A total of 71 patients diagnosed with aCRC at the Peking Union Medical College Hospital (Beijing, China) between June 2013 and February 2018 were enrolled in the present study. Tissue samples were obtained from patients following radical CRC resection or biopsy during colonoscopy. All patients provided written informed consent. The inclusion criteria were as follows: i) All patients were histologically diagnosed with aCRC and cancer stage was evaluated according to the American Joint Committee on Cancer, 7th edition (20); ii) all patients received oxaliplatin plus 5-FU regimens in accordance with the National Comprehensive Cancer Network guideline (21), including mFOLFOX6, which consisted of 85 mg/m² oxaliplatin on day 1, 400 mg/m² leucovorin on day 1, 400 mg/m² intravenous (IV) bolus 5-FU on day 1, and then 1,200 mg/m²/day for 48 h by continuous IV infusion, repeating every 2 weeks and evaluated every 4 cycles using the Response Evaluation Criteria in Solid Tumors (RECIST 1.1) (22), or XELOX, which consists of 130 mg/m² IV oxaliplatin on day 1 and 1,000 mg/m² oral capecitabine twice daily for 14 days, repeating every 3 weeks and evaluated every 2–3 cycles using RECIST; iii) non-responders to chemotherapy (NRCs) included patients with stage IV CRC who received first-line chemotherapy and were identified as having progressive disease when evaluated the first time (after 4 cycles of mFOLFOX6 or 2–3 cycles of XELOX), and patients with stage III CRC who received adjuvant chemotherapy following surgery and who relapsed within 6 months. Chemotherapy responders (CRs) referred to patients with stage IV CRC who received first-line chemotherapy and who were evaluated with partial regression or complete regression at most. The chemotherapy responses were evaluated independently in all patients by two oncologists using RECIST. If their conclusions were different, a third oncologist independently evaluated the case, and the three oncologists consulted and obtained the final result together. Patients with CRC who were identified as stable after chemotherapy or as progressive disease following the first evaluation were excluded from the present study.

IHC staining was performed on 4-µm thick sections of paraffin-embedded tissue samples to detect the protein expression levels of α-SMA, p-AKT, p-ERK and survivin. Briefly, sections were deparaffinized with xylene, rehydrated in decreasing gradient of alcohol and immersed in distilled water for 5 min. Sections were then incubated with EDTA antigen retrieval buffer (pH 8.0) in a microwave for 15 min. Sections were treated with 3% hydrogen peroxide for 25 min at room temperature in the dark to block endogenous peroxidase activity, washed with PBS three times for 5 min and blocked with 3% bovine serum albumin at room temperature for 30 min. Sections were incubated with the following primary antibodies at 4°C overnight: Mouse anti-human α-SMA (cat. no. GB13044, 1:1,000; Servicebio), mouse anti-human p-ERK (cat. no. GB13004; 1:1,000; Servicebio), rabbit anti-human p-AKT (cat. no. GB13012-3; 1:100; Servicebio) and rabbit anti-human survivin (cat. no. 10508-1-AP; 1:100; ProteinTech Group, Inc.). Sections were then incubated with goat anti-mouse secondary antibody (cat. no. GB23301; 1:200; Servicebio; Startech) or goat anti-rabbit secondary antibody (cat. no. GB13044; 1:1,000; Servicebio; Startech) at room temperature for 50 min. Sections were developed using 3,3′-diaminobenzidine and counterstained with hematoxylin at room temperature for 5 min. Images were captured under light microscopy.

