NeuRGI model and OntoVAE model performance.

(A) Gene Ontology Biological Processes (GO-BP) enrichment of 293 PU-learning negative genes. These genes are not related to any neutrophil pathway. (B) Dot plot showing the feature importance of the baseline model using Gini coefficient with all features included. (C) Feature ablation studies. The line plots depict changes in the model’s evaluation metrics—AUC, AUC-PR, ACC, MCC, F1 Score, and Mean (average of these five metrics)—as features are sequentially removed in reverse order of their importance. The results indicate that the model achieves optimal performance when the last seven features are excluded. (D) Density distribution of NeuRGI scores across different gene categories, including predicted function (red, 4,786 genes), predicted non-function (blue, 4,734 genes), predicted uncertain (grey, 9,768 genes), known positive (light red, 293 genes), and PU-learning negative (purple, 293 genes). (E) Pie chart illustrating the NeuRGI classification of 70 CellOracle-predicted TFs and 36 SCENIC-predicted TFs. (F) The boxplot illustrates the transformed feature values for 70 CellOracle-predicted TFs (up) and 36 SCENIC-predicted TFs (down) in four feature groups. The p value was calculated using the Student’s t-test. (G) Heatmap illustrates the results of Kmeans clustering applied to the top 100 GO-BP terms associated with a set of 4,786 predictive functional genes. Rows and columns represent unique GO terms, with color intensity indicating similarity, where darker red denotes greater similarity. (H) The trend of test loss in OntoVAE model training. After 300 epochs of model training, the loss value no longer decreases and remains stable. (I) Scatter plot showing example pathway activity scores retrieved from OntoVAE model. The ‘neutrophil activation’ pathway is especially active in neutrophils and the ‘regulation of mononuclear cell proliferation’ pathway is especially active in monocytes. (PDF)