Structured explainability for interactions in deep learning models applied to pathogen phenotype prediction

解释和理解基因组区域的潜在相互作用对于正确的病原体表型表征至关重要，例如预测生物体的毒力或对药物的耐药性。由于数据的高维性，现有的基因组序列基础大规模数据分类方法在可解释性方面面临挑战，这使得分类决策难以可视化、访问和证明。这在存在相互作用的情况下尤其如此，例如基因组区域。为了应对这些挑战，我们将开发变量选择和结构化可解释性的方法，以捕获重要输入变量之间的相互作用：更具体地说，我们将解决这些挑战(i)在融合广义线性混合模型和结构化预测器深度变体的二元结果的深度混合模型框架内。因此，我们将统计逻辑回归模型与深度学习相结合，以解开基因组数据中复杂的相互作用。当没有明确的公式输入可用于模型时，我们特别启用估计，例如与基因组学数据相关。此外，（ii）我们将扩展分类决策的可解释性方法，如分层相关传播来解释这些相互作用。在模型和可解释性水平上研究这两种互补的方法，我们的主要目标是制定和假设结构化的解释，不仅给出分类决策的一阶单变量解释，而且考虑它们之间的相互作用。虽然我们的方法是由我们的基因组数据驱动的，但它们可以很有用，并扩展到其他对相互作用感兴趣的应用领域。

Explaining and understanding the underlying interactions of genomic regions are crucial for proper pathogen phenotype characterization such as predicting the virulence of an organism or the resistance to drugs. Existing methods for classifying the underlying large-scale data of genome sequences face challenges with regard to explainability due to the high dimensionality of data, making it difficult to visualize, access and justify classification decisions. This is particularly the case in the presence of interactions, such as of genomic regions. To address these challenges, we will develop methods for variable selection and structured explainability that capture the interactions of important input variables: More specifically, we address these challenges (i) within a deep mixed models framework for binary outcomes fusing generalized linear mixed models and a deep variant of structured predictors. We thereby combine statistical logistic regression models with deep learning for disentangling complex interactions in genomic data. We particularly enable estimation when no explicitly formulated inputs are available for the models, as for instance relevant with genomics data. Further, (ii), we will extend methods for explainability of classification decisions such as layerwise relevance propagation to explain these interactions. Investigating these two complementary approaches on both the model and explainability levels, it is our main objective to formulate and postulate structured explanations that not only give first-order, single variable explanations of classification decisions, but also regard their interactions. While our methods are motivated by our genomic data, they can be useful and extended to other application areas in which interactions are of interest.