Backgrounds and Aims Microsatellite instability (MSI) has become an important biological marker for clinical diagnosis, adjuvant therapy, and prognostic guidance in colorectal cancer (CRC). Microsatellite instability often accompanies the loss of DNA mismatch repair proteins (dMMR). Currently, the diagnosis of mismatch repair protein deficiency mainly relies on the results of pathological immunohistochemistry for four repair proteins (MLH1, MSH2, MSH6, and PMS2), and MSI has become an important biological marker for immunotherapy in CRC. However, there are few studies on precise MSI prediction models and new signature genes. With the development of artificial intelligence in medicine, precise prediction and data mining have become research hotspots. The aim of this study was to establish a neural network model for MSI prediction and to discern new MSI signature genes.Methods Three CRC GEO datasets (GSE39582, GSE29638, and GSE75315) were used as model training sets, and one TCGA CRC dataset was used as an independent external validation set. Based on the sequencing data and microarray data of the datasets, a neural network prediction model for CRC MSI was established using differential analysis, random forest algorithm, and elastic backpropagation algorithm. Traditional machine learning models for MSI were established using K-nearest neighbor algorithm (KNN) and support vector machine (SVM) algorithm. The prediction ability of the models was evaluated using confusion matrices, receiver operating characteristic (ROC) curves, and the area under the curve (AUC).Results In the training set, a total of 787 cases were included, including 111 cases (14.10%) of microsatellite instability-high (MSI-H) and 676 cases (85.90%) of microsatellite instability-low/microsatellite stability (MSI-L/MSS). In the validation set, 389 cases in the TCGA dataset were finally included, including 67 cases (17.22%) of MSI-H and 322 cases (82.78%) of MSI-L/MSS. One hundred MSI-related genes were identified by differential analysis, including 61 up-regulated genes and 39 down-regulated genes. By combining differential analysis and random forest algorithm, the top 30 most significant MSI-related genes were screened out. Based on the expression matrix of the MSI-related genes, a neural network prediction model was established using 23 gene expression matrices. The model showed accurate prediction ability in both the training set (sensitivity: 0.993, specificity: 0.973, diagnostic coincidence rate: 0.990, AUC: 0.991) and the validation set (sensitivity: 0.950, specificity: 0.828, diagnostic coincidence rate: 0.933, AUC 0.922). Moreover, compared with other machine learning models, the neural network model demonstrated more accurate prediction ability in predicting MSI.Conclusion The neural network prediction model combined with tissue deep sequencing can assist clinicians in diagnosing the MSI status of CRC, and provide references and decision-making basis for the selection of tumor immunotherapy schemes. At the same time, the identified MSI signature genes provide clues and directions for in-depth research on related functions and mechanisms.