Learning Chemical Representations and Retrosynthetic Agents to aid Closed Loop Drug Discovery

这项研究将专注于开发由机器人科学家实现的帮助闭环药物发现的算法。重点将放在从药效团学习改进的、可解释的化学表示和综合可行的从头分子设计。学习给定任务的最佳化学表示将受到奥卡姆剃须刀的数学形式化的启发。这些概念跨越算法复杂性和类似的可判定，尽管版本较弱，例如最小消息长度(MML)、最小描述长度(MDL)原则以及常规和上下文无关的语言。还将研究确定表示n个化合物的最佳方式的问题。最初的研究将考虑最短语法问题的解决方案作为候选表示。在被认为代表化学的化合物语料库上解决这个问题可能会产生改进的官能团。我们将研究这些思想在图灵可判定语言中的扩展。改进的化学表示法将对化学的各个方面产生积极影响。新的官能团可以用来进一步解释化学性质、行为和治疗效果。描述一组药物分子的最短语法可能会捕捉到那些对药物活性重要的模式，假设这些模式在上下文无关的语言中是可识别的。设计具有化学合成知识的模型将需要逻辑代理。这些模型将通过在合成过程中良好的化学表示而得到改进。将探索归纳逻辑编程，以确定化学合成所需的一组有效规则。重点将放在所使用的分子表示法上。这一表象将考虑分子背景。将探索反向合成树搜索和强化学习。这类算法的开发将有助于化学和药物发现的速度。对该项目的期望包括开发符合Occam‘s Razor的替代化学表示法(考虑到分子背景)、从药效团构建潜在药物分子的生成模型以及基于知识的逆合成模型。

The study will focus on the development of algorithms aiding closed-loop drug discovery enabled by a robotic scientist. A focus will be placed on learning improved, interpretable chemical representations and synthetically feasible de-novo molecular design from a pharmacophore. Learning optimal chemical representations for given tasks will be inspired by mathematical formalisations of Occam's razor. These concepts span algorithmic complexity and analogous decidable, albeit weaker versions such as the minimum message length (MML), minimum description length (MDL) principle, and regular and context-free languages. The problem of determining the optimal way to represent n compounds will also be investigated. Initial studies will consider solutions to the shortest grammar problem as a candidate representation. Solutions to this problem over a corpus of compounds deemed representative of chemistry may produce improved functional groups. Extensions of these ideas to the Turing decidable languages will be investigated. Improved chemical representations would positively affect all aspects of chemistry. New functional groups could serve to further explain chemical properties, behaviour and therapeutic effect. The shortest grammar which describes a set of drug molecules would likely capture those patterns important for drug activity, assuming such patterns are recognisable within a context free language.Design of models with knowledge of chemical synthesis will require logical agents. These models will be improved by good chemical representations in the synthetic process. Inductive logic programming will be explored in determining a set of valid rules required for chemical synthesis. A focus will be placed upon the molecular representation used. This representation will account for molecular context. Retrosynthetic tree search and reinforcement learning will be explored.Development of such algorithms would aid the speed of chemical and pharmaceutical discovery. Expectations for the project include the development of alternative chemical representations in line with Occam's Razor which account for molecular context, generative models which construct potential drug molecules from a pharmacophore and knowledge based retrosynthetic models.