A Siamese neural network for immunotherapy response prediction

Immune checkpoint inhibitors (ICIs) have emerged as groundbreaking treatments for varioustypes of cancer. However, only a portion of patients with solid tumors experience positive outcomesfrom ICI therapies, which sometimes come with significant side effects and costs. Therefore, accurateprediction of ICI response is crucial. The scarcity of patient samples treated with ICIs, which have bothclinical outcomes and transcriptomic data, has posed a critical challenge in developing predictors forpatient response. To address this issue, we first identify pathway-based biomarkers and estimate thepathway expression levels, effectively reducing the dimensionality of the data. We utilized Reactomepathways and conducted single-sample gene set enrichment analysis (ssGSEA) to compute theexpression levels of pathways. The normalized enrichment score (NES) employed as each sample'spathway expression levels. Second, we propose a novel model that applies a homogeneous Siameseneural network, which takes the pathway expression levels of two patients as input to predict thesimilarity of their ICI response. As a result, we generate a new dataset and increase the sample size.This dataset contains pairs of patients and the binary similarity of their ICI response. We utilize a datasetcomprising ۹۱ ICI-treated patient samples. Remarkably, our model achieves an accuracy of ACC =۰.۷۸, outperforming the last existing model that uses logistic regression with an accuracy of ACC =۰.۷۳.

کلیدواژه ها:

Siamese neural network ، immune checkpoint inhibitors ، pathway expression ، machine learning