Deep GONet: Self-Explainable Phenotype Prediction

A report on the paper "Deep GONet: self-explainable deep neural network based on Gene Ontology for phenotype prediction from gene expression data" is provided below:

This paper proposes a new deep learning model called Deep GONet for phenotype prediction based on gene expression data. In the discussion section, the authors emphasize that the goal of interpreting this model is to explain its¹ operational mechanism, rather than necessarily how biological processes function. They point out that there may not always be a direct relationship between the biologically interpreted functions and the predicted phenotype, but this does not necessarily mean that the predictions are unreliable. The model seeks to find correlations between input and output, not causal relationships. If a function that appears unrelated to the phenotype is returned, it is possible that this function has an indirect correlation or is linked to the phenotype through an unknown causal relationship. However, the more biological functions in the interpretation that are consistent with the phenotype, the greater the confidence in the model's predictions. Conversely, a major inconsistency between the interpretation and existing biological knowledge raises questions about the model's reliability and suggests the possibility of overfitting or bias in the training data.

The authors believe that although model interpretation is not a direct tool for biological discoveries, certain aspects of their neural network, such as high-weighted noGO connections and neurons diverted from GO terms, can be investigated in this regard. These elements connect to probes that lack annotations, and examining why they are used in prediction and how their expression is combined in the hidden layer could lead to new biological insights. Probes connected to a similar neuron may have biological functions related to the predicted phenotype. Therefore, the Deep GONet model can contribute to the enrichment of Gene Ontology by providing new hypotheses that require further validation through biological experiments.

In the conclusion section, the authors state that Deep GONet is a self-explainable deep learning model whose prediction performance is comparable to classical deep learning and machine learning methods. The architecture of this model is designed to be interpretable and understandable for biologists, as its structure reflects the biological knowledge they utilize. Each layer of the model corresponds to a level of the Gene Ontology, and each neuron corresponds to a GO term. The use of a custom regularization method helps the model better consider this knowledge by focusing on the actual connections between biological elements. Experiments conducted in the field of cancer diagnosis demonstrate how a simple interpretation of the model and its predictions can be provided, making it understandable for physicians and biologists. Although the current architecture of Deep GONet is based on GO-BP, other ontologies structured as a directed acyclic graph (DAG), such as GO-CC and GO-MF, can also be used within this framework. Furthermore, the model has the potential to be applied to other gene expression datasets and other prediction tasks, such as determining the type of cancer or prognosis, although this would require retraining the model.

In future plans, the authors intend to improve Deep GONet by adding neurons to manage genes without GO annotations, as well as a second branch to display biological pathways in order to enrich the interpretations. They also plan to further investigate the relationship between the activation of a neuron and the activation of its corresponding biological function.

The source of this report is the paper "Deep GONet: self-explainable deep neural network based on Gene Ontology for phenotype prediction from gene expression data."²