Augmenting Predictive Accuracy in Pharmacological Mechanism of Action Models

Background: Despite the success of Deep Neural Networks in forecasting the therapeutic efficacy of oncological agents, the lack of explainability of their decision-making process is a significant challenge. They further grapple with two principal challenges: first, the deterministic nature of feature compression, which fails to capture the stochastic essence of data distributions during the inferential process; and second, the inherent limitations of multi-class classification models, which tend to overlook the nuanced interplay of "pro" and "anti" dynamics intrinsic to the target classes. Methods: We propose constructing a composite, interpretable neural network architecture designed to estimate the Mechanism of Action based on a reduced gene expression profile. Furthermore, we introduce a novel ternary loss function that exploits the inherent relationships among the related MoAs, thereby enhancing the model's predictive accuracy and overall efficacy. Findings: The model expedited the training process, reducing the computational time when compared with alternative tabular learning methodologies. It also demonstrated performance comparable to state-of-the-art deep learning frameworks designed for tabular data processing, even when applied to a condensed gene expression signature. Moreover, leveraging gene expression profiles has enabled the model to predict Mechanisms of Action with a training dataset encompassing approximately 5,000 distinct pharmaceutical compounds. Interpretation: The proposed model effectively predicts the Mechanism of Action of a drug while also allowing the assessment of the importance of the feature attributes. This method remains robust even when applied to a reduced dataset.