All slides were independently evaluated by two pathologists who had no knowledge of the clinical information. If their conclusions were different, a third pathologist independently evaluated the case, and the three pathologists discussed and obtained the final result together. Positive α-SMA expression was localized in the cytoplasm of CAFs; positive survivin expression was localized in the cytoplasm of CRC cells; positive p-AKT and p-ERK expression was localized in the nucleus of CRC cells due to AKT and ERK nuclear translocation. The staining intensity of α-SMA expression was visually scored and classified as follows: 0, no staining; 1, light brown staining; and 2, brown staining. The proportion of cells positively stained for α-SMA was graded as follows: 0, no positively stained CAFs/total CAFs in high-power fields (HPFs); 1, positively stained CAFs/total CAFs in HPFs <10%; 2, positively stained CAFs/total CAFs in HPFs between 10 and 50%; and 3, positively stained CAFs/total CAFs in HPFs >50%. The staining intensity of p-AKT, p-ERK and survivin was visually scored and classified as follows: 0, no staining; 1, light brown staining; and 2, brown staining. The proportion of cells positively stained for p-AKT, p-ERK or survivin was graded as follows: 0, no positively stained CRC cells/total CRC cells in HPFs; 1, positively stained CRC cells/total CRC cells in HPFs <10%; 2, positively stained CRC cells/total CRC cells in HPFs between 10 and 50%; 3, positively stained CRC cells/total CRC cells in HPFs >50%. The staining index (SI) of α-SMA, p-AKT, p-ERK and survivin was calculated as follows: SI = staining intensity × proportion of positive cells. The results were scored from 0 to 6. Low expression was defined as an SI of 0 to 2 and high expression was defined as an SI of 3 to 6 (23).

Follow-ups were designed to assess the progression-free survival (PFS), disease-free survival (DFS) and overall survival (OS) of the patients. The latest follow-up of the patients was performed on February 28, 2019.

Expression levels of α-SMA, p-AKT, survivin and p-ERK were used as features for this model and scored as 1 or 2 (low and high expression, respectively). First, the accuracy of these four features as predictive biomarkers for chemotherapy efficiency in patients with aCRC who were identified as NRCs or CRs was analyzed. Subsequently, the two features of highest accuracy were selected, integrated and their accuracy was analyzed. The selection was then increased to three and four features in order to analyze the predictive accuracy.

An artificial neural network is a type of artificial intelligence; it has been recently widely used in numerous fields, including clinical diagnosis, screening and prognosis prediction (24). As PNNs use the advantages of a radial basis network and classical probability density estimation theory, they offer particularly significant advantages in pattern classification. The PNN classification was therefore used in the present study as the model for classification problem.

The MATLAB Neural Network tool cabinet provides the function ‘newpnn’ to create probabilistic neural networks. Assuming that the input vector is P and the target is T, a two-layer neural network named ‘net’ can be created by using net = newpnn (P, T, SPREAD), where P is a matrix of R feature quantities of the input vector of the Q group, R rows and Q columns, T is a target matrix of S rows and Q columns composed of Q group classification vectors, S is a classification number, and S is 2 if it is a class 2 problem. If the sample of the Jth column belongs to the class i, the element of the Jth column of the ith row of the matrix T is 1, and the remaining columns of the row are ~0. SPREAD is the spreading coefficient of the Gaussian radial basis function. The default value is 0.1. If the SPREAD value is ~0, the created probabilistic neural network can be used as the nearest neighbor classifier.

The MATLAB tool cabinet provides the function y=sim (net, q) by using the probabilistic neural network model created by newpnn, which predicts the target value y of the input vector q using the sim function. This function can therefore be used to verify and check the network model.

The MATLAB tool cabinet also provides the function plotconfusion (y, T), where y is the target output value predicted by the network model and T is the known target value for comparison. In case the target value has only two categories and the function calculates the predictive values as 0 and 1, this function will calculate the number of correct samples and false samples, the accuracy and the error rate.

By using the three functions aforementioned provided by the MATLAB tool cabinet, the classification problem of predicting the target value of 1 (NRCs) or 0 (CRs) based on the four eigenvalues of the provided cases can be accomplished.

The data from the 71 patients were first divided into two parts as follows: 43 patients were used for modeling and the remaining 28 were used for prediction. The data from the 43 patients were divided into two parts by 1, 3, 5 …, 2, 4, 6 …, each of which is 22 lines and 21 lines, and the 22 lines were then further interlaced to obtain two lines of 11, 11 of which were combined with 21 rows into a 32-line dataset (74.4% of the total samples) for modeling, and the remaining 11 patients (25.6% of the total samples) were used for model validation.

The clinical and follow-up data were analyzed using SPSS 24.0 (IBM Corp.). The χ² test and Fisher's exact test were used to assess the association between patient clinical characteristics and the expression of α-SMA, p-AKT, p-ERK and survivin. P<0.05 was considered to indicate a statistically significant difference.

A total of 71 patients with pathologically confirmed aCRC who received oxaliplatin plus 5-FU chemotherapy were enrolled in the present study. Amongst them, 26 patients were defined as NRCs and 45 patients were defined as CRs (Table I). There were 38 men and 33 women, and the median age of these patients was 62 years old. There was no significant difference between pathology type or primary tumor location and the chemotherapy response (P=0.144 and P=0.853, respectively). A flowchart describing the study design is presented in Fig. 1.

