摘要：

在量子化学领域, 分子性质预测是一项基础而关键的任务, 广泛应用于药物发现、化学合成预测等多个领域。随着人工智能的发展, 深度学习方法在该领域得到了广泛应用。然而, 当前的方法往往采用微观和宏观视图两种极端的抽象层次来对分子性质进行建模, 导致难以推广到分布之外样本的挑战。化学的介观视图提供了一个有益的中间层次, 通过包含与性质相关的功能基团的介观成分来描述分子性质。通过考虑这些介观成分, 并从因果关系的角度对其进行建模, 可以更加关注与性质相关的功能基团。为了实现该目标, 提出一种介观成分识别模型。该模型基于分子数据的介观因果生成过程和变分自编码器的框架, 通过学习与分子性质相关的介观成分的表示, 实现对分子性质的预测。首先假设原子隐变量遵循高斯分布和语义隐子结构遵循多元伯努利分布, 将分子数据输入神经网络来识别原子隐变量和语义隐子结构。接着利用识别出来的原子隐变量和语义隐子结构来预测分子性质。为了能够识别出原子隐变量和语义隐子结构, 利用变分下界和稀疏项来构造模型的损失函数。实验结果表明, 该模型不仅在性能上取得先进的结果, 而且提供了深入的解释, 为模型预测提供了更全面的理解, 提高分子性质预测的准确性和泛化能力。

关键词: 分子属性预测, 因果, 分布外泛化, 图表征, 图神经网络

Abstract:

In the field of quantum chemistry, molecular property prediction is a fundamental and critical task, which is widely used in many fields such as drug discovery and chemical synthesis prediction. With the development of artificial intelligence, deep learning methods have been widely used in this field. However, current methods often adopt two extreme levels of abstraction, namely micro- and macro-views, to model molecular properties, posing challenges in generalizing to out-of-distribution samples. The mesoscopic view of chemistry provides a beneficial intermediate level for describing molecular properties through mesoscopic components containing functional groups associated with these properties. By considering these mesoscopic components and modeling them from a causal perspective, more attention can be paid to the functional groups related to these properties. To achieve this goal, this study proposes a Mesoscopic Component Identification(MCI) model. This model is based on a mesoscopic causal generative process that uses molecular data and a framework of variational autoencoders. The proposed model predicts molecular properties by learning the representation of mesoscopic components related to molecular properties. Initially, the model assumes that the atomic latent variables and semantic latent substructure follow Gaussian and multivariate Bernoulli distributions, respectively. Molecular data are then input into a neural network to identify the atomic latent variables and semantic latent substructure. Next, the identified atomic latent variables and semantic latent substructures are used to predict molecular properties. To identify the substructures of the atomic and semantic latent variables, variational lower bounds and sparse terms are used to construct the loss function of the model. Experiments demonstrate that our model not only achieves state-of-the-art performance but also offers in-depth explanations that provide a more comprehensive understanding of model predictions and improve the accuracy and generalization ability of molecular property predictions.

Key words: molecular property prediction, causal, out-of-distribution generalization, graph representation, Graph Neural Network(GNN)