A deep learning model for the prediction of pathogenic POLE mutations and microsatellite instability in colorectal cancer from digital pathology images

A deep learning model for the prediction of pathogenic POLE mutations and microsatellite instability in colorectal cancer from digital pathology images | Research Square window.SnipcartSettings = { analytics: { enabled: false } }; (function() { var accessVector = localStorage.getItem('access_vector') || ''; window.dataLayer = window.dataLayer || []; if (accessVector) { window.dataLayer.push({ user: { profile: { profileInfo: { snid: accessVector } } } }); } })(); (function(w,d,s,l,i){w[l]=w[l]||[];w[l].push({'gtm.start':new Date().getTime(),event:'gtm.js'});var f=d.getElementsByTagName(s)[0],j=d.createElement(s),dl=l!='dataLayer'?'&l='+l:'';j.async=true;j.src='https://www.googletagmanager.com/gtm.js?id='+i+dl;f.parentNode.insertBefore(j,f);})(window,document,'script','dataLayer','GTM-K279D39R'); Browse Preprints In Review Journals COVID-19 Preprints AJE Video Bytes Research Tools Research Promotion AJE Professional Editing AJE Rubriq About Preprint Platform In Review Editorial Policies Our Team Advisory Board Help Center Sign In Submit a Preprint Cite Share Download PDF Article A deep learning model for the prediction of pathogenic POLE mutations and microsatellite instability in colorectal cancer from digital pathology images Ting Xu, Jinze Yu, Zhongwu Li, Luxin Tan, Shuo Li, Haoyi Zhou, and 5 more This is a preprint; it has not been peer reviewed by a journal. https://doi.org/ 10.21203/rs.3.rs-6692980/v1 This work is licensed under a CC BY 4.0 License Status: Under Revision Version 1 posted 9 You are reading this latest preprint version Abstract POLE-mutant colorectal cancers (CRCs) exhibit high tumor mutational burden (TMB) and immunogenicity, yet their clinical detection remains challenging due to cost and complexity. We developed whole-slide image cohorts of POLE-mutant, MSI-H, and MSS&TMB-L CRCs and trained an attention-based deep learning model (CLAM) to identify MSI-H and POLE mutations. POLE-mutant CRCs showed distinct pathological features, including poor differentiation, lymphocytic infiltration, Crohn’s-like reaction, and solid growth patterns. In binary classification, CLAM achieved an AUC of 0.9568 (95% CI: 0.9404–0.9722) in internal validation but dropped to 0.8193 (0.7006–0.9381) externally due to data heterogeneity. To improve generalizability, we introduced cross-domain feature alignment and adversarial training, creating the CDA-CLAM model. In ternary classification, CDA-CLAM achieved macro-average AUCs of 0.9638 (0.9453–0.9823) in cross-validation and 0.9323 (0.8693–0.9932) in independent testing. External validation class-wise AUCs were 0.9674 (POLE), 0.9674 (MSI-H), and 0.9091 (MSS&TMB-L), demonstrating enhanced robustness. Our model leverages interpretable attention maps from H&E-stained slides to predict POLE and MSI-H status in CRC, offering a cost-effective diagnostic tool. Biological sciences/Cancer Health sciences/Medical research Physical sciences/Mathematics and computing deep learning POLE mutation colorectal cancer pathology images Figures Figure 1 Figure 2 Figure 3 Figure 4 Figure 5 Figure 6 What is already known on this topic Both MSI-H and POLE-mutant colorectal cancers exhibit high tumor mutational burden (TMB) and an active immunological microenvironment, serving as critical biomarkers for predicting immunotherapy benefits in colorectal cancer. However, the current testing rates for MSI and POLE mutations are suboptimal. What this study adds We established the largest known cohort of colorectal cancer cases with pathogenic POLE mutations and collected their histopathological images. A multi-class molecular label prediction model using an attention-based deep learning approach could effectively predicts MSI-H and pathogenic POLE mutations from H&E-stained histological slides. Cross-domain feature alignment and adversarial training techniques were able to address the issues of clinical cohort data scarcity and distribution bias and improve performance and generalizability. How this study might affect research, practice or policy Given the robust performance and generalizability of the CDA-CLAM model, our research proposes a potential tool for prescreening POLE mutation and MSI-H in colorectal cancer. 1. Introduction Colorectal cancer (CRC) is the third most common malignancy and the third leading cause of cancer-related mortality worldwide 1 . Immunotherapy, usually known as immune checkpoint inhibitors (ICI), has brought about a landmark change for oncotherapy in recent years. However, most patients with colorectal cancer could not benefit from immunotherapy. Mismatch repair deficiency (dMMR), which occurs in approximately 15% of colorectal cancers, can lead to cumulative genetic mutations, especially frameshift mutations in tumor cells, and thus causes high microsatellite instability (MSI-H) and hypermutation phenotype. 2 , 3 MSI-H is the only clinically approved biomarker predicting response to immunotherapy in CRC. 4 Polymerase epsilon (POLE) is one of the eukaryotic high-fidelity DNA polymerases responsible for DNA replication and repair. 5 Alterations in POLE that impair the DNA editing function can also result in the accumulation of missense mutations and improved immune recognition. Increasing evidence suggests that POLE mutations are qualified to predict the benefit of ICI in solid tumors, including CRC. 6 Identifying MSI and POLE mutation is critical to determine whether to choose ICI treatment. MSI testing with immunohistochemistry or PCR-based methods has already been recommended for all CRC patients in international guidelines. However, the actual testing rate of MSI in clinical practice is still suboptimal, with the overall testing rate in Europe and Asia being only 44% and 35%, respectively, mainly for economic and technical reasons. 7 Moreover, POLE gene sequencing has yet to be clinically available. It was reported that 3%-6·3% of CRC had POLE mutations. 8 However, only POLE mutations in the DNA binding and catalytic site of the exonuclease domain could drive an ultra-mutational phenotype and improve the treatment outcomes of ICI. 9 Thus, next-generation sequencing and bioinformatics are required to identify the pathogenic POLE mutations and verify the high TMB status of the tumors. The extremely low mutation frequency of pathogenic POLE mutation (~ 0·3%) and high detection costs restrict the utilization of POLE mutation detection in clinical practice. 10 Therefore, there is an unmet clinical need to develop a cost-efficient and easily accessible tool to screen out potential MSI and POLE mutant CRC for further confirmatory molecular detection. Hematoxylin and eosin (H&E) histology images contain not only the morphological characteristics of tumor cells but also the composition, distribution, and interaction of mesenchyme cells, which hold vital clues to the underlying molecular biology of tumors. MSI-H is associated with poor differentiation, mucinous adenocarcinoma, tumor-infiltrating lymphocytes, and Crohn ’s-like reaction in CRC. 11 Similarly, POLE mutations in CRC and endometrial can also be characterized by an inflammatory tumor microenvironment, including prominent peritumoral and tumor-infiltrating lymphocytes compared to POLE wild-type and POLE non-proofreading mutant tumors. 12 , 13 Deep learning methods could extract multi-scale image features and have shown great potential in tumor diagnosis, grading, classification, and prognosis prediction. 14 Recently, Coudray et al. trained a deep convolutional neural network on the WSIs that could predict the most commonly mutated driver genes in lung adenocarcinoma with the area under the receiver operating characteristic curve (AUC) value ranging from 0·773 to 0·856. 15 Several deep learning models are further proposed to predict the microsatellite status with the H&E stained WSIs of gastric and colorectal cancer, with an AUC ranging from 0·779 to 0·96. 16 – 18 Therefore, we hypothesize that a WSI-based deep learning model could capture the genotype-phenotype corrections in POLE mutated and MSI-H CRC. However, because of the low prevalence of MSI-H and POLE mutations, it is necessary to collect data from diverse hospitals and research institutions, which inevitably introduces the data distribution shift problem, particularly in the distribution of data categories across different sources. Existing WSI classification deep learning models tend to capture data source-specific features and overfit training data, impairing generalization performance on external validation datasets from different sources. To mitigate model performance degradation or collapse on datasets with distribution biases, we propose a multi-level cross-domain feature alignment method to eliminate variability among data sources and encourage the model to learn more robust features relevant to the underlying pathological patterns. 2. Materials and Methods 2.1. Patient Cohorts and Dataset Collection This study collected two proprietary CRC WSI datasets, including an internal dataset ( POLE ) for model development and an external validation dataset ( POLEex ) for prospective evaluation. Both datasets comprised three CRC cohorts: an MSS&TMB-L, an MSI-H, and a POLE mutant (POLEmut) cohort. The POLEmut cohort enrolled CRC with POLE mutations defined as oncogenic mutations by referring to the POLE functional mutation list on OncoKB [POLE (oncokb.org)]. 10 CRC cases confirmed as microsatellite instability by PCR (polymerase chain reaction) or NGS (next-generation sequencing) testing were included in the MSI-H group. And patients with a MSS status and a low TMB (<20 Mutations/Mb) were classified to the MSS&TMB-L group, as colorectal cancer with a pathogenic POLE mutation could also exhibit a microsatellite stable phenotype. 19 Because of the extremely low frequency of POLE mutations in CRC, we retrospectively reviewed the NGS data of 35108 CRC patients from two hospitals (Peking University Cancer Hospital and Peking University Shougang Hospital) and three domestic genetic databases (brbiotech.com, 3dmedcare.com.cn, and genecast.com.cn) from January 2015 to February 2022. Representative formalin-fixed paraffin-embedded H&E-stained slides from each patient were reviewed and selected by a junior pathologist, with a subsequent quality assessment procedure by a senior pathologist to exclude poorly stained or folded slides. All H&E-stained slides were derived from primary colorectal tumors. Of the 261 cases with pathogenic POLE mutations, archived HE-stained tumor slides were available for 129 patients. Moreover, a total of 142 and 235 HE-stained slides were respectively included in the MSI and MSS&TMB-L cohorts of the internal dataset, respectively. For the external validation of the deep learning model, an external validation dataset ( POLEex ) was prospectively retrieved from Peking University Cancer Hospital between March 2022 and March 2023. The detailed composition of categories and data sources for the datasets is shown in Fig. 1 . In addition, the publicly available Camelyon17 training set, 20 comprising 100 slides from each of five different institutions, was included to validate the deep learning model's performance in distinguishing data sources. More detailed descriptions of the dataset were provided in the appendix. This study was reviewed and approved by the Ethics Committee of the Peking University Cancer Hospital. Informed consent was waived because patients were not directly recruited for this study. 2.2. Model development pipeline The entire framework of our deep learning model, CDA-CLAM, is illustrated in Fig. 2 and includes the following components: WSI preprocessing, model development and adversarial training, model evaluation, and heatmap visualization. WSI Preprocessing The WSI preprocessing stage involved tissue segmentation, stain normalization, patch cropping, and feature extraction. In the tissue segmentation stage, median blurring was first performed on the WSIs. Then, contours and binary masks for tissue regions in WSIs were detected and segmented by Otsu thresholding and a morphological closing operation on the down-sampled WSI images. Stain normalization, was then performed to standardize the color appearance of the WSIs. 21 Subsequently, the stain-normalized tissue regions were cropped into patches of \(\:256\times\:256\) pixels at a magnification of \(\:20\times\:\) and 0·25 MPP(microns-per-pixel), and encoded into 512-dimensional feature vectors using CTransPath. 22 The average number of patches in each slide is 6,440 (ranging from 35 to 18,229). Model Development and Adversarial Training The CDA-CLAM model, developed from the CLAM model 23 incorporated multi-level cross-domain feature alignment and adversarial training. CLAM model is a highly performant attention-based WSI classification model that achieves state-of-the-art performance in the balance of accuracy and efficiency. CLAM contains two variants, CLAM-SB and CLAM-MB. CLAM-SB aggregates patch features into a single global feature representing the entire slide with a patch-wise attention mechanism, making it ideal for binary classification tasks. In contrast, CLAM-MB extracts multiple global features, each corresponding to a different class, allowing selective focus on distinct image characteristics, which is better suited for multi-category classification. In our study, we used CLAM-SB for binary classification (MSS&TMB-L vs. MSI-H & POLEmut) and CLAM-MB for ternary classification (MSS&TMB-L vs. MSI-H vs. POLEmut), ensuring precise and efficient molecular classification of colorectal cancer. Though the CLAM model excels in various WSI classification tasks, it is not equipped to handle distribution biases across different data sources, leading to performance degradation or model collapse in multi-center studies. To overcome this limitation, we proposed the CDA-CLAM model, incorporating multi-level cross-domain feature alignment and adversarial training. Our approach included local-level patch-wise alignment, feature-level prototypical alignment, and global-level slide-wise alignment. These techniques ensured comprehensive alignment of the multi-source features at various granularities. The CDA-CLAM model was trained in an adversarial manner, incorporating a domain discriminator trained simultaneously with a slide classifier. The slide classifier selectively focused on the most informative and domain-invariant features pertinent to the molecular classification task. The domain discriminator identified domain-specific features across various data sources to determine the input features' domain. Detailed descriptions of the cross-domain feature alignment techniques and the adversarial training method were provided in the appendix. This approach effectively eliminated the interference of domain-specific features, enabling robust diagnostic feature extraction and maintaining high diagnostic performance even in the presence of dataset biases. Model Evaluation The model was evaluated by five-fold cross-validation on the internal POLE dataset and external validation on the external POLEex dataset. In the internal five-fold cross-validation, the POLE dataset was randomly split into five equal subsets, or "folds". Each fold served as the validation set once, while the remaining four folds were used as the training set. Consequently, the model underwent five training and validation cycles. The model's performance was averaged over these iterations to provide an overall estimate of its performance on the internal dataset. For external validation, each of the five models trained during the cross-validation process was used to predict the molecular labels of the POLEex dataset, resulting in a total of five prediction results. The final prediction results were obtained through model ensemble, achieved by averaging the predictions of these five models. Heatmap Visualization : The CDA-CLAM model utilized an attention mechanism to aggregate patch features and generate features for entire slides. The attention score assigned to each patch reflected its significance in the final decision-making process: patches with higher attention values were deemed more important for the model's predictions. In the visualization process, these attention scores were mapped to RGB color space, and overlayed onto the slide image to highlight the patches that contributed to the model's decisions. The top five patches with the highest attention scores for each category were then extracted and reviewed by pathologists. For external validation, the attention scores from the five models are averaged to produce the final overall attention scores in heatmap visualizations. Computational Hardware and Software The entire process was implemented using Python 3.8 and Pytorch 1.11. The experiments were conducted on a Nvidia Tesla V100 PCI-E 32GB GPU with CUDA 11.3 and cuDNN 8.2. For WSI processing, we used openside-python 1.1.2. Annotation tasks were performed with the Automated Slide Analysis Platform (ASAP) version 1.9. Statistical analysis was performed with GraphPad Prism 9.0 and R 4.3. 