Elucidating Fingerprints – Towards a Holistic Explanatory Toolbox for Molecular Machine Learning

该提案的中心点是在结构层面上开发可解释和可解释的分子机器学习。在该项目中，将开发、调整和使用广泛使用的分子表示来训练高度稳健但准确的模型（例如梯度增强算法）。从这些模型开始，将采用开源软件管道将特征重要性，影响力，相互依赖性以及模型置信度映射回分子结构，为训练有素的化学家提供分子和反应设计的简单处理。这项工作的一个重要部分将涉及基于分析结果的可视化开发，一方面提供高度的准确性，另一方面易于理解任何在分子科学领域工作的科学家。这些工具应可用于研究和改进基础数据集以及分子设计。除了单个分子的着色和可视化之外，还应该开发关于官能团的一般影响的统计评估方法，以便可以导出进一步反应设计的规则。最后，这些规则应在实验室中使用，以验证在本提案过程中开发的解释方法。通过这些目标，该提案旨在实现PP的以下总体目标：“最先进的ML算法的应用-可解释的AI”，“（特定领域）分子表示的开发-一般改进的分子表示”和“分子特性的预测，理解和解释-改进当前应用”。在此范围内，高度关注定量产量预测的解释和解释模型，以便在MML的这一欠发达领域内找到系统改进的方法，这也被定义为本PP的主要主题。

The central point of this proposal is the development of out-of-the-box for interpretable and Explainable Molecular Machine Learning on a structural level. Within this project broadly utilized molecular representations will be developed, adapted and used to train highly robust but accurate models (e.g. Gradient Boost algorithms). Starting from these models an open-source software pipeline will be employed to map feature importance, influence, interdependencies, as well as model confidences back to the molecular structure giving trained chemists a plain handle for molecular and reaction design. An important part of this work will involve the development of visualization based on analytic results that provide a high degree of accuracy on the one hand and are easy to understand for any scientist working in the field of molecular science on the other hand. Those tools shall be usable to investigate and improve underlaying datasets as well as for molecular design. In addition to the coloration and visualization of individual molecules, methods of statistical evaluation regarding the general influence of functional groups should be developed, so that rules for further reaction design can be derived. Finally, these rules should be used in the laboratory to validate the explanatory methods developed within the course of this proposal. By these objectives the proposal aims on fulfilling the following of the PPs general goals: “Application of state-of-the-art ML algorithms – Explainable AI”, “Development of (domain specific) molecular representations – Generally improved molecular representations” and “Prediction, understanding and interpretation of molecular properties – Improvement of current applications”. Within this scope a high focus lies on the interpretation and explanation models for quantitative yield prediction to find handles for a systematic improvement within this underdeveloped area of MML which also has defined as a major topic of this PP.