2259P - An artificial neural network system to predict the fraction of type I polarized macrophage

Background

Type I polarized macrophages (M1) play an anti-tumor role in the tumor microenvironment (TME), and more M1 infiltration means better prognosis. Previous studies have developed algorithms to predict the fraction of M1 based on transcriptomic data, whereas few study focus on DNA methylation data.

Methods

A macrophage multiplex immunofluorescence (mIF) panel (including DAPI, panCK, CD68, CD86, IRF5, CD163 and CD206) and DNA methylation sequencing were used in lung adenocarcinoma (LUAD) patients’ formalin-fixed paraffin-embedded samples. The fraction of M1 (with more positive M1 markers [CD86 and IRF5] than M2 markers [CD163 and CD206]) was calculated based on all macrophages (DAPI⁺panCK^-CD68⁺). Patients were randomly divided into training cohort (70%) and test cohort (30%). The Pearson correlation and random forest algorithm were used to screened genes to build artificial neural network (ANN) models in the training cohort. Three hundred times of three-fold cross validation were performed to obtain 900 ANN models and test the accuracy of them. The final predicted value of the ANN system were equal to the weighted average of output values of all 900 ANN models. Finally, the accuracy of the ANN system was tested validated in the test cohort.

Results

Forty patients with LUAD were enrolled, and the median fraction of M1 was 3.27% (range: 0% to 30.91%). A total of 43 genes were with p-value < 0.001 were selected by Pearson correlation, and the top eight genes (C10orf90, ITGAX, TMED9, RGN, CENPF, LGI4, PPCS and KIAA1211) from random forest algorithm were used to build 900 three-layer ANN models. Each ANN model had eight input nodes, five hidden nodes and one output node. The frequency of ANN model whose weight > 0.70 (the accuracy > 0.70) was 74.78% (673/900). The Pearson correlation coefficient between the predicted value and the true M1 fraction was 0.89 (R² = 0.79, p < 0.01) in the training cohort and 0.80 (R² = 0.65, p < 0.01) in the test cohort.

Conclusions

DNA methylation data is more suitable to establish models than RNA data. In this study, an ANN system based on DNA methylation data was constructed and showed a potential value in predicting the fraction of M1 in LUAD, which could help to further understand the TME and identify patients that would benefit from immunotherapy.

Clinical trial identification

Editorial acknowledgement

Legal entity responsible for the study

The authors.

Funding

Has not received any funding.

Disclosure

All authors have declared no conflicts of interest.