Deep learning models have emerged as rapid, accurate, and effective approaches for clinical decisions. Through a combination of drug screening and deep learning models, drugs that may benefit patients before and after surgery can be discovered to reduce the risk of complications or speed recovery. However, most existing drug prediction methods have high data requirements and lack interpretability, which has a limited role in adjuvant surgical treatment. To address these limitations, the authors propose the attention-based convolution transpositional interfusion network (ACTIN) for flexible and efficient drug discovery. ACTIN leverages the graph convolution and the transformer mechanism, utilizing drug and transcriptome data to assess the impact of chemical pharmacophores containing certain elements on gene expression. Remarkably, just with only 393 training instances, only one-tenth of the other models, ACTIN achieves state-of-the-art performance, demonstrating its effectiveness even with limited data. By incorporating chemical element embedding disparity and attention mechanism-based parameter analysis, it identifies the possible pharmacophore containing certain elements that could interfere with specific cell lines, which is particularly valuable for screening useful pharmacophores for new drugs tailored to adjuvant surgical treatment. To validate its reliability, the authors conducted comprehensive examinations by utilizing transcriptome data from the lung tissue of fatal COVID-19 patients as additional input for ACTIN, the authors generated novel lead chemicals that align with clinical evidence. In summary, ACTIN offers insights into the perturbation biases of elements within pharmacophore on gene expression, which holds the potential for guiding the development of new drugs that benefit surgical treatment.

An abstract is unavailable. This article is available as a PDF only.