborators working on neglected, rare or common disease targets as well as used it for our internal drug discovery projects. In Phase I we also created error checking and correction software. We also built and validated Bayesian models with the datasets that were collected and cleaned. And, in addition, we developed new data visualization tools. The software can be used to create selections of these models for sharing with collaborators as needed and for scoring new molecules and visualizing the multiple outputs in various formats. In Phase II, we have developed Assay CentralÒ into a production tool which is easy to deploy, built on industry standard technologies, provided graphical display of models and information on model applicability. Importantly, we identified that customers wanted us to provide them with the results! We developed our fee-for-service consulting services model using Assay CentralÒ to solve their problems and this has expanded our revenues annually. In Phase II we evaluated additional ML algorithms and molecular descriptors with manually curated datasets as well as compared algorithms across over 5000 auto-curated datasets from ChEMBL. This illustrated the utility of access to multiple algorithms and how the Bayesian algorithm was generally comparable to these other ML algorithms. This also motivated us to develop new software to integrate these algorithms. We have also explored finding rare disease datasets and applying our data curation and ML approach to them. With these and additional collaborations, as well as internal projects on Alzheimer’s disease (through a NIH NIGMS supplement) we have been able to repurpose already approved drugs for several targets for this and other diseases. For multiple projects we have performed several rounds of model building and fed data back into the models to enable improved predictions. Finally, we have developed prototype tools to enable us to develop automated molecule designs, assess their synthesizability and perform retrosynthetic analysis. These combined efforts dramatically increased the number of projects we were able to work on (and ultimately publish to raise our visibility), created new spin off products as collections of models (MegaTransÒ, MegaToxÒ and MegaPredictÒ), molecule related IP, and generated employment. In Phase IIB we now propose a focus on steps to aid commercialization and further development of these technologies. We have identified that developing auto-curation software for dealing with complex biological data in unstructured databases will be a competitive advantage. We have also recognized that for many diseases we can have a complete or near complete collection of targets which may enable us to understand how a molecule may interfere with biological pathways from structure alone and this can be applied to complex diseases and “adverse outcome pathways” in toxicology. We also propose integrating state of the art multi-objective generative models for molecule design into our Assay Central computational software in order to complement our analog generation and retrosynthesis tools created in Phase II and aid in molecule optimization. We will validate this capability using some of the hit molecules identified in Phase II for different targets including human acetylcholinesterase. Assay Central would then have a full suite of integrated capabilities from data curation through to molecule design and retrosynthetic analysis and will enable us to attract larger deals with companies.