The expression of α-SMA, p-AKT, p-ERK and survivin was analyzed by IHC in paraffin-embedded specimens from the patients (Fig. 2). The results detailed in Table I demonstrated that the high expression levels of α-SMA, survivin and p-AKT in patients with CRC were associated with oxaliplatin plus 5-FU resistance (NCRs; P<0.001, P<0.001 and P=0.023, respectively). There was no significant difference between p-ERK expression and chemotherapy response (P=0.227).

Among the 71 patients, there were 61 patients (NRCs:CR ratio, 16:45) with stage IV CRC receiving oxaliplatin and 5-FU as first-line chemotherapy, and 10 patients (NRC:CR ratio, 9:1) with stage III with CRC receiving oxaliplatin and 5-FU as adjuvant chemotherapy. Since stage IV patients with CRC and stage III patients with CRC had different prognoses, the association between PFS, DFS and OS and α-SMA, p-AKT, p-ERK and survivin expression for these patients were separately analyzed using Kaplan-Meier curves.

The results presented in Figs. 3 and 4 demonstrated that in patients with stage IV CRC, the high expression of α-SMA was associated with a significantly reduced PFS time (5.5 vs. 15.0 months; P=0.005; Fig. 3A), and although a shorter OS time was observed (30.7 vs. 35.0 months; P=0.218; Fig. 4A), this difference was not statistically significant. Furthermore, high survivin expression was associated with a significantly reduced PFS time (5.5 vs. 15.0 months; P=0.001; Fig. 3C) and a significantly reduced OS time (15.0 vs. 50.0 months; P<0.001; Fig. 4C). There was no significant association between p-AKT and p-ERK expression and PFS (6.0 vs. 11.0 months and 6.0 vs. 11.0 months; P=0.516 and P=0.466, respectively; Fig. 3B and D) or OS (26.7 vs. 35.0 months and 32.7 vs. 35.0 months; P=0.429 and P=0.446, respectively; Fig. 4B and D) times.

The results presented in Fig. 5 demonstrated that in patients with stage III CRC, there was no significant association between α-SMA, p-AKT, survivin and p-ERK expression and PFS time (6.2 vs. 6.6 months, 5.0 vs. 6.6 months, 6.6 vs. 3.0 months and 5.0 vs. 6.6 months; P=0.949, P=0.158, P=0.427 and P=0.280, respectively; Fig. 5A-D). The results from Fig. 6 also revealed that in patients with stage III CRC, there was no significant association between p-AKT, survivin and p-ERK expression and OS time (15.0 vs. 31.0 months, 24.0 vs. 38.0 months and 12.0 vs. 31.0 months; P=0.207, P=0.264 and P=0.353, respectively; Fig. 6A-C). In patients with stage III CRC, only 1 patient had high α-SMA expression (Fig. 6D) and since this patient was still alive during the study, the association between α-SMA expression and median OS could therefore not be determined.

The error rates of α-SMA, p-AKT, survivin and p-ERK expressions were first analyzed as predictive biomarkers for chemotherapy response in the 43 patients involved in this model. The error rates of individual feature were 28, 35 and 30% for α-SMA, p-AKT and survivin expression, whereas p-ERK expression did not have discriminating ability (Table II) and was not satisfactory for predicting chemotherapy response. Two features with the lowest error rate, α-SMA and survivin expression, were selected from the four features and were used to detect the predictive accuracy of the integrated model. The error rate of the combined model was 26% (Table III). When p-AKT expression was added to this model, the error rate decreased to 21% (Table IV). When p-ERK expression was added to this model, the error rate decreased to 16% and the accuracy was >80% (Table V).

Based on these results, the PNN module of the MATLAB tool cabinet was used to establish the model regarding the classification problems. The data from 32 patients were used for modeling, the data from 11 patients were used for validation and the data from 28 patients were used for prediction. The results of modeling, validation and prediction are presented in Table VI. The results demonstrated that the modeling accuracy involving a single feature was ~70%. When the features number increased to 4, the modeling accuracy was 83.7%, the validation accuracy was 92.9% and the prediction accuracy was 85.7%.