2.3. Evaluation and Statistical Analysis Patient demographics, clinical characteristics, and pathological features among the MSS&TMB-L, the MSI-H, and the POLEmut cohorts were compared using Chi-squared tests. Model performance was evaluated with the Area Under the Receiver Operating Characteristic (ROC) curve (AUROC, or AUC) and accuracy. For the multi-class classification task, the ROC curves were plotted with the one-vs-rest strategy, and the AUC was calculated for each class. The overall ROC curves and AUC and accuracy values for the multi-class classification task and the multi-fold cross-validation were plotted or calculated with the macro-average strategy: ROC curves were plotted for each class or each fold separately, then interpolated to obtain the overall ROC curve. The AUC and accuracy values of each class or each fold were averaged to obtain the overall AUC and accuracy values. The 95% confidence intervals (CI) for accuracy were calculated with the bootstrapping method, and the 95% CI for AUC values and the statistical significance of differences between AUC values were determined using the DeLong test. 24 3. Result 3.1. Patient cohorts and histopathological features For the internal POLE dataset, the clinicopathological characteristics of the three cohorts were summarized in Table 1 . No overlapping cases were found among the three cohorts. Twelve recurrent functional POLE alterations were included in the POLEmut cohort, with p.286R, p.V411L, and p.A456P being the most common hotspot mutations, consistent with previous studies. The median TMB was 252·24 mutations/Mb (ranging from 44·97 to 719 mutations/Mb) for the 79 patients with available TMB information. And all POLE mutant CRC were MSS. Detailed information on mutations and TMB was provided in Supplementary Table 1. The median age of patients in the MSS&TMB-L, MSI-H, and POLEmut cohorts were 58, 56, and 49 years, respectively. The proportion of male patients was 61·7% in the MSS&TMB-L cohort, 64·1% in the MSI-H cohort, and 76·0% in the POLEmut cohort. MSI-H and POLE mutant CRC more frequently originated from the right-side colon compared with MSS&TMB-L CRC (70·4%, 58·7% versus 23·5%, p<0·0001). The majority of MSI-H and POLE mutant CRCs were stage Ⅰ/Ⅱ, while most of the MSS&TMB-L CRCs were locally advanced or metastatic diseases (p<0·0001). Table 1 Clinicopathological characteristics of the internal POLE dataset MSS&TMB-L (n = 235) MSI-H (n = 142) POLEmut (n = 129) P value Age(Average) 56·59 59·49 47·50 <0·0001 Gender 0·0194 Male 145(61·7%) 91(64·1%) 98(76·0%) Female 90(38·3%) 51(35·9%) 31(24·0%) Primary tumor location <0·0001 Right-side colon 55(23·5%) 100(70·4%) 64(58·7%) Left-side colon 98(41·9%) 34(23·9%) 30(27·5%) Rectum 81(34·9%) 8(5·6%) 15(13·8%) Missing 1 0 15(13·8%) Histology <0·0001 Adenocarcinoma 185(78·7%) 85(59·9%) 76(58·9%) Mucinous/ signet-ring cell carcinoma 23(9·8%) 36(25·4%) 26(20·2%) Adenocarcinoma with mucin-producing 27(11·5%) 21(14·8%) 27(20·9%) Tumor differentiation 0·0008 Well/Mediate 203(86·4%) 104(73·2%) 93(72·1%) Poor 32(13·6%) 38(26·8%) 36(27·9%) Stage <0·0001 Ⅰ/Ⅲ 41(18·0%) 87(61·3%) 47(58·8%) Ⅲ 122(53·5%) 49(34·5%) 14(17·5%) Ⅳ 65(28·5%) 6(4·2%) 19(23·8%) Missing 7 0 49 TILs grading 0·0148 0 150(63·8%) 69(48·6%) 66(51.2%) 1 78(33·2%) 65(45·8%) 53(41.1%) 2 7(3·0%) 8(5·6%) 10(7.8%) Crohn’s-like reaction <0·0001 Grade 0 77(42·8%) 37(29·8%) 14(21·9%) Grade 1 96(53·3%) 66(53·2%) 25(39·1%) Grade 2 7(3·9%) 21(16·9%) 25(39·1%) NE 55 18 65 Tumor budding 0·9712 Low 169(97·7%) 121(96·8%) 64(98·5%) Intermediate 3(1·7%) 3(2·4%) 1(1·5%) High 1(0·6%) 1(0·8%) 0(0) NE 62 17 65 Tumour invasive pattern 0·0011 Circumscribed 45(24·7%) 55(44·4%) 26(38·8%) Infiltrative 137(75·3%) 69(55·6%) 41(61·2%) NE 53 18 62 Solid growth (≥ 10%) 0·0008 Yes 18(7·7%) 24(16·9%) 27(20·9%) No 217(92·3%) 118(83·1%) 102(79·1%) An expert pathologist, blinded to the molecular labels, evaluated the histomorphological characteristics of the three cohorts. Compared to MSS&TMB-L CRC, MSI-H and POLE mutant tumors were more likely to be poorly differentiated (13·6% for MSS&TMB-L, 26·8% for MSI-H, and 27·9% for POLEmut, p = 0·0008). Among the POLE mutant and MSI-H CRCs, 20·2% and 25·4% were diagnosed as mucinous or signet-ring cell carcinoma; 20·9% and 14·8% were adenocarcinomas with mucin-production. In the MSS&TMB-L cohort, these pathological types accounted for 9·9% and 11·5%, respectively, significantly lower than the other two groups (p<0·0001). Additionally, there was a remarkable difference in the invasive margin morphology among the three cohorts. Circumscribed margin was most common in MSI-H CRC, while infiltrative margin was most frequently observed in MSS&TMB-L CRC (p = 0·0011). Over 10% solid growth was observed in 16·9% and 20·9% of MSI-H and POLE mutant CRC, respectively, compared to only 7·7% in the MSS&TMB-L cohort (p = 0·0008). Moreover, POLE mutant and MSI-H CRCs were strongly associated with an inflammatory tumor microenvironment, as indicated by the prominent tumor-infiltrating lymphocytes (p = 0·0148) and Crohn’s 's-like reaction (p<0·0001). There was no significant difference in tumor budding among the three groups (p = 0·9712). The clinicopathological information of the POLEex dataset was provided in Supplementary Table 2. 3.2. Data source heterogeneity could impair model performance As previously noted, our internal dataset originates from multiple hospitals and institutions, resulting in data source heterogeneity. To assess the extent to which heterogeneity and joint distribution skewness might impair the performance of our deep learning model in classifying MSI-H and POLE mutations, we curated a smaller but higher-quality BCH-POLE dataset by exclusively selecting slides from the Peking University Cancer Hospital out of the multi-source POLE dataset. Given the limited number of POLE-mutated slides in the BCH-POLE dataset, we conducted only the binary classification task of MSS&TMB-L vs. MSI-H and POLE. We trained two original CLAM models on the POLE dataset and the BCH-POLE dataset, respectively. Internal cross-validation and independent external validation were then employed to compare the performance of models trained on different datasets. As shown in Fig. 3 , the CLAM model trained on the BCH-POLE dataset, exhibited a notable decrease in internal cross-validation performance compared to the model trained on the multi-source POLE dataset (AUC: 0·8228(95% CI: 0·7869-0·8587) versus 0·9568 (95% CI: 0·9404-0·9722), p < 0·0001). However, in the independent external cohort ( POLEex ), different outcomes emerged. The CLAM model trained on the multi-source POLE dataset had a reduced AUC (0·8193, 95% CI: 0·7006–0·9381) compared with the model trained on the BCH-POLE dataset (AUC = 0·8563, 95% CI: 0·7527–0·9599), though this difference was not statistically significant (p = 0·2195). Given the substantial difference in the number of training samples in the two datasets, these results suggested that data heterogeneity and joint distribution bias led to severe overfitting, adversely affecting the model’s generalization ability. Therefore, we hypothesized that training deep learning models with a multi-source dataset could mislead the model to identify domain-specific characteristics, resulting in model overfitting and reduced generalization performance. 3.3. Deep learning consistently predicts data source To substantiate our hypothesis, we trained the CLAM model using our multi-source POLE dataset and a publicly accessible Camelyon17 dataset, focusing on the task of identifying data sources with five-fold validation. Our findings, as illustrated in Fig. 4 , indicated that the deep learning model achieved high accuracy in identifying data sources in both the POLE and Camelyon17 datasets, with AUC of 0·9880 (95% CI: 0·9781–0·9973) and 0·9997 (95% CI: 0·9993–1·0000), respectively. Additionally, the accuracy of deep learning models in determining the data source was not affected by staining inconsistencies among different sources. The AUC values of the CLAM model were similar when the trained on stained-normalized and non-stained-normalized POLE datasets (AUC = 0·9880 (95% CI: 0·9781–0·9973) and 0·9946, (95% CI: 0·9901–0·9989), respectively). These results indicated that the deep learning model is capable of capturing data source-specific image characteristics, and the commonly used stain normalization technique can not eliminate the effect of data source heterogeneity. As described in Section 3.2, data source heterogeneity deteriorates the model’s generalization performance. 3.4. Cross-domain alignment and adversarial training improve model generalization To address the deteriorated model generalization performance caused by data source distribution biases, we developed the CLAM-CDA model with cross-domain alignment and adversarial training. We compared the model performance of the CLAM and CDA-CLAM models (Table 2 ). In the internal validation of the binary classification task (Fig. 5 A), the CDA-CLAM model trained on the POLE dataset achieved an AUC of 0·9543 (95% CI: 0·9376-0·9710), comparable to that of the CLAM model trained on the POLE dataset (AUC = 0·9568, 95% CI: 0·9404-0·9733). Both models significantly outperformed the CLAM model trained with the BCH-POLE dataset (AUC = 0·8517, 95% CI: 0·8185-0·8850). Furthermore, in the external validation performed on the POLEex dataset (Fig. 5 B), the CDA-CLAM demonstrated a significantly higher AUC (0·8676; 95% CI: 0·7687-0·9665) compared to the CLAM model trained on the POLE dataset (0·8193; 95%CI: 0·7006–0·9381) (p = 0·044). Table 2 Performance of the proposed CDA-CLAM and the original CLAM models in the two classification tasks. Task Model Internal Test Set: POLE External Test Set: POLEex AUC Accuracy AUC Accuracy Binary (MSS&TMB-L vs. MSI-H & POLEmut) CLAM (trained on BCH-POLE ) 0·8517 (0·8185-0·8850) 0·7510 (0·6668-0·8349) 0·8519 (0·7401-0·9636) 0·8500 (0·7597-0·9403) CLAM (trained on POLE ) 0·9568 (0·9404-0·9733) 0·7510 (0·6680-0·8334) 0·8193 (0·7006-0·9381) 0·7333 (0·6214-0·8452) CDA-CLAM 0·9543 (0·9376-0·9710) 0·8775 (0·8489-0·9060) 0·8676 (0·7687-0·9665) 0·8333 (0·7390-0·9276) Ternary (MSS&TMB-L vs. MSI-H vs. POLEmut) CLAM (trained on POLE ) 0·9595 (0·9408-0·9782) 0·8735 (0·8457-0·9013) 0·8929 (0·8041-0·9769) 0·8556 (0·7682-0·9429) CDA-CLAM 0·9638 (0·9453-0·9823) 0·8880 (0·8614-0·9147) 0·9323 (0·8693-0·9932) 0·8778 (0·7965-0·9590) Similarly, in the ternary classification task (Fig. 5 C and D ), both the CLAM and CDA-CLAM models achieved excellent accuracy and AUC in the internal cross-validation. In the external validation cohort, the CDA-CLAM model achieved an AUC of 0·9323 (95% CI: 0·8693-0·9932), outperforming the CLAM model, which had an AUC of 0·8929 (95% CI: 0·8041-0·9769). Class-wise AUCs in external validation for the CDA-CLAM model tended to be higher than those of the CLAM model (p = 0·0533, 0·0762 and 0·3244 for the MSS&TMB-L, MSI-H and POLE categories, respectively) (Supplementary Table 5). Detailed performance metrics of the proposed CDA-CLAM in both cross-validation and external validation were provided in the Supplementary materials. The confusion matrix illustrated the concordance and discordance between the true molecular labels and the predicted molecular classes by the CDA-CLAM model (Fig. 5 E, F, G, H). For both internal and external datasets, the ternary CDA-CLAM model achieved higher sensitivity in identifying MSI-H and POLE mutant CRC compared to the binary classifier. In the internal POLE dataset, the binary classifier identified 257 out of 271 (94·8%) MSI-H or POLE mutant CRC cases, whereas the ternary classifier identified 267 out of 271 (98·5%) MSI-H or POLE mutant CRC cases. In the independent external dataset, 7 (21·2%) and 3 (9·1%) MSI-H or POLE mutant cases were misclassified as MSS&TMB-L by the binary and ternary classifiers, respectively. For the ternary classifier, the most common error was misclassifying the MSS&TMB-L subtype as the MSI-H subtype. This error is considered acceptable in clinical practice, as the primary goal for the model is to identify as many MSI-H and POLE mutations as possible, even if it results in a decrease in specificity. Our findings suggested that the proposed CDA-CLAM model effectively utilized the expanded dataset while mitigating the impact of dataset distribution bias, leading to enhanced performance. This aligns with our objective: by leveraging the cross-domain alignment and adversarial training, we can effectively reduce the effect of dataset bias and prevent the model from overfitting to the trivial solution that differentiates data sources within the internal development dataset. Consequently, the model could generalize better to the external validation dataset. 3.5. Interpretability To interpret the morphological-molecular correlation tracked by the CDA-CLAM model, we first focused on the CDA-CLAM model’s high-confidence predictions of POLE mutation (Fig. 6 A) and MSI-H (Fig. 6 B). From each category, we selected the five image sections that garnered the highest attention scores. In POLE mutant tumors, the most informative patches predominantly displayed necrotic tissues accompanied by various types of inflammatory cell infiltration such as neutrophils, eosinophils, mononuclear lymphocytes, plasma cells, and macrophages. Additionally, some POLE mutant tiles exhibited tumor cells with irregularly sized vacuolated nuclei, prominent nucleoli, and solid growth architecture. The most informative patches of MSI-H depicted densely proliferating round or oval tumor cells arranged in an incomplete glandular manner. The tumor cells appeared relatively uniform with vacuolar nuclei and visible nucleoli. This analysis helps connect observable features of tumors with their underlying genetic profiles, aiding in their classification and understanding their biological behavior. 4. Discussion In this study, we demonstrated that the proposed CDA-CLAM model can effectively predict MSI-H status and pathogenic POLE mutations directly from the H&E-stained CRC tumor images. To our knowledge, we assembled the largest pathological dataset of POLE-mutated colorectal cancer and developed the first deep learning model capable of distinguishing POLE mutations from H&E-stained whole slide images in CRC. Through within-cohort cross-validation, our model showed robust performance across multiple clinical and pathologic subgroups (Supplementary Fig. 3). Furthermore, we enhanced the generalization performance by integrating the attention-based model with cross-domain feature alignment and adversarial training techniques. The resulting CDA-CLAM model demonstrated promising performance in an independent external cohort. Determining MSI-H status holds significant clinical importance, as MSI-H is approved as the only indicator of response to ICIs in metastatic CRC and the only molecular biomarker directing the neoadjuvant/adjuvant treatment of locally advanced CRC. 25 – 28 In early-stage CRC, pathogenic POLE mutations are associated with an excellent prognosis, even potentially surpassing that of MSI-H CRC. 13 Furthermore, CRC and endometrial cancers harboring POLE mutations, which lead to DNA repair deficiencies, have shown at least comparable responses to immune checkpoint inhibitors as MSI-H tumors. 9 Therefore, it is natural to raise the idea of expanding the treatment paradigms of MSI-H CRC to CRC with pathogenic POLE mutations in the future. The initial challenge lies in accurately identifying tumors with these specific molecular characteristics. The MSI-H and pathogenic POLE mutation prediction task poses several critical challenges for deep learning model development, including limited training data, biased category and data source distribution. In clinical settings, the number of MSI-H and POLE mutant samples is significantly lower than MSS&TMB-L samples. As reported in previous studies, weakly supervised attention-based deep learning models can achieve data-efficient whole-slide training with limited available samples. 23 , 29 CLAM, 23 which is adopted as the base model in this study, is able to accomplish such a task with relatively low computational requirements. Instead of attempting to classify each individual patch within gigapixel WSIs to differentiate MSS&TMB-L, MSI-H, and POLE mutations, CLAM captures contextual information across the entire slide by globally aggregating patch features with an attention mechanism. Moreover, local-level classification can be further applied to better elucidate the relationship between diagnostic patches and the overall slide classification result. This holistic approach enhances the model's ability to capture the critical contextual image information for molecular classification tasks. Addressing the data scarcity in MSI-H and POLE mutant slides necessitates aggregating data from diverse medical centers, posing an additional challenge to the deep learning models. Data source discrepancies can stem from various factors, including variations in specimen handling, staining techniques, and differences in scanner resolution and magnification across different medical centers, and such data distribution bias is often overlooked in deep learning models. 30 Our experiments have demonstrated that such dataset bias can significantly degrade the model’s generalization performance. To alleviate such a problem, we propose the CDA-CLAM model, which integrates domain generalization techniques with the original CLAM model, including cross-domain feature alignment and adversarial training. This approach aims to handle variability in image characteristics, learn robust image features and enhance generalization performance, even when trained on the biased datasets with relatively low computational requirements. Deep learning methods, while powerful, are often criticized for their black-box nature, where the decision-making process of the model lacks transparency for pathological interpretation. In our study, we address this limitation by overlaying attention score heatmaps on the slides. These heatmaps visually highlight tiles that significantly influence each molecular classification. Previous deep learning models for molecular classification typically employed a two-step strategy: a former classifier for tissue classification and a latter classifier for molecular label determination. Usually, only the tumor tiles with or without mucin tiles were reserved for the second part of the workflow. Contrarily, our attention-based model's heatmaps reveal that influential patches extend beyond tumor and mucin regions to include the invasive margin, necrotic components, and tumor mesenchyme. This observation aligns with pathological principles, where genetic events not only impact tumor cell characteristics but also have a prominent role in shaping the tumor microenvironment. POLE mutant and MSI-H tumors share histopathological similarities such as poor differentiation, nuclear atypia, and inflammatory cell infiltration. Interestingly, our model distinguishes these subtypes, primarily attributing more prominent poor tumor differentiation and dirty necrosis in POLE mutant tumors, as highlighted by the attention score heatmap. These findings demonstrate that our attention-based deep learning model effectively uncovers the interplay between molecular events and pathomorphology, thereby achieving accurate molecular classification based on pathological images. Our study has several limitations. First, despite establishing the largest known POLE mutant mCRC cohort, the dataset is still relatively small. Second, though MSI-H and POLE mutations occur in multiple types of solid tumors, our model was trained exclusively on a CRC cohort, limiting its applicability to other tumor types. Additional training data and future algorithmic enhancements can be made to improve model performance and broaden its applicability beyond CRC. Lastly, although the patch-wise attention-based model demonstrates high performance in our study, the aggregation of patch features through attention mechanisms can be considered a weighted sum of these features, which may not adequately capture the complex contextual information necessary for distinguishing different molecular characteristics. Therefore, Transformer-based models, known for their superior contextual modeling capability, could potentially enhance performance in these tasks and the overall model accuracy. 5. Conclusion In summary, we proposed an attention-based deep learning model capable of extracting interpretable image information from digital H&E-stained tumor slides, and distinguishing POLE mutant and MSI-H colorectal cancer while eliminating the effect of data source heterogeneity. The model's high performance suggests its potential as a prescreening tool for identifying immunotherapy-sensitive CRC in clinical practice. Abbreviations AUC area under the receiver operating characteristic curve CRC Colorectal cancer dMMR Mismatch repair deficiency H&E Hematoxylin and eosin ICI immune checkpoint inhibitor MSI microsatellite instability MSS Microsatellite stable PCR polymerase chain reaction TILs tumor-infiltrating lymphocytes TMB tumor mutation burden WSI Whole slide image Declarations 6. Author contribution T.X. and JZ.Y. wrote the main manuscript text. XC.W., J.L., and JX.L. substantial contributions to the conception OR design of the work. T.X., JZ.Y., and LX.T. contributed to the acquisition of the data sets and finished the analysis. T.X. and L.S. contributed to the interpretation of data. JZ.Y., S.L., and HY.Z. contributed to the creation of the deep learning model used in the work. T.X., JZ.Y, and ZH.W. have drafted the work and substantively revised it. All authors have approved the submitted version. 7. Competing interests All authors declare no competing interests. 8. Data Availability The clinical information and all whole-slide images in this study are not publicly available due to human subject privacy protection. However, if you wish to access our data solely for scientific research purposes, the corresponding author can share the relevant data. 9. Code availability The underlying code for this study is not publicly available, but can be made available to qualified researchers on reasonable request from the corresponding author. 10. Acknowledgments We thank all the participants, pathologists and medical technologists for helping with the collection of the samples for this study. This study was supported by the National Science and Technology Major Project (No. 2022ZD0117800), Beijing Natural Science Foundation (7234357), and the Young Elite Scientists Sponsorship Program by CAST (No. 2023QNRC001). References Siegel RL, Miller KD, Fuchs HE, Jemal A. Cancer Statistics, 2021. CA Cancer J Clin. 2021;71(1):7–33. Shia J. The diversity of tumours with microsatellite instability: molecular mechanisms and impact upon microsatellite instability testing and mismatch repair protein immunohistochemistry. Histopathology. 2021;78(4):485–97. Shinbrot E, Henninger EE, Weinhold N, Covington KR, Göksenin AY, Schultz N, et al. Exonuclease mutations in DNA polymerase epsilon reveal replication strand specific mutation patterns and human origins of replication. Genome Res. 2014;24(11):1740–50. Ganesh K, Stadler ZK, Cercek A, Mendelsohn RB, Shia J, Segal NH, et al. Immunotherapy in colorectal cancer: rationale, challenges and potential. Nat Rev Gastroenterol Hepatol. 2019;16(6):361–75. Church DN, Briggs SE, Palles C, Domingo E, Kearsey SJ, Grimes JM, et al. DNA polymerase ε and δ exonuclease domain mutations in endometrial cancer. Hum Mol Genet. 2013;22(14):2820–8. Ma X, Dong L, Liu X, Ou K, Yang L. POLE/POLD1 mutation and tumor immunotherapy. J Exp Clin Cancer Res. 2022;41(1):216. Jindal K, Mendoza L, Merot JL, Anger C, Pino A. 372P Microsatellite instability testing practices in colorectal cancer patients, in Europe and Asia. Annals of Oncology. 2022;33:S706. Lu W, Huang Z, Wang J, Chen Y, Huang M. 359P Analysis of TMB and tumor microenvironment in Polymerase epsilon (POLE) deficient colorectal cancer. Annals of Oncology. 2022;33:S700. Rousseau B, Bieche I, Pasmant E, Hamzaoui N, Leulliot N, Michon L, et al. PD-1 Blockade in Solid Tumors with Defects in Polymerase Epsilon. Cancer Discov. 2022;12(6):1435–48. Campbell BB, Light N, Fabrizio D, Zatzman M, Fuligni F, de Borja R, et al. Comprehensive Analysis of Hypermutation in Human Cancer. Cell. 2017;171(5):1042-56.e10. Mei WJ, Mi M, Qian J, Xiao N, Yuan Y, Ding PR. Clinicopathological characteristics of high microsatellite instability/mismatch repair-deficient colorectal cancer: A narrative review. Front Immunol. 2022;13:1019582. Van Gool IC, Ubachs JEH, Stelloo E, de Kroon CD, Goeman JJ, Smit V, et al. Blinded histopathological characterisation of POLE exonuclease domain-mutant endometrial cancers: sheep in wolf's clothing. Histopathology. 2018;72(2):248–58. Mo S, Ma X, Li Y, Zhang L, Hou T, Han-Zhang H, et al. Somatic POLE exonuclease domain mutations elicit enhanced intratumoral immune responses in stage II colorectal cancer. J Immunother Cancer. 2020;8(2). Jiang Y, Yang M, Wang S, Li X, Sun Y. Emerging role of deep learning-based artificial intelligence in tumor pathology. Cancer Commun (Lond). 2020;40(4):154–66. Coudray N, Ocampo PS, Sakellaropoulos T, Narula N, Snuderl M, Fenyö D, et al. Classification and mutation prediction from non-small cell lung cancer histopathology images using deep learning. Nat Med. 2018;24(10):1559–67. Yamashita R, Long J, Longacre T, Peng L, Berry G, Martin B, et al. Deep learning model for the prediction of microsatellite instability in colorectal cancer: a diagnostic study. Lancet Oncol. 2021;22(1):132–41. Hinata M, Ushiku T. Detecting immunotherapy-sensitive subtype in gastric cancer using histologic image-based deep learning. Sci Rep. 2021;11(1):22636. Echle A, Grabsch HI, Quirke P, van den Brandt PA, West NP, Hutchins GGA, et al. Clinical-Grade Detection of Microsatellite Instability in Colorectal Tumors by Deep Learning. Gastroenterology. 2020;159(4):1406-16.e11. Comprehensive molecular characterization of human colon and rectal cancer. Nature. 2012;487(7407):330–7. Bandi P, Geessink O, Manson Q, Van Dijk M, Balkenhol M, Hermsen M, et al. From Detection of Individual Metastases to Classification of Lymph Node Status at the Patient Level: The CAMELYON17 Challenge. IEEE Trans Med Imaging. 2019;38(2):550–60. Vahadane A, Peng T, Sethi A, Albarqouni S, Wang L, Baust M, et al. Structure-Preserving Color Normalization and Sparse Stain Separation for Histological Images. IEEE Trans Med Imaging. 2016;35(8):1962–71. Wang X, Yang S, Zhang J, Wang M, Zhang J, Yang W, et al. Transformer-based unsupervised contrastive learning for histopathological image classification. Med Image Anal. 2022;81:102559. Lu MY, Williamson DFK, Chen TY, Chen RJ, Barbieri M, Mahmood F. Data-efficient and weakly supervised computational pathology on whole-slide images. Nat Biomed Eng. 2021;5(6):555–70. Sun X, Xu W. Fast Implementation of DeLong's Algorithm for Comparing the Areas Under Correlated Receiver Operating Characteristic Curves. IEEE Signal Processing Letters. 2014;21(11):1389–93. Chalabi M, Fanchi LF, Dijkstra KK, Van den Berg JG, Aalbers AG, Sikorska K, et al. Neoadjuvant immunotherapy leads to pathological responses in MMR-proficient and MMR-deficient early-stage colon cancers. Nat Med. 2020;26(4):566–76. Hu H, Kang L, Zhang J, Wu Z, Wang H, Huang M, et al. Neoadjuvant PD-1 blockade with toripalimab, with or without celecoxib, in mismatch repair-deficient or microsatellite instability-high, locally advanced, colorectal cancer (PICC): a single-centre, parallel-group, non-comparative, randomised, phase 2 trial. Lancet Gastroenterol Hepatol. 2022;7(1):38–48. Cohen R, Taieb J, Fiskum J, Yothers G, Goldberg R, Yoshino T, et al. Microsatellite Instability in Patients With Stage III Colon Cancer Receiving Fluoropyrimidine With or Without Oxaliplatin: An ACCENT Pooled Analysis of 12 Adjuvant Trials. J Clin Oncol. 2021;39(6):642–51. Baxter NN, Kennedy EB, Bergsland E, Berlin J, George TJ, Gill S, et al. Adjuvant Therapy for Stage II Colon Cancer: ASCO Guideline Update. J Clin Oncol. 2022;40(8):892–910. Schirris Y, Gavves E, Nederlof I, Horlings HM, Teuwen J. DeepSMILE: Contrastive self-supervised pre-training benefits MSI and HRD classification directly from H&E whole-slide images in colorectal and breast cancer. Med Image Anal. 2022;79:102464. Howard FM, Dolezal J, Kochanny S, Schulte J, Chen H, Heij L, et al. The impact of site-specific digital histology signatures on deep learning model accuracy and bias. Nat Commun. 2021;12(1):4423. Additional Declarations No competing interests reported. Supplementary Files SupplementaryMaterials.docx Cite Share Download PDF Status: Under Revision Version 1 posted Editorial decision: Revision requested 17 Aug, 2025 Reviews received at journal 16 Aug, 2025 Reviews received at journal 15 Aug, 2025 Reviewers agreed at journal 01 Aug, 2025 Reviewers agreed at journal 31 Jul, 2025 Reviewers invited by journal 29 Jul, 2025 Editor assigned by journal 29 Jul, 2025 Submission checks completed at journal 07 Jul, 2025 First submitted to journal 18 May, 2025 You are reading this latest preprint version Research Square lets you share your work early, gain feedback from the community, and start making changes to your manuscript prior to peer review in a journal. As a division of Research Square Company, we’re committed to making research communication faster, fairer, and more useful. We do this by developing innovative software and high quality services for the global research community. Our growing team is made up of researchers and industry professionals working together to solve the most critical problems facing scientific publishing. Also discoverable on Platform About Our Team In Review Editorial Policies Advisory Board Help Center Resources Author Services Accessibility API Access RSS feed Manage Cookie Preferences © Research Square 2026 | ISSN 2693-5015 (online) Privacy Policy Terms of Service Do Not Sell My Personal Information {"props":{"pageProps":{"initialData":{"identity":"rs-6692980","acceptedTermsAndConditions":true,"allowDirectSubmit":false,"archivedVersions":[],"articleType":"Article","associatedPublications":[],"authors":[{"id":492613824,"identity":"34154580-afd1-4d70-a256-c7964056fcf6","order_by":0,"name":"Ting Xu","email":"","orcid":"","institution":"Peking University Cancer Hospital and Institute","correspondingAuthor":false,"prefix":"","firstName":"Ting","middleName":"","lastName":"Xu","suffix":""},{"id":492613825,"identity":"47b151a9-f1ec-414b-8995-6e55f6c7b98b","order_by":1,"name":"Jinze Yu","email":"","orcid":"","institution":"Beihang University","correspondingAuthor":false,"prefix":"","firstName":"Jinze","middleName":"","lastName":"Yu","suffix":""},{"id":492613826,"identity":"c220f24a-7e62-4cb1-b519-f0c3f6388e6d","order_by":2,"name":"Zhongwu Li","email":"","orcid":"","institution":"Peking University Cancer Hospital and Institute","correspondingAuthor":false,"prefix":"","firstName":"Zhongwu","middleName":"","lastName":"Li","suffix":""},{"id":492613827,"identity":"779b22cc-faa5-43a7-af84-8e1f9f9399c7","order_by":3,"name":"Luxin Tan","email":"","orcid":"","institution":"Peking University Cancer Hospital and Institute","correspondingAuthor":false,"prefix":"","firstName":"Luxin","middleName":"","lastName":"Tan","suffix":""},{"id":492613828,"identity":"e97ff3c4-e923-4470-80f3-880bbf61c8be","order_by":4,"name":"Shuo Li","email":"","orcid":"","institution":"Beihang University","correspondingAuthor":false,"prefix":"","firstName":"Shuo","middleName":"","lastName":"Li","suffix":""},{"id":492613829,"identity":"5cb18251-1cdb-47d8-b1b3-58b2725d7ab4","order_by":5,"name":"Haoyi Zhou","email":"","orcid":"","institution":"Beihang University","correspondingAuthor":false,"prefix":"","firstName":"Haoyi","middleName":"","lastName":"Zhou","suffix":""},{"id":492613830,"identity":"cf6805e2-6da4-4de1-838f-c890db6ea29e","order_by":6,"name":"Zhenghang Wang","email":"","orcid":"","institution":"Peking University Cancer Hospital and Institute","correspondingAuthor":false,"prefix":"","firstName":"Zhenghang","middleName":"","lastName":"Wang","suffix":""},{"id":492613831,"identity":"b1cd4482-0571-49c5-95d6-8075f0f63e86","order_by":7,"name":"Lin Shen","email":"","orcid":"","institution":"Peking University Cancer Hospital and Institute","correspondingAuthor":false,"prefix":"","firstName":"Lin","middleName":"","lastName":"Shen","suffix":""},{"id":492613832,"identity":"0d7b1972-b008-4bee-9404-497767307cea","order_by":8,"name":"Jianxin Li","email":"","orcid":"","institution":"Beihang University","correspondingAuthor":false,"prefix":"","firstName":"Jianxin","middleName":"","lastName":"Li","suffix":""},{"id":492613833,"identity":"9cd1f398-d01e-412d-980e-332fb92b0dd9","order_by":9,"name":"Xicheng Wang","email":"","orcid":"","institution":"Peking University Cancer Hospital and Institute","correspondingAuthor":false,"prefix":"","firstName":"Xicheng","middleName":"","lastName":"Wang","suffix":""},{"id":492613834,"identity":"a4b73dae-d4eb-49e6-947a-e64e3fcf7d73","order_by":10,"name":"Jian Li","email":"data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAAZAAAAAyAQMAAABI0h/eAAAABlBMVEX///8AAABVwtN+AAAACXBIWXMAAA7EAAAOxAGVKw4bAAAAtElEQVRIiWNgGAWjYDADfmbmww9I0yLZzpZmQJoWg/M8ChLEuUe6+eHDLxV35IwP8zAYMNTYRBN2z5xjxsYyZ54Zmx3mPfCA4VhabgNB99zIYZOWbDucuO0wX4IBY8NhYrX8O1y/uZnHQIJoLZIfGw4nGDATq0VyRpqxMcOxw4YzDgMDOYEYv/BLJD98+KPmsDx//+HDDz7U2BDWAgLMPDBWAjHKQYDxB7EqR8EoGAWjYGQCADpFPWUEPkzcAAAAAElFTkSuQmCC","orcid":"","institution":"Peking University Cancer Hospital and Institute","correspondingAuthor":true,"prefix":"","firstName":"Jian","middleName":"","lastName":"Li","suffix":""}],"badges":[],"createdAt":"2025-05-18 17:08:08","currentVersionCode":1,"declarations":"","doi":"10.21203/rs.3.rs-6692980/v1","doiUrl":"https://doi.org/10.21203/rs.3.rs-6692980/v1","draftVersion":[],"editorialEvents":[],"editorialNote":"","failedWorkflow":false,"files":[{"id":88003935,"identity":"cd50bac6-06e4-48dd-b68c-f9bd28335e58","added_by":"auto","created_at":"2025-07-31 10:34:38","extension":"jpeg","order_by":1,"title":"Figure 1","display":"","copyAsset":false,"role":"figure","size":82396,"visible":true,"origin":"","legend":"\u003cp\u003eDetailed composition of categories and data sources of the datasets. (A) Data source and category distribution of the internal POLE dataset. (B\u0026amp;C) Category distribution of the internal dataset POLE and the external validation dataset POLEex.\u003c/p\u003e","description":"","filename":"floatimage1.jpeg","url":"https://assets-eu.researchsquare.com/files/rs-6692980/v1/b3925d57a1ccd1af4879f8b3.jpeg"},{"id":88005480,"identity":"3c5e45e7-aeeb-4db9-b90d-861b718e0444","added_by":"auto","created_at":"2025-07-31 10:42:38","extension":"jpeg","order_by":2,"title":"Figure 2","display":"","copyAsset":false,"role":"figure","size":186928,"visible":true,"origin":"","legend":"\u003cp\u003eThe entire framework of the proposed CDA-CLAM model\u003c/p\u003e","description":"","filename":"floatimage2.jpeg","url":"https://assets-eu.researchsquare.com/files/rs-6692980/v1/71ca5e4c0af9d91fbc7d1d83.jpeg"},{"id":88003940,"identity":"b883810a-6db6-4849-b3cf-e509b0987fb7","added_by":"auto","created_at":"2025-07-31 10:34:38","extension":"jpeg","order_by":3,"title":"Figure 3","display":"","copyAsset":false,"role":"figure","size":48952,"visible":true,"origin":"","legend":"\u003cp\u003ePerformance of the CLAM model trained on the multi-source POLE dataset and the single-source BCH-POLE sub-dataset. (A) Internal cross-validation of CLAM model trained on BCH-POLE and POLE dataset. (B) External validation of CLAM model trained on BCH-POLE and POLE dataset.\u003c/p\u003e","description":"","filename":"floatimage3.jpeg","url":"https://assets-eu.researchsquare.com/files/rs-6692980/v1/998bd9f43c23c1491c5df91d.jpeg"},{"id":88005796,"identity":"06d6c1a7-d389-45e5-b017-00d3d02e88de","added_by":"auto","created_at":"2025-07-31 10:50:38","extension":"jpeg","order_by":4,"title":"Figure 4","display":"","copyAsset":false,"role":"figure","size":50839,"visible":true,"origin":"","legend":"\u003cp\u003eROC curves of the CLAM model in identifying data sources for different datasets. (A) five-fold validation of the POLE dataset without stain normalization; (B) five-fold validation of the POLE dataset with stain normalization; (C) five-fold validation of the Camelyon17 dataset with stain normalization.\u003c/p\u003e","description":"","filename":"floatimage4.jpeg","url":"https://assets-eu.researchsquare.com/files/rs-6692980/v1/8d083cda77e0b4126b244fe1.jpeg"},{"id":88003944,"identity":"f83df4ec-0b0e-42e3-b34d-ffb4b3d4e90e","added_by":"auto","created_at":"2025-07-31 10:34:38","extension":"jpeg","order_by":5,"title":"Figure 5","display":"","copyAsset":false,"role":"figure","size":89302,"visible":true,"origin":"","legend":"\u003cp\u003eROC curves of the CDA-CLAM model and the original CLAM in internal and external validation for the binary and ternary classification task. (A) The internal validation performance in detecting MSI-H and POLE mutation (binary classification) of the CDA-CLAM model trained with the entire POLE dataset and the CLAM model trained with the entire POLE dataset and the BCH-POLE sub-dataset; (B) The ROC curves of external validation in detecting MSI-H and POLE mutation (binary classification) of the CDA-CLAM model and CLAM model (C) The ROC curves of CDA-CLAM and CLAM for detecting MSS\u0026amp;TMB-L, MSI-H, and POLE mutant cases (ternary classification) in the internal dataset. (D) The ROC curves of CDA-CLAM and CLAM for detecting MSS\u0026amp;TMB-L, MSI-H, and POLE mutant cases (ternary classification) in the external dataset. (E) Confusion matrix of binary classification CDA-CLAM model in cross-validation of internal dataset. (F) Confusion matrix of binary classification CDA-CLAM model in the external dataset. (G) Confusion matrix of ternary classification CDA-CLAM model in cross-validation of internal dataset. (G) Confusion matrix of ternary classification CDA-CLAM model in the external dataset.\u003c/p\u003e","description":"","filename":"floatimage5.jpeg","url":"https://assets-eu.researchsquare.com/files/rs-6692980/v1/1358582dda780b0314b77a4e.jpeg"},{"id":88005483,"identity":"e4695483-2485-4762-8939-aed7fdfff86a","added_by":"auto","created_at":"2025-07-31 10:42:38","extension":"jpeg","order_by":6,"title":"Figure 6","display":"","copyAsset":false,"role":"figure","size":157647,"visible":true,"origin":"","legend":"\u003cp\u003eHigh attention score patches of POLE mutant and MSI-H CRC. (A) Five patches with the highest attention scores from five cases which were predicted as the most likely POLE mutant CRC by the CDA-CLAM model. (B) Five patches with the highest attention scores from five cases which were predicted as the most likely MSI-H CRC by the CDA-CLAM model.\u003c/p\u003e","description":"","filename":"floatimage6.jpeg","url":"https://assets-eu.researchsquare.com/files/rs-6692980/v1/a4e685a45109913a8f5ad111.jpeg"},{"id":88006630,"identity":"e371f5ec-82b8-431e-b681-5efe0bc70364","added_by":"auto","created_at":"2025-07-31 11:06:39","extension":"pdf","order_by":0,"title":"","display":"","copyAsset":false,"role":"manuscript-pdf","size":1618430,"visible":true,"origin":"","legend":"","description":"","filename":"manuscript.pdf","url":"https://assets-eu.researchsquare.com/files/rs-6692980/v1/b274f538-50fb-4577-a991-6e9bf1fa828c.pdf"},{"id":88003945,"identity":"3fb290ac-3050-4e4a-89df-11207564a549","added_by":"auto","created_at":"2025-07-31 10:34:38","extension":"docx","order_by":0,"title":"","display":"","copyAsset":false,"role":"supplement","size":2556740,"visible":true,"origin":"","legend":"","description":"","filename":"SupplementaryMaterials.docx","url":"https://assets-eu.researchsquare.com/files/rs-6692980/v1/add6a9d0a8f6356d6a1ac3d5.docx"}],"financialInterests":"No competing interests reported.","formattedTitle":"A deep learning model for the prediction of pathogenic POLE mutations and microsatellite instability in colorectal cancer from digital pathology images","fulltext":[{"header":"What is already known on this topic","content":"\u003cp\u003eBoth MSI-H and POLE-mutant colorectal cancers exhibit high tumor mutational burden (TMB) and an active immunological microenvironment, serving as critical biomarkers for predicting immunotherapy benefits in colorectal cancer. However, the current testing rates for MSI and POLE mutations are suboptimal.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eWhat this study adds\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eWe established the largest known cohort of colorectal cancer cases with pathogenic POLE mutations and collected their histopathological images. A multi-class molecular label prediction model using an attention-based deep learning approach could effectively predicts MSI-H and pathogenic POLE mutations from H\u0026amp;E-stained histological slides. Cross-domain feature alignment and adversarial training techniques were able to address the issues of clinical cohort data scarcity and distribution bias and improve performance and generalizability.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eHow this study might affect research, practice or policy\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eGiven the robust performance and generalizability of the CDA-CLAM model, our research proposes a potential tool for prescreening POLE mutation and MSI-H in colorectal cancer.\u003c/p\u003e"},{"header":"1. Introduction","content":"\u003cp\u003eColorectal cancer (CRC) is the third most common malignancy and the third leading cause of cancer-related mortality worldwide\u003csup\u003e\u003cspan citationid=\"CR1\" class=\"CitationRef\"\u003e1\u003c/span\u003e\u003c/sup\u003e. Immunotherapy, usually known as immune checkpoint inhibitors (ICI), has brought about a landmark change for oncotherapy in recent years. However, most patients with colorectal cancer could not benefit from immunotherapy. Mismatch repair deficiency (dMMR), which occurs in approximately 15% of colorectal cancers, can lead to cumulative genetic mutations, especially frameshift mutations in tumor cells, and thus causes high microsatellite instability (MSI-H) and hypermutation phenotype.\u003csup\u003e\u003cspan citationid=\"CR2\" class=\"CitationRef\"\u003e2\u003c/span\u003e, \u003cspan citationid=\"CR3\" class=\"CitationRef\"\u003e3\u003c/span\u003e\u003c/sup\u003e MSI-H is the only clinically approved biomarker predicting response to immunotherapy in CRC.\u003csup\u003e\u003cspan citationid=\"CR4\" class=\"CitationRef\"\u003e4\u003c/span\u003e\u003c/sup\u003e Polymerase epsilon (POLE) is one of the eukaryotic high-fidelity DNA polymerases responsible for DNA replication and repair.\u003csup\u003e\u003cspan citationid=\"CR5\" class=\"CitationRef\"\u003e5\u003c/span\u003e\u003c/sup\u003e Alterations in POLE that impair the DNA editing function can also result in the accumulation of missense mutations and improved immune recognition. Increasing evidence suggests that POLE mutations are qualified to predict the benefit of ICI in solid tumors, including CRC.\u003csup\u003e\u003cspan citationid=\"CR6\" class=\"CitationRef\"\u003e6\u003c/span\u003e\u003c/sup\u003e Identifying MSI and POLE mutation is critical to determine whether to choose ICI treatment.\u003c/p\u003e\u003cp\u003e MSI testing with immunohistochemistry or PCR-based methods has already been recommended for all CRC patients in international guidelines. However, the actual testing rate of MSI in clinical practice is still suboptimal, with the overall testing rate in Europe and Asia being only 44% and 35%, respectively, mainly for economic and technical reasons.\u003csup\u003e\u003cspan citationid=\"CR7\" class=\"CitationRef\"\u003e7\u003c/span\u003e\u003c/sup\u003e Moreover, POLE gene sequencing has yet to be clinically available. It was reported that 3%-6\u0026middot;3% of CRC had POLE mutations.\u003csup\u003e\u003cspan citationid=\"CR8\" class=\"CitationRef\"\u003e8\u003c/span\u003e\u003c/sup\u003e However, only POLE mutations in the DNA binding and catalytic site of the exonuclease domain could drive an ultra-mutational phenotype and improve the treatment outcomes of ICI.\u003csup\u003e\u003cspan citationid=\"CR9\" class=\"CitationRef\"\u003e9\u003c/span\u003e\u003c/sup\u003e Thus, next-generation sequencing and bioinformatics are required to identify the pathogenic POLE mutations and verify the high TMB status of the tumors. The extremely low mutation frequency of pathogenic POLE mutation (~\u0026thinsp;0\u0026middot;3%) and high detection costs restrict the utilization of POLE mutation detection in clinical practice.\u003csup\u003e\u003cspan citationid=\"CR10\" class=\"CitationRef\"\u003e10\u003c/span\u003e\u003c/sup\u003e Therefore, there is an unmet clinical need to develop a cost-efficient and easily accessible tool to screen out potential MSI and POLE mutant CRC for further confirmatory molecular detection.\u003c/p\u003e\u003cp\u003eHematoxylin and eosin (H\u0026amp;E) histology images contain not only the morphological characteristics of tumor cells but also the composition, distribution, and interaction of mesenchyme cells, which hold vital clues to the underlying molecular biology of tumors. MSI-H is associated with poor differentiation, mucinous adenocarcinoma, tumor-infiltrating lymphocytes, and Crohn \u0026rsquo;s-like reaction in CRC.\u003csup\u003e\u003cspan citationid=\"CR11\" class=\"CitationRef\"\u003e11\u003c/span\u003e\u003c/sup\u003e Similarly, POLE mutations in CRC and endometrial can also be characterized by an inflammatory tumor microenvironment, including prominent peritumoral and tumor-infiltrating lymphocytes compared to POLE wild-type and POLE non-proofreading mutant tumors.\u003csup\u003e\u003cspan citationid=\"CR12\" class=\"CitationRef\"\u003e12\u003c/span\u003e, \u003cspan citationid=\"CR13\" class=\"CitationRef\"\u003e13\u003c/span\u003e\u003c/sup\u003e Deep learning methods could extract multi-scale image features and have shown great potential in tumor diagnosis, grading, classification, and prognosis prediction.\u003csup\u003e\u003cspan citationid=\"CR14\" class=\"CitationRef\"\u003e14\u003c/span\u003e\u003c/sup\u003e Recently, Coudray et al. trained a deep convolutional neural network on the WSIs that could predict the most commonly mutated driver genes in lung adenocarcinoma with the area under the receiver operating characteristic curve (AUC) value ranging from 0\u0026middot;773 to 0\u0026middot;856.\u003csup\u003e\u003cspan citationid=\"CR15\" class=\"CitationRef\"\u003e15\u003c/span\u003e\u003c/sup\u003e Several deep learning models are further proposed to predict the microsatellite status with the H\u0026amp;E stained WSIs of gastric and colorectal cancer, with an AUC ranging from 0\u0026middot;779 to 0\u0026middot;96.\u003csup\u003e\u003cspan additionalcitationids=\"CR17\" citationid=\"CR16\" class=\"CitationRef\"\u003e16\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR18\" class=\"CitationRef\"\u003e18\u003c/span\u003e\u003c/sup\u003e\u003c/p\u003e\u003cp\u003eTherefore, we hypothesize that a WSI-based deep learning model could capture the genotype-phenotype corrections in POLE mutated and MSI-H CRC. However, because of the low prevalence of MSI-H and POLE mutations, it is necessary to collect data from diverse hospitals and research institutions, which inevitably introduces the data distribution shift problem, particularly in the distribution of data categories across different sources. Existing WSI classification deep learning models tend to capture data source-specific features and overfit training data, impairing generalization performance on external validation datasets from different sources. To mitigate model performance degradation or collapse on datasets with distribution biases, we propose a multi-level cross-domain feature alignment method to eliminate variability among data sources and encourage the model to learn more robust features relevant to the underlying pathological patterns.\u003c/p\u003e"},{"header":"2. Materials and Methods","content":"\u003cdiv id=\"Sec3\" class=\"Section2\"\u003e\u003ch2\u003e2.1. Patient Cohorts and Dataset Collection\u003c/h2\u003e\u003cp\u003eThis study collected two proprietary CRC WSI datasets, including an internal dataset (\u003cem\u003ePOLE\u003c/em\u003e) for model development and an external validation dataset (\u003cem\u003ePOLEex\u003c/em\u003e) for prospective evaluation. Both datasets comprised three CRC cohorts: an MSS\u0026amp;TMB-L, an MSI-H, and a POLE mutant (POLEmut) cohort. The POLEmut cohort enrolled CRC with POLE mutations defined as oncogenic mutations by referring to the POLE functional mutation list on OncoKB [POLE (oncokb.org)].\u003csup\u003e\u003cspan citationid=\"CR10\" class=\"CitationRef\"\u003e10\u003c/span\u003e\u003c/sup\u003e CRC cases confirmed as microsatellite instability by PCR (polymerase chain reaction) or NGS (next-generation sequencing) testing were included in the MSI-H group. And patients with a MSS status and a low TMB (\u0026lt;20 Mutations/Mb) were classified to the MSS\u0026amp;TMB-L group, as colorectal cancer with a pathogenic POLE mutation could also exhibit a microsatellite stable phenotype.\u003csup\u003e\u003cspan citationid=\"CR19\" class=\"CitationRef\"\u003e19\u003c/span\u003e\u003c/sup\u003e\u003c/p\u003e\u003cp\u003eBecause of the extremely low frequency of POLE mutations in CRC, we retrospectively reviewed the NGS data of 35108 CRC patients from two hospitals (Peking University Cancer Hospital and Peking University Shougang Hospital) and three domestic genetic databases (brbiotech.com, 3dmedcare.com.cn, and genecast.com.cn) from January 2015 to February 2022. Representative formalin-fixed paraffin-embedded H\u0026amp;E-stained slides from each patient were reviewed and selected by a junior pathologist, with a subsequent quality assessment procedure by a senior pathologist to exclude poorly stained or folded slides. All H\u0026amp;E-stained slides were derived from primary colorectal tumors. Of the 261 cases with pathogenic POLE mutations, archived HE-stained tumor slides were available for 129 patients. Moreover, a total of 142 and 235 HE-stained slides were respectively included in the MSI and MSS\u0026amp;TMB-L cohorts of the internal dataset, respectively. For the external validation of the deep learning model, an external validation dataset (\u003cem\u003ePOLEex\u003c/em\u003e) was prospectively retrieved from Peking University Cancer Hospital between March 2022 and March 2023. The detailed composition of categories and data sources for the datasets is shown in Fig.\u0026nbsp;\u003cspan refid=\"Fig1\" class=\"InternalRef\"\u003e1\u003c/span\u003e.\u003c/p\u003e\u003cp\u003e\u003c/p\u003e\u003cp\u003eIn addition, the publicly available \u003cem\u003eCamelyon17\u003c/em\u003e training set,\u003csup\u003e\u003cspan citationid=\"CR20\" class=\"CitationRef\"\u003e20\u003c/span\u003e\u003c/sup\u003e comprising 100 slides from each of five different institutions, was included to validate the deep learning model's performance in distinguishing data sources. More detailed descriptions of the dataset were provided in the appendix.\u003c/p\u003e\u003cp\u003e This study was reviewed and approved by the Ethics Committee of the Peking University Cancer Hospital. Informed consent was waived because patients were not directly recruited for this study.\u003c/p\u003e\u003c/div\u003e\u003cdiv id=\"Sec4\" class=\"Section2\"\u003e\u003ch2\u003e2.2. Model development pipeline\u003c/h2\u003e\u003cp\u003eThe entire framework of our deep learning model, CDA-CLAM, is illustrated in Fig.\u0026nbsp;\u003cspan refid=\"Fig2\" class=\"InternalRef\"\u003e2\u003c/span\u003e and includes the following components: WSI preprocessing, model development and adversarial training, model evaluation, and heatmap visualization.\u003c/p\u003e\u003cp\u003e\u003c/p\u003e\u003cp\u003e\u003cstrong\u003eWSI Preprocessing\u003c/strong\u003e\u003cp\u003eThe WSI preprocessing stage involved tissue segmentation, stain normalization, patch cropping, and feature extraction. In the tissue segmentation stage, median blurring was first performed on the WSIs. Then, contours and binary masks for tissue regions in WSIs were detected and segmented by Otsu thresholding and a morphological closing operation on the down-sampled WSI images. Stain normalization, was then performed to standardize the color appearance of the WSIs.\u003csup\u003e\u003cspan citationid=\"CR21\" class=\"CitationRef\"\u003e21\u003c/span\u003e\u003c/sup\u003e Subsequently, the stain-normalized tissue regions were cropped into patches of \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:256\\times\\:256\\)\u003c/span\u003e\u003c/span\u003e pixels at a magnification of \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:20\\times\\:\\)\u003c/span\u003e\u003c/span\u003e and 0\u0026middot;25 MPP(microns-per-pixel), and encoded into 512-dimensional feature vectors using CTransPath.\u003csup\u003e\u003cspan citationid=\"CR22\" class=\"CitationRef\"\u003e22\u003c/span\u003e\u003c/sup\u003e The average number of patches in each slide is 6,440 (ranging from 35 to 18,229).\u003c/p\u003e\u003c/p\u003e\u003cp\u003e\u003cstrong\u003eModel Development and Adversarial Training\u003c/strong\u003e\u003cp\u003eThe CDA-CLAM model, developed from the CLAM model\u003csup\u003e\u003cspan citationid=\"CR23\" class=\"CitationRef\"\u003e23\u003c/span\u003e\u003c/sup\u003e incorporated multi-level cross-domain feature alignment and adversarial training. CLAM model is a highly performant attention-based WSI classification model that achieves state-of-the-art performance in the balance of accuracy and efficiency. CLAM contains two variants, CLAM-SB and CLAM-MB. CLAM-SB aggregates patch features into a single global feature representing the entire slide with a patch-wise attention mechanism, making it ideal for binary classification tasks. In contrast, CLAM-MB extracts multiple global features, each corresponding to a different class, allowing selective focus on distinct image characteristics, which is better suited for multi-category classification. In our study, we used CLAM-SB for binary classification (MSS\u0026amp;TMB-L vs. MSI-H \u0026amp; POLEmut) and CLAM-MB for ternary classification (MSS\u0026amp;TMB-L vs. MSI-H vs. POLEmut), ensuring precise and efficient molecular classification of colorectal cancer.\u003c/p\u003e\u003c/p\u003e\u003cp\u003eThough the CLAM model excels in various WSI classification tasks, it is not equipped to handle distribution biases across different data sources, leading to performance degradation or model collapse in multi-center studies. To overcome this limitation, we proposed the CDA-CLAM model, incorporating multi-level cross-domain feature alignment and adversarial training. Our approach included local-level patch-wise alignment, feature-level prototypical alignment, and global-level slide-wise alignment. These techniques ensured comprehensive alignment of the multi-source features at various granularities.\u003c/p\u003e\u003cp\u003eThe CDA-CLAM model was trained in an adversarial manner, incorporating a domain discriminator trained simultaneously with a slide classifier. The slide classifier selectively focused on the most informative and domain-invariant features pertinent to the molecular classification task. The domain discriminator identified domain-specific features across various data sources to determine the input features' domain. Detailed descriptions of the cross-domain feature alignment techniques and the adversarial training method were provided in the appendix. This approach effectively eliminated the interference of domain-specific features, enabling robust diagnostic feature extraction and maintaining high diagnostic performance even in the presence of dataset biases.\u003c/p\u003e\u003cp\u003e\u003cstrong\u003eModel Evaluation\u003c/strong\u003e\u003cp\u003eThe model was evaluated by five-fold cross-validation on the internal POLE dataset and external validation on the external \u003cem\u003ePOLEex\u003c/em\u003e dataset. In the internal five-fold cross-validation, the \u003cem\u003ePOLE\u003c/em\u003e dataset was randomly split into five equal subsets, or \"folds\". Each fold served as the validation set once, while the remaining four folds were used as the training set. Consequently, the model underwent five training and validation cycles. The model's performance was averaged over these iterations to provide an overall estimate of its performance on the internal dataset. For external validation, each of the five models trained during the cross-validation process was used to predict the molecular labels of the \u003cem\u003ePOLEex\u003c/em\u003e dataset, resulting in a total of five prediction results. The final prediction results were obtained through model ensemble, achieved by averaging the predictions of these five models.\u003c/p\u003e\u003c/p\u003e\u003cp\u003e\u003cb\u003eHeatmap Visualization\u003c/b\u003e: The CDA-CLAM model utilized an attention mechanism to aggregate patch features and generate features for entire slides. The attention score assigned to each patch reflected its significance in the final decision-making process: patches with higher attention values were deemed more important for the model's predictions. In the visualization process, these attention scores were mapped to RGB color space, and overlayed onto the slide image to highlight the patches that contributed to the model's decisions. The top five patches with the highest attention scores for each category were then extracted and reviewed by pathologists. For external validation, the attention scores from the five models are averaged to produce the final overall attention scores in heatmap visualizations.\u003c/p\u003e\u003cp\u003e\u003cstrong\u003eComputational Hardware and Software\u003c/strong\u003e\u003cp\u003eThe entire process was implemented using Python 3.8 and Pytorch 1.11. The experiments were conducted on a Nvidia Tesla V100 PCI-E 32GB GPU with CUDA 11.3 and cuDNN 8.2. For WSI processing, we used openside-python 1.1.2. Annotation tasks were performed with the Automated Slide Analysis Platform (ASAP) version 1.9. Statistical analysis was performed with GraphPad Prism 9.0 and R 4.3.\u003c/p\u003e\u003c/p\u003e\u003c/div\u003e\u003cdiv id=\"Sec5\" class=\"Section2\"\u003e\u003ch2\u003e2.3. Evaluation and Statistical Analysis\u003c/h2\u003e\u003cp\u003ePatient demographics, clinical characteristics, and pathological features among the MSS\u0026amp;TMB-L, the MSI-H, and the POLEmut cohorts were compared using Chi-squared tests. Model performance was evaluated with the Area Under the Receiver Operating Characteristic (ROC) curve (AUROC, or AUC) and accuracy. For the multi-class classification task, the ROC curves were plotted with the one-vs-rest strategy, and the AUC was calculated for each class. The overall ROC curves and AUC and accuracy values for the multi-class classification task and the multi-fold cross-validation were plotted or calculated with the macro-average strategy: ROC curves were plotted for each class or each fold separately, then interpolated to obtain the overall ROC curve. The AUC and accuracy values of each class or each fold were averaged to obtain the overall AUC and accuracy values. The 95% confidence intervals (CI) for accuracy were calculated with the bootstrapping method, and the 95% CI for AUC values and the statistical significance of differences between AUC values were determined using the DeLong test.\u003csup\u003e\u003cspan citationid=\"CR24\" class=\"CitationRef\"\u003e24\u003c/span\u003e\u003c/sup\u003e\u003c/p\u003e\u003c/div\u003e"},{"header":"3. Result","content":"\u003cdiv id=\"Sec7\" class=\"Section2\"\u003e\u003ch2\u003e3.1. Patient cohorts and histopathological features\u003c/h2\u003e\u003cp\u003eFor the internal \u003cem\u003ePOLE\u003c/em\u003e dataset, the clinicopathological characteristics of the three cohorts were summarized in Table\u0026nbsp;\u003cspan refid=\"Tab1\" class=\"InternalRef\"\u003e1\u003c/span\u003e. No overlapping cases were found among the three cohorts. Twelve recurrent functional POLE alterations were included in the POLEmut cohort, with p.286R, p.V411L, and p.A456P being the most common hotspot mutations, consistent with previous studies. The median TMB was 252\u0026middot;24 mutations/Mb (ranging from 44\u0026middot;97 to 719 mutations/Mb) for the 79 patients with available TMB information. And all POLE mutant CRC were MSS. Detailed information on mutations and TMB was provided in Supplementary Table\u0026nbsp;1. The median age of patients in the MSS\u0026amp;TMB-L, MSI-H, and POLEmut cohorts were 58, 56, and 49 years, respectively. The proportion of male patients was 61\u0026middot;7% in the MSS\u0026amp;TMB-L cohort, 64\u0026middot;1% in the MSI-H cohort, and 76\u0026middot;0% in the POLEmut cohort. MSI-H and POLE mutant CRC more frequently originated from the right-side colon compared with MSS\u0026amp;TMB-L CRC (70\u0026middot;4%, 58\u0026middot;7% versus 23\u0026middot;5%, p\u0026lt;0\u0026middot;0001). The majority of MSI-H and POLE mutant CRCs were stage Ⅰ/Ⅱ, while most of the MSS\u0026amp;TMB-L CRCs were locally advanced or metastatic diseases (p\u0026lt;0\u0026middot;0001).\u003c/p\u003e\u003cp\u003e\u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"Yes\" id=\"Tab1\" border=\"1\"\u003e\u003ccaption language=\"En\"\u003e\u003cdiv class=\"CaptionNumber\"\u003eTable 1\u003c/div\u003e\u003cdiv class=\"CaptionContent\"\u003e\u003cp\u003eClinicopathological characteristics of the internal \u003cem\u003ePOLE\u003c/em\u003e dataset\u003c/p\u003e\u003c/div\u003e\u003c/caption\u003e\u003ccolgroup cols=\"5\"\u003e\u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e\u003cdiv align=\"left\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e\u003cdiv align=\"left\" class=\"colspec\" colname=\"c3\" colnum=\"3\"\u003e\u003c/div\u003e\u003cdiv align=\"left\" class=\"colspec\" colname=\"c4\" colnum=\"4\"\u003e\u003c/div\u003e\u003cdiv align=\"left\" class=\"colspec\" colname=\"c5\" colnum=\"5\"\u003e\u003c/div\u003e\u003cthead\u003e\u003ctr\u003e\u003cth align=\"left\" colname=\"c1\"\u003e\u0026nbsp;\u003c/th\u003e\u003cth align=\"left\" colname=\"c2\"\u003e\u003cp\u003eMSS\u0026amp;TMB-L\u003c/p\u003e\u003cp\u003e(n\u0026thinsp;=\u0026thinsp;235)\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c3\"\u003e\u003cp\u003eMSI-H\u003c/p\u003e\u003cp\u003e(n\u0026thinsp;=\u0026thinsp;142)\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c4\"\u003e\u003cp\u003ePOLEmut\u003c/p\u003e\u003cp\u003e(n\u0026thinsp;=\u0026thinsp;129)\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c5\"\u003e\u003cp\u003eP value\u003c/p\u003e\u003c/th\u003e\u003c/tr\u003e\u003c/thead\u003e\u003ctbody\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eAge(Average)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003e56\u0026middot;59\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e59\u0026middot;49\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u003cp\u003e47\u0026middot;50\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c5\"\u003e\u003cp\u003e\u0026lt;0\u0026middot;0001\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eGender\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u0026nbsp;\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u0026nbsp;\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u0026nbsp;\u003c/td\u003e\u003ctd align=\"left\" colname=\"c5\"\u003e\u003cp\u003e0\u0026middot;0194\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eMale\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003e145(61\u0026middot;7%)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e91(64\u0026middot;1%)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u003cp\u003e98(76\u0026middot;0%)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c5\"\u003e\u0026nbsp;\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eFemale\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003e90(38\u0026middot;3%)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e51(35\u0026middot;9%)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u003cp\u003e31(24\u0026middot;0%)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c5\"\u003e\u0026nbsp;\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003ePrimary tumor location\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u0026nbsp;\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u0026nbsp;\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u0026nbsp;\u003c/td\u003e\u003ctd align=\"left\" colname=\"c5\"\u003e\u003cp\u003e\u0026lt;0\u0026middot;0001\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eRight-side colon\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003e55(23\u0026middot;5%)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e100(70\u0026middot;4%)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u003cp\u003e64(58\u0026middot;7%)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c5\"\u003e\u0026nbsp;\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eLeft-side colon\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003e98(41\u0026middot;9%)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e34(23\u0026middot;9%)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u003cp\u003e30(27\u0026middot;5%)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c5\"\u003e\u0026nbsp;\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eRectum\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003e81(34\u0026middot;9%)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e8(5\u0026middot;6%)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u003cp\u003e15(13\u0026middot;8%)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c5\"\u003e\u0026nbsp;\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eMissing\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003e1\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e0\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u003cp\u003e15(13\u0026middot;8%)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c5\"\u003e\u0026nbsp;\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eHistology\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u0026nbsp;\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u0026nbsp;\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u0026nbsp;\u003c/td\u003e\u003ctd align=\"left\" colname=\"c5\"\u003e\u003cp\u003e\u0026lt;0\u0026middot;0001\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eAdenocarcinoma\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003e185(78\u0026middot;7%)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e85(59\u0026middot;9%)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u003cp\u003e76(58\u0026middot;9%)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c5\"\u003e\u0026nbsp;\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eMucinous/ signet-ring cell carcinoma\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003e23(9\u0026middot;8%)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e36(25\u0026middot;4%)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u003cp\u003e26(20\u0026middot;2%)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c5\"\u003e\u0026nbsp;\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eAdenocarcinoma with mucin-producing\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003e27(11\u0026middot;5%)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e21(14\u0026middot;8%)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u003cp\u003e27(20\u0026middot;9%)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c5\"\u003e\u0026nbsp;\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eTumor differentiation\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u0026nbsp;\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u0026nbsp;\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u0026nbsp;\u003c/td\u003e\u003ctd align=\"left\" colname=\"c5\"\u003e\u003cp\u003e0\u0026middot;0008\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eWell/Mediate\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003e203(86\u0026middot;4%)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e104(73\u0026middot;2%)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u003cp\u003e93(72\u0026middot;1%)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c5\"\u003e\u0026nbsp;\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003ePoor\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003e32(13\u0026middot;6%)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e38(26\u0026middot;8%)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u003cp\u003e36(27\u0026middot;9%)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c5\"\u003e\u0026nbsp;\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eStage\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u0026nbsp;\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u0026nbsp;\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u0026nbsp;\u003c/td\u003e\u003ctd align=\"left\" colname=\"c5\"\u003e\u003cp\u003e\u0026lt;0\u0026middot;0001\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eⅠ/Ⅲ\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003e41(18\u0026middot;0%)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e87(61\u0026middot;3%)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u003cp\u003e47(58\u0026middot;8%)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c5\"\u003e\u0026nbsp;\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eⅢ\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003e122(53\u0026middot;5%)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e49(34\u0026middot;5%)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u003cp\u003e14(17\u0026middot;5%)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c5\"\u003e\u0026nbsp;\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eⅣ\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003e65(28\u0026middot;5%)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e6(4\u0026middot;2%)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u003cp\u003e19(23\u0026middot;8%)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c5\"\u003e\u0026nbsp;\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eMissing\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003e7\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e0\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u003cp\u003e49\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c5\"\u003e\u0026nbsp;\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eTILs grading\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u0026nbsp;\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u0026nbsp;\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u0026nbsp;\u003c/td\u003e\u003ctd align=\"left\" colname=\"c5\"\u003e\u003cp\u003e0\u0026middot;0148\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003e0\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003e150(63\u0026middot;8%)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e69(48\u0026middot;6%)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u003cp\u003e66(51.2%)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c5\"\u003e\u0026nbsp;\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003e1\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003e78(33\u0026middot;2%)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e65(45\u0026middot;8%)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u003cp\u003e53(41.1%)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c5\"\u003e\u0026nbsp;\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003e2\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003e7(3\u0026middot;0%)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e8(5\u0026middot;6%)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u003cp\u003e10(7.8%)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c5\"\u003e\u0026nbsp;\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eCrohn\u0026rsquo;s-like reaction\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u0026nbsp;\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u0026nbsp;\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u0026nbsp;\u003c/td\u003e\u003ctd align=\"left\" colname=\"c5\"\u003e\u003cp\u003e\u0026lt;0\u0026middot;0001\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eGrade 0\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003e77(42\u0026middot;8%)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e37(29\u0026middot;8%)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u003cp\u003e14(21\u0026middot;9%)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c5\"\u003e\u0026nbsp;\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eGrade 1\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003e96(53\u0026middot;3%)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e66(53\u0026middot;2%)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u003cp\u003e25(39\u0026middot;1%)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c5\"\u003e\u0026nbsp;\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eGrade 2\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003e7(3\u0026middot;9%)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e21(16\u0026middot;9%)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u003cp\u003e25(39\u0026middot;1%)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c5\"\u003e\u0026nbsp;\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eNE\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003e55\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e18\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u003cp\u003e65\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c5\"\u003e\u0026nbsp;\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eTumor budding\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u0026nbsp;\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u0026nbsp;\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u0026nbsp;\u003c/td\u003e\u003ctd align=\"left\" colname=\"c5\"\u003e\u003cp\u003e0\u0026middot;9712\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eLow\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003e169(97\u0026middot;7%)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e121(96\u0026middot;8%)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u003cp\u003e64(98\u0026middot;5%)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c5\"\u003e\u0026nbsp;\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eIntermediate\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003e3(1\u0026middot;7%)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e3(2\u0026middot;4%)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u003cp\u003e1(1\u0026middot;5%)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c5\"\u003e\u0026nbsp;\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eHigh\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003e1(0\u0026middot;6%)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e1(0\u0026middot;8%)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u003cp\u003e0(0)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c5\"\u003e\u0026nbsp;\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eNE\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003e62\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e17\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u003cp\u003e65\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c5\"\u003e\u0026nbsp;\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eTumour invasive pattern\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u0026nbsp;\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u0026nbsp;\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u0026nbsp;\u003c/td\u003e\u003ctd align=\"left\" colname=\"c5\"\u003e\u003cp\u003e0\u0026middot;0011\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eCircumscribed\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003e45(24\u0026middot;7%)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e55(44\u0026middot;4%)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u003cp\u003e26(38\u0026middot;8%)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c5\"\u003e\u0026nbsp;\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eInfiltrative\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003e137(75\u0026middot;3%)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e69(55\u0026middot;6%)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u003cp\u003e41(61\u0026middot;2%)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c5\"\u003e\u0026nbsp;\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eNE\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003e53\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e18\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u003cp\u003e62\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c5\"\u003e\u0026nbsp;\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eSolid growth (\u0026ge;\u0026thinsp;10%)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u0026nbsp;\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u0026nbsp;\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u0026nbsp;\u003c/td\u003e\u003ctd align=\"left\" colname=\"c5\"\u003e\u003cp\u003e0\u0026middot;0008\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eYes\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003e18(7\u0026middot;7%)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e24(16\u0026middot;9%)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u003cp\u003e27(20\u0026middot;9%)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c5\"\u003e\u0026nbsp;\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eNo\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003e217(92\u0026middot;3%)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e118(83\u0026middot;1%)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u003cp\u003e102(79\u0026middot;1%)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c5\"\u003e\u0026nbsp;\u003c/td\u003e\u003c/tr\u003e\u003c/tbody\u003e\u003c/colgroup\u003e\u003c/table\u003e\u003c/div\u003e\u003c/p\u003e\u003cp\u003eAn expert pathologist, blinded to the molecular labels, evaluated the histomorphological characteristics of the three cohorts. Compared to MSS\u0026amp;TMB-L CRC, MSI-H and POLE mutant tumors were more likely to be poorly differentiated (13\u0026middot;6% for MSS\u0026amp;TMB-L, 26\u0026middot;8% for MSI-H, and 27\u0026middot;9% for POLEmut, p\u0026thinsp;=\u0026thinsp;0\u0026middot;0008). Among the POLE mutant and MSI-H CRCs, 20\u0026middot;2% and 25\u0026middot;4% were diagnosed as mucinous or signet-ring cell carcinoma; 20\u0026middot;9% and 14\u0026middot;8% were adenocarcinomas with mucin-production. In the MSS\u0026amp;TMB-L cohort, these pathological types accounted for 9\u0026middot;9% and 11\u0026middot;5%, respectively, significantly lower than the other two groups (p\u0026lt;0\u0026middot;0001). Additionally, there was a remarkable difference in the invasive margin morphology among the three cohorts. Circumscribed margin was most common in MSI-H CRC, while infiltrative margin was most frequently observed in MSS\u0026amp;TMB-L CRC (p\u0026thinsp;=\u0026thinsp;0\u0026middot;0011). Over 10% solid growth was observed in 16\u0026middot;9% and 20\u0026middot;9% of MSI-H and POLE mutant CRC, respectively, compared to only 7\u0026middot;7% in the MSS\u0026amp;TMB-L cohort (p\u0026thinsp;=\u0026thinsp;0\u0026middot;0008). Moreover, POLE mutant and MSI-H CRCs were strongly associated with an inflammatory tumor microenvironment, as indicated by the prominent tumor-infiltrating lymphocytes (p\u0026thinsp;=\u0026thinsp;0\u0026middot;0148) and Crohn\u0026rsquo;s 's-like reaction (p\u0026lt;0\u0026middot;0001). There was no significant difference in tumor budding among the three groups (p\u0026thinsp;=\u0026thinsp;0\u0026middot;9712). The clinicopathological information of the \u003cem\u003ePOLEex\u003c/em\u003e dataset was provided in Supplementary Table\u0026nbsp;2.\u003c/p\u003e\u003c/div\u003e\u003cdiv id=\"Sec8\" class=\"Section2\"\u003e\u003ch2\u003e3.2. Data source heterogeneity could impair model performance\u003c/h2\u003e\u003cp\u003eAs previously noted, our internal dataset originates from multiple hospitals and institutions, resulting in data source heterogeneity. To assess the extent to which heterogeneity and joint distribution skewness might impair the performance of our deep learning model in classifying MSI-H and POLE mutations, we curated a smaller but higher-quality \u003cem\u003eBCH-POLE\u003c/em\u003e dataset by exclusively selecting slides from the Peking University Cancer Hospital out of the multi-source \u003cem\u003ePOLE\u003c/em\u003e dataset. Given the limited number of POLE-mutated slides in the BCH-POLE dataset, we conducted only the binary classification task of MSS\u0026amp;TMB-L vs. MSI-H and POLE.\u003c/p\u003e\u003cp\u003eWe trained two original CLAM models on the \u003cem\u003ePOLE\u003c/em\u003e dataset and the \u003cem\u003eBCH-POLE\u003c/em\u003e dataset, respectively. Internal cross-validation and independent external validation were then employed to compare the performance of models trained on different datasets. As shown in Fig.\u0026nbsp;\u003cspan refid=\"Fig3\" class=\"InternalRef\"\u003e3\u003c/span\u003e, the CLAM model trained on the \u003cem\u003eBCH-POLE\u003c/em\u003e dataset, exhibited a notable decrease in internal cross-validation performance compared to the model trained on the multi-source \u003cem\u003ePOLE\u003c/em\u003e dataset (AUC: 0\u0026middot;8228(95% CI: 0\u0026middot;7869-0\u0026middot;8587) versus 0\u0026middot;9568 (95% CI: 0\u0026middot;9404-0\u0026middot;9722), p\u0026thinsp;\u0026lt;\u0026thinsp;0\u0026middot;0001). However, in the independent external cohort (\u003cem\u003ePOLEex\u003c/em\u003e), different outcomes emerged. The CLAM model trained on the multi-source \u003cem\u003ePOLE\u003c/em\u003e dataset had a reduced AUC (0\u0026middot;8193, 95% CI: 0\u0026middot;7006\u0026ndash;0\u0026middot;9381) compared with the model trained on the \u003cem\u003eBCH-POLE dataset\u003c/em\u003e (AUC\u0026thinsp;=\u0026thinsp;0\u0026middot;8563, 95% CI: 0\u0026middot;7527\u0026ndash;0\u0026middot;9599), though this difference was not statistically significant (p\u0026thinsp;=\u0026thinsp;0\u0026middot;2195). Given the substantial difference in the number of training samples in the two datasets, these results suggested that data heterogeneity and joint distribution bias led to severe overfitting, adversely affecting the model\u0026rsquo;s generalization ability.\u003c/p\u003e\u003cp\u003e\u003c/p\u003e\u003cp\u003eTherefore, we hypothesized that training deep learning models with a multi-source dataset could mislead the model to identify domain-specific characteristics, resulting in model overfitting and reduced generalization performance.\u003c/p\u003e\u003c/div\u003e\u003cdiv id=\"Sec9\" class=\"Section2\"\u003e\u003ch2\u003e3.3. Deep learning consistently predicts data source\u003c/h2\u003e\u003cp\u003eTo substantiate our hypothesis, we trained the CLAM model using our multi-source \u003cem\u003ePOLE\u003c/em\u003e dataset and a publicly accessible \u003cem\u003eCamelyon17\u003c/em\u003e dataset, focusing on the task of identifying data sources with five-fold validation. Our findings, as illustrated in Fig.\u0026nbsp;\u003cspan refid=\"Fig4\" class=\"InternalRef\"\u003e4\u003c/span\u003e, indicated that the deep learning model achieved high accuracy in identifying data sources in both the \u003cem\u003ePOLE\u003c/em\u003e and \u003cem\u003eCamelyon17\u003c/em\u003e datasets, with AUC of 0\u0026middot;9880 (95% CI: 0\u0026middot;9781\u0026ndash;0\u0026middot;9973) and 0\u0026middot;9997 (95% CI: 0\u0026middot;9993\u0026ndash;1\u0026middot;0000), respectively. Additionally, the accuracy of deep learning models in determining the data source was not affected by staining inconsistencies among different sources. The AUC values of the CLAM model were similar when the trained on stained-normalized and non-stained-normalized \u003cem\u003ePOLE\u003c/em\u003e datasets (AUC\u0026thinsp;=\u0026thinsp;0\u0026middot;9880 (95% CI: 0\u0026middot;9781\u0026ndash;0\u0026middot;9973) and 0\u0026middot;9946, (95% CI: 0\u0026middot;9901\u0026ndash;0\u0026middot;9989), respectively).\u003c/p\u003e\u003cp\u003e\u003c/p\u003e\u003cp\u003eThese results indicated that the deep learning model is capable of capturing data source-specific image characteristics, and the commonly used stain normalization technique can not eliminate the effect of data source heterogeneity. As described in Section 3.2, data source heterogeneity deteriorates the model\u0026rsquo;s generalization performance.\u003c/p\u003e\u003c/div\u003e\u003cdiv id=\"Sec10\" class=\"Section2\"\u003e\u003ch2\u003e3.4. Cross-domain alignment and adversarial training improve model generalization\u003c/h2\u003e\u003cp\u003eTo address the deteriorated model generalization performance caused by data source distribution biases, we developed the CLAM-CDA model with cross-domain alignment and adversarial training. We compared the model performance of the CLAM and CDA-CLAM models (Table\u0026nbsp;\u003cspan refid=\"Tab2\" class=\"InternalRef\"\u003e2\u003c/span\u003e ). In the internal validation of the binary classification task (Fig.\u0026nbsp;\u003cspan refid=\"Fig5\" class=\"InternalRef\"\u003e5\u003c/span\u003eA), the CDA-CLAM model trained on the \u003cem\u003ePOLE\u003c/em\u003e dataset achieved an AUC of 0\u0026middot;9543 (95% CI: 0\u0026middot;9376-0\u0026middot;9710), comparable to that of the CLAM model trained on the \u003cem\u003ePOLE\u003c/em\u003e dataset (AUC\u0026thinsp;=\u0026thinsp;0\u0026middot;9568, 95% CI: 0\u0026middot;9404-0\u0026middot;9733). Both models significantly outperformed the CLAM model trained with the \u003cem\u003eBCH-POLE\u003c/em\u003e dataset (AUC\u0026thinsp;=\u0026thinsp;0\u0026middot;8517, 95% CI: 0\u0026middot;8185-0\u0026middot;8850). Furthermore, in the external validation performed on the POLEex dataset (Fig.\u0026nbsp;\u003cspan refid=\"Fig5\" class=\"InternalRef\"\u003e5\u003c/span\u003eB), the CDA-CLAM demonstrated a significantly higher AUC (0\u0026middot;8676; 95% CI: 0\u0026middot;7687-0\u0026middot;9665) compared to the CLAM model trained on the \u003cem\u003ePOLE\u003c/em\u003e dataset (0\u0026middot;8193; 95%CI: 0\u0026middot;7006\u0026ndash;0\u0026middot;9381) (p\u0026thinsp;=\u0026thinsp;0\u0026middot;044).\u003c/p\u003e\u003cp\u003e\u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"Yes\" id=\"Tab2\" border=\"1\"\u003e\u003ccaption language=\"En\"\u003e\u003cdiv class=\"CaptionNumber\"\u003eTable 2\u003c/div\u003e\u003cdiv class=\"CaptionContent\"\u003e\u003cp\u003ePerformance of the proposed CDA-CLAM and the original CLAM models in the two classification tasks.\u003c/p\u003e\u003c/div\u003e\u003c/caption\u003e\u003ccolgroup cols=\"6\"\u003e\u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e\u003cdiv align=\"left\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e\u003cdiv align=\"left\" class=\"colspec\" colname=\"c3\" colnum=\"3\"\u003e\u003c/div\u003e\u003cdiv align=\"left\" class=\"colspec\" colname=\"c4\" colnum=\"4\"\u003e\u003c/div\u003e\u003cdiv align=\"left\" class=\"colspec\" colname=\"c5\" colnum=\"5\"\u003e\u003c/div\u003e\u003cdiv align=\"left\" class=\"colspec\" colname=\"c6\" colnum=\"6\"\u003e\u003c/div\u003e\u003ctbody\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\" morerows=\"1\" rowspan=\"2\"\u003e\u003cp\u003eTask\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\" morerows=\"1\" rowspan=\"2\"\u003e\u003cp\u003eModel\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colspan=\"2\" nameend=\"c4\" namest=\"c3\"\u003e\u003cp\u003eInternal Test Set: \u003cem\u003ePOLE\u003c/em\u003e\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colspan=\"2\" nameend=\"c6\" namest=\"c5\"\u003e\u003cp\u003eExternal Test Set: \u003cem\u003ePOLEex\u003c/em\u003e\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003eAUC\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u003cp\u003eAccuracy\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c5\"\u003e\u003cp\u003eAUC\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c6\"\u003e\u003cp\u003eAccuracy\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\" morerows=\"2\" rowspan=\"3\"\u003e\u003cp\u003eBinary\u003c/p\u003e\u003cp\u003e(MSS\u0026amp;TMB-L vs.\u003c/p\u003e\u003cp\u003eMSI-H \u0026amp; POLEmut)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003eCLAM\u003c/p\u003e\u003cp\u003e(trained on \u003cem\u003eBCH-POLE\u003c/em\u003e)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e0\u0026middot;8517\u003c/p\u003e\u003cp\u003e(0\u0026middot;8185-0\u0026middot;8850)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u003cp\u003e0\u0026middot;7510\u003c/p\u003e\u003cp\u003e(0\u0026middot;6668-0\u0026middot;8349)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c5\"\u003e\u003cp\u003e0\u0026middot;8519\u003c/p\u003e\u003cp\u003e(0\u0026middot;7401-0\u0026middot;9636)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c6\"\u003e\u003cp\u003e0\u0026middot;8500\u003c/p\u003e\u003cp\u003e(0\u0026middot;7597-0\u0026middot;9403)\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003eCLAM\u003c/p\u003e\u003cp\u003e(trained on \u003cem\u003ePOLE\u003c/em\u003e)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e0\u0026middot;9568\u003c/p\u003e\u003cp\u003e(0\u0026middot;9404-0\u0026middot;9733)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u003cp\u003e0\u0026middot;7510\u003c/p\u003e\u003cp\u003e(0\u0026middot;6680-0\u0026middot;8334)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c5\"\u003e\u003cp\u003e0\u0026middot;8193\u003c/p\u003e\u003cp\u003e(0\u0026middot;7006-0\u0026middot;9381)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c6\"\u003e\u003cp\u003e0\u0026middot;7333\u003c/p\u003e\u003cp\u003e(0\u0026middot;6214-0\u0026middot;8452)\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003eCDA-CLAM\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e0\u0026middot;9543\u003c/p\u003e\u003cp\u003e(0\u0026middot;9376-0\u0026middot;9710)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u003cp\u003e0\u0026middot;8775\u003c/p\u003e\u003cp\u003e(0\u0026middot;8489-0\u0026middot;9060)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c5\"\u003e\u003cp\u003e0\u0026middot;8676\u003c/p\u003e\u003cp\u003e(0\u0026middot;7687-0\u0026middot;9665)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c6\"\u003e\u003cp\u003e0\u0026middot;8333\u003c/p\u003e\u003cp\u003e(0\u0026middot;7390-0\u0026middot;9276)\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\" morerows=\"1\" rowspan=\"2\"\u003e\u003cp\u003eTernary\u003c/p\u003e\u003cp\u003e(MSS\u0026amp;TMB-L vs.\u003c/p\u003e\u003cp\u003eMSI-H vs.\u003c/p\u003e\u003cp\u003ePOLEmut)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003eCLAM\u003c/p\u003e\u003cp\u003e(trained on \u003cem\u003ePOLE\u003c/em\u003e)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e0\u0026middot;9595\u003c/p\u003e\u003cp\u003e(0\u0026middot;9408-0\u0026middot;9782)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u003cp\u003e0\u0026middot;8735\u003c/p\u003e\u003cp\u003e(0\u0026middot;8457-0\u0026middot;9013)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c5\"\u003e\u003cp\u003e0\u0026middot;8929\u003c/p\u003e\u003cp\u003e(0\u0026middot;8041-0\u0026middot;9769)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c6\"\u003e\u003cp\u003e0\u0026middot;8556\u003c/p\u003e\u003cp\u003e(0\u0026middot;7682-0\u0026middot;9429)\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003eCDA-CLAM\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e0\u0026middot;9638\u003c/p\u003e\u003cp\u003e(0\u0026middot;9453-0\u0026middot;9823)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u003cp\u003e0\u0026middot;8880\u003c/p\u003e\u003cp\u003e(0\u0026middot;8614-0\u0026middot;9147)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c5\"\u003e\u003cp\u003e0\u0026middot;9323\u003c/p\u003e\u003cp\u003e(0\u0026middot;8693-0\u0026middot;9932)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c6\"\u003e\u003cp\u003e0\u0026middot;8778\u003c/p\u003e\u003cp\u003e(0\u0026middot;7965-0\u0026middot;9590)\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003c/tbody\u003e\u003c/colgroup\u003e\u003c/table\u003e\u003c/div\u003e\u003c/p\u003e\u003cp\u003e\u003c/p\u003e\u003cp\u003eSimilarly, in the ternary classification task (Fig.\u0026nbsp;\u003cspan refid=\"Fig5\" class=\"InternalRef\"\u003e5\u003c/span\u003eC \u003cb\u003eand D\u003c/b\u003e), both the CLAM and CDA-CLAM models achieved excellent accuracy and AUC in the internal cross-validation. In the external validation cohort, the CDA-CLAM model achieved an AUC of 0\u0026middot;9323 (95% CI: 0\u0026middot;8693-0\u0026middot;9932), outperforming the CLAM model, which had an AUC of 0\u0026middot;8929 (95% CI: 0\u0026middot;8041-0\u0026middot;9769). Class-wise AUCs in external validation for the CDA-CLAM model tended to be higher than those of the CLAM model (p\u0026thinsp;=\u0026thinsp;0\u0026middot;0533, 0\u0026middot;0762 and 0\u0026middot;3244 for the MSS\u0026amp;TMB-L, MSI-H and POLE categories, respectively) (Supplementary Table\u0026nbsp;5). Detailed performance metrics of the proposed CDA-CLAM in both cross-validation and external validation were provided in the Supplementary materials.\u003c/p\u003e\u003cp\u003eThe confusion matrix illustrated the concordance and discordance between the true molecular labels and the predicted molecular classes by the CDA-CLAM model (Fig.\u0026nbsp;\u003cspan refid=\"Fig5\" class=\"InternalRef\"\u003e5\u003c/span\u003eE, F, G, H). For both internal and external datasets, the ternary CDA-CLAM model achieved higher sensitivity in identifying MSI-H and POLE mutant CRC compared to the binary classifier. In the internal \u003cem\u003ePOLE\u003c/em\u003e dataset, the binary classifier identified 257 out of 271 (94\u0026middot;8%) MSI-H or POLE mutant CRC cases, whereas the ternary classifier identified 267 out of 271 (98\u0026middot;5%) MSI-H or POLE mutant CRC cases. In the independent external dataset, 7 (21\u0026middot;2%) and 3 (9\u0026middot;1%) MSI-H or POLE mutant cases were misclassified as MSS\u0026amp;TMB-L by the binary and ternary classifiers, respectively. For the ternary classifier, the most common error was misclassifying the MSS\u0026amp;TMB-L subtype as the MSI-H subtype. This error is considered acceptable in clinical practice, as the primary goal for the model is to identify as many MSI-H and POLE mutations as possible, even if it results in a decrease in specificity.\u003c/p\u003e\u003cp\u003eOur findings suggested that the proposed CDA-CLAM model effectively utilized the expanded dataset while mitigating the impact of dataset distribution bias, leading to enhanced performance. This aligns with our objective: by leveraging the cross-domain alignment and adversarial training, we can effectively reduce the effect of dataset bias and prevent the model from overfitting to the trivial solution that differentiates data sources within the internal development dataset. Consequently, the model could generalize better to the external validation dataset.\u003c/p\u003e\u003c/div\u003e\u003cdiv id=\"Sec11\" class=\"Section2\"\u003e\u003ch2\u003e3.5. Interpretability\u003c/h2\u003e\u003cp\u003eTo interpret the morphological-molecular correlation tracked by the CDA-CLAM model, we first focused on the CDA-CLAM model\u0026rsquo;s high-confidence predictions of POLE mutation (Fig.\u0026nbsp;\u003cspan refid=\"Fig6\" class=\"InternalRef\"\u003e6\u003c/span\u003eA) and MSI-H (Fig.\u0026nbsp;\u003cspan refid=\"Fig6\" class=\"InternalRef\"\u003e6\u003c/span\u003eB). From each category, we selected the five image sections that garnered the highest attention scores.\u003c/p\u003e\u003cp\u003e\u003c/p\u003e\u003cp\u003eIn POLE mutant tumors, the most informative patches predominantly displayed necrotic tissues accompanied by various types of inflammatory cell infiltration such as neutrophils, eosinophils, mononuclear lymphocytes, plasma cells, and macrophages. Additionally, some POLE mutant tiles exhibited tumor cells with irregularly sized vacuolated nuclei, prominent nucleoli, and solid growth architecture. The most informative patches of MSI-H depicted densely proliferating round or oval tumor cells arranged in an incomplete glandular manner. The tumor cells appeared relatively uniform with vacuolar nuclei and visible nucleoli. This analysis helps connect observable features of tumors with their underlying genetic profiles, aiding in their classification and understanding their biological behavior.\u003c/p\u003e\u003c/div\u003e"},{"header":"4. Discussion","content":"\u003cp\u003eIn this study, we demonstrated that the proposed CDA-CLAM model can effectively predict MSI-H status and pathogenic POLE mutations directly from the H\u0026amp;E-stained CRC tumor images. To our knowledge, we assembled the largest pathological dataset of POLE-mutated colorectal cancer and developed the first deep learning model capable of distinguishing POLE mutations from H\u0026amp;E-stained whole slide images in CRC. Through within-cohort cross-validation, our model showed robust performance across multiple clinical and pathologic subgroups (Supplementary Fig.\u0026nbsp;3). Furthermore, we enhanced the generalization performance by integrating the attention-based model with cross-domain feature alignment and adversarial training techniques. The resulting CDA-CLAM model demonstrated promising performance in an independent external cohort.\u003c/p\u003e\u003cp\u003eDetermining MSI-H status holds significant clinical importance, as MSI-H is approved as the only indicator of response to ICIs in metastatic CRC and the only molecular biomarker directing the neoadjuvant/adjuvant treatment of locally advanced CRC.\u003csup\u003e\u003cspan additionalcitationids=\"CR26 CR27\" citationid=\"CR25\" class=\"CitationRef\"\u003e25\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR28\" class=\"CitationRef\"\u003e28\u003c/span\u003e\u003c/sup\u003e In early-stage CRC, pathogenic POLE mutations are associated with an excellent prognosis, even potentially surpassing that of MSI-H CRC.\u003csup\u003e\u003cspan citationid=\"CR13\" class=\"CitationRef\"\u003e13\u003c/span\u003e\u003c/sup\u003e Furthermore, CRC and endometrial cancers harboring POLE mutations, which lead to DNA repair deficiencies, have shown at least comparable responses to immune checkpoint inhibitors as MSI-H tumors.\u003csup\u003e\u003cspan citationid=\"CR9\" class=\"CitationRef\"\u003e9\u003c/span\u003e\u003c/sup\u003e Therefore, it is natural to raise the idea of expanding the treatment paradigms of MSI-H CRC to CRC with pathogenic POLE mutations in the future. The initial challenge lies in accurately identifying tumors with these specific molecular characteristics.\u003c/p\u003e\u003cp\u003eThe MSI-H and pathogenic POLE mutation prediction task poses several critical challenges for deep learning model development, including limited training data, biased category and data source distribution. In clinical settings, the number of MSI-H and POLE mutant samples is significantly lower than MSS\u0026amp;TMB-L samples. As reported in previous studies, weakly supervised attention-based deep learning models can achieve data-efficient whole-slide training with limited available samples.\u003csup\u003e\u003cspan citationid=\"CR23\" class=\"CitationRef\"\u003e23\u003c/span\u003e, \u003cspan citationid=\"CR29\" class=\"CitationRef\"\u003e29\u003c/span\u003e\u003c/sup\u003e CLAM,\u003csup\u003e\u003cspan citationid=\"CR23\" class=\"CitationRef\"\u003e23\u003c/span\u003e\u003c/sup\u003e which is adopted as the base model in this study, is able to accomplish such a task with relatively low computational requirements. Instead of attempting to classify each individual patch within gigapixel WSIs to differentiate MSS\u0026amp;TMB-L, MSI-H, and POLE mutations, CLAM captures contextual information across the entire slide by globally aggregating patch features with an attention mechanism. Moreover, local-level classification can be further applied to better elucidate the relationship between diagnostic patches and the overall slide classification result. This holistic approach enhances the model's ability to capture the critical contextual image information for molecular classification tasks. Addressing the data scarcity in MSI-H and POLE mutant slides necessitates aggregating data from diverse medical centers, posing an additional challenge to the deep learning models. Data source discrepancies can stem from various factors, including variations in specimen handling, staining techniques, and differences in scanner resolution and magnification across different medical centers, and such data distribution bias is often overlooked in deep learning models.\u003csup\u003e\u003cspan citationid=\"CR30\" class=\"CitationRef\"\u003e30\u003c/span\u003e\u003c/sup\u003e Our experiments have demonstrated that such dataset bias can significantly degrade the model\u0026rsquo;s generalization performance. To alleviate such a problem, we propose the CDA-CLAM model, which integrates domain generalization techniques with the original CLAM model, including cross-domain feature alignment and adversarial training. This approach aims to handle variability in image characteristics, learn robust image features and enhance generalization performance, even when trained on the biased datasets with relatively low computational requirements.\u003c/p\u003e\u003cp\u003eDeep learning methods, while powerful, are often criticized for their black-box nature, where the decision-making process of the model lacks transparency for pathological interpretation. In our study, we address this limitation by overlaying attention score heatmaps on the slides. These heatmaps visually highlight tiles that significantly influence each molecular classification. Previous deep learning models for molecular classification typically employed a two-step strategy: a former classifier for tissue classification and a latter classifier for molecular label determination. Usually, only the tumor tiles with or without mucin tiles were reserved for the second part of the workflow. Contrarily, our attention-based model's heatmaps reveal that influential patches extend beyond tumor and mucin regions to include the invasive margin, necrotic components, and tumor mesenchyme. This observation aligns with pathological principles, where genetic events not only impact tumor cell characteristics but also have a prominent role in shaping the tumor microenvironment. POLE mutant and MSI-H tumors share histopathological similarities such as poor differentiation, nuclear atypia, and inflammatory cell infiltration. Interestingly, our model distinguishes these subtypes, primarily attributing more prominent poor tumor differentiation and dirty necrosis in POLE mutant tumors, as highlighted by the attention score heatmap. These findings demonstrate that our attention-based deep learning model effectively uncovers the interplay between molecular events and pathomorphology, thereby achieving accurate molecular classification based on pathological images.\u003c/p\u003e\u003cp\u003eOur study has several limitations. First, despite establishing the largest known POLE mutant mCRC cohort, the dataset is still relatively small. Second, though MSI-H and POLE mutations occur in multiple types of solid tumors, our model was trained exclusively on a CRC cohort, limiting its applicability to other tumor types. Additional training data and future algorithmic enhancements can be made to improve model performance and broaden its applicability beyond CRC. Lastly, although the patch-wise attention-based model demonstrates high performance in our study, the aggregation of patch features through attention mechanisms can be considered a weighted sum of these features, which may not adequately capture the complex contextual information necessary for distinguishing different molecular characteristics. Therefore, Transformer-based models, known for their superior contextual modeling capability, could potentially enhance performance in these tasks and the overall model accuracy.\u003c/p\u003e"},{"header":"5. Conclusion","content":"\u003cp\u003eIn summary, we proposed an attention-based deep learning model capable of extracting interpretable image information from digital H\u0026amp;E-stained tumor slides, and distinguishing POLE mutant and MSI-H colorectal cancer while eliminating the effect of data source heterogeneity. The model's high performance suggests its potential as a prescreening tool for identifying immunotherapy-sensitive CRC in clinical practice.\u003c/p\u003e"},{"header":"Abbreviations","content":"\u003ctable border=\"1\" cellspacing=\"0\" cellpadding=\"0\"\u003e\n \u003ctbody\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eAUC\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003earea under the receiver operating characteristic curve\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eCRC\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eColorectal cancer\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003edMMR\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eMismatch repair deficiency\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eH\u0026amp;E\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eHematoxylin and eosin\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eICI\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eimmune checkpoint inhibitor\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eMSI\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003emicrosatellite instability\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eMSS\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eMicrosatellite stable\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003ePCR\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003epolymerase chain reaction\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eTILs\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003etumor-infiltrating lymphocytes\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eTMB\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003etumor mutation burden\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eWSI\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eWhole slide image\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003c/tbody\u003e\n\u003c/table\u003e"},{"header":"Declarations","content":"\u003ch2\u003e6.\u0026nbsp;Author contribution\u003c/h2\u003e\n\u003cp\u003eT.X. and JZ.Y. wrote the main manuscript text. \u0026nbsp;XC.W., J.L., and JX.L. substantial contributions to the conception OR design of the work. T.X., \u0026nbsp;JZ.Y., and LX.T. \u0026nbsp;contributed to the acquisition of the data sets and finished the analysis. T.X. and L.S. contributed to the interpretation of data. JZ.Y., S.L., and HY.Z. contributed to the creation of the deep learning model used in the work. T.X., JZ.Y, and ZH.W. have drafted the work and substantively revised it. All authors have approved the submitted version.\u003c/p\u003e\n\u003ch2\u003e7.\u0026nbsp;Competing interests\u003c/h2\u003e\n\u003cp\u003eAll authors declare no competing interests.\u003c/p\u003e\n\u003ch2\u003e8.\u0026nbsp;Data Availability\u003c/h2\u003e\n\u003cp\u003eThe clinical information and all whole-slide images in this study are not publicly available due to human subject privacy protection. However, if you wish to access our data solely for scientific research purposes, the corresponding author can share the relevant data.\u003c/p\u003e\n\u003ch2\u003e9.\u0026nbsp; Code availability\u003c/h2\u003e\n\u003cp\u003eThe underlying code for this study is not publicly available, but can be made available to qualified researchers on reasonable request from the corresponding author.\u003c/p\u003e\n\u003ch2\u003e10. Acknowledgments\u003c/h2\u003e\n\u003cp\u003eWe thank all the participants, pathologists and medical technologists for helping with the collection of the samples for this study. This study was supported by the National Science and Technology Major Project (No. 2022ZD0117800), Beijing Natural Science Foundation (7234357), and the Young Elite Scientists Sponsorship Program by CAST (No. 2023QNRC001).\u003c/p\u003e"},{"header":"References","content":"\u003col\u003e\u003cli\u003e\u003cspan\u003eSiegel RL, Miller KD, Fuchs HE, Jemal A. Cancer Statistics, 2021. CA Cancer J Clin. 2021;71(1):7\u0026ndash;33.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eShia J. The diversity of tumours with microsatellite instability: molecular mechanisms and impact upon microsatellite instability testing and mismatch repair protein immunohistochemistry. Histopathology. 2021;78(4):485\u0026ndash;97.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eShinbrot E, Henninger EE, Weinhold N, Covington KR, G\u0026ouml;ksenin AY, Schultz N, et al. Exonuclease mutations in DNA polymerase epsilon reveal replication strand specific mutation patterns and human origins of replication. Genome Res. 2014;24(11):1740\u0026ndash;50.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eGanesh K, Stadler ZK, Cercek A, Mendelsohn RB, Shia J, Segal NH, et al. Immunotherapy in colorectal cancer: rationale, challenges and potential. Nat Rev Gastroenterol Hepatol. 2019;16(6):361\u0026ndash;75.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eChurch DN, Briggs SE, Palles C, Domingo E, Kearsey SJ, Grimes JM, et al. DNA polymerase ε and δ exonuclease domain mutations in endometrial cancer. Hum Mol Genet. 2013;22(14):2820\u0026ndash;8.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eMa X, Dong L, Liu X, Ou K, Yang L. POLE/POLD1 mutation and tumor immunotherapy. J Exp Clin Cancer Res. 2022;41(1):216.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eJindal K, Mendoza L, Merot JL, Anger C, Pino A. 372P Microsatellite instability testing practices in colorectal cancer patients, in Europe and Asia. Annals of Oncology. 2022;33:S706.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eLu W, Huang Z, Wang J, Chen Y, Huang M. 359P Analysis of TMB and tumor microenvironment in Polymerase epsilon (POLE) deficient colorectal cancer. Annals of Oncology. 2022;33:S700.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eRousseau B, Bieche I, Pasmant E, Hamzaoui N, Leulliot N, Michon L, et al. PD-1 Blockade in Solid Tumors with Defects in Polymerase Epsilon. Cancer Discov. 2022;12(6):1435\u0026ndash;48.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eCampbell BB, Light N, Fabrizio D, Zatzman M, Fuligni F, de Borja R, et al. Comprehensive Analysis of Hypermutation in Human Cancer. Cell. 2017;171(5):1042-56.e10.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eMei WJ, Mi M, Qian J, Xiao N, Yuan Y, Ding PR. Clinicopathological characteristics of high microsatellite instability/mismatch repair-deficient colorectal cancer: A narrative review. Front Immunol. 2022;13:1019582.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eVan Gool IC, Ubachs JEH, Stelloo E, de Kroon CD, Goeman JJ, Smit V, et al. Blinded histopathological characterisation of POLE exonuclease domain-mutant endometrial cancers: sheep in wolf's clothing. Histopathology. 2018;72(2):248\u0026ndash;58.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eMo S, Ma X, Li Y, Zhang L, Hou T, Han-Zhang H, et al. Somatic POLE exonuclease domain mutations elicit enhanced intratumoral immune responses in stage II colorectal cancer. J Immunother Cancer. 2020;8(2).\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eJiang Y, Yang M, Wang S, Li X, Sun Y. Emerging role of deep learning-based artificial intelligence in tumor pathology. Cancer Commun (Lond). 2020;40(4):154\u0026ndash;66.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eCoudray N, Ocampo PS, Sakellaropoulos T, Narula N, Snuderl M, Feny\u0026ouml; D, et al. Classification and mutation prediction from non-small cell lung cancer histopathology images using deep learning. Nat Med. 2018;24(10):1559\u0026ndash;67.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eYamashita R, Long J, Longacre T, Peng L, Berry G, Martin B, et al. Deep learning model for the prediction of microsatellite instability in colorectal cancer: a diagnostic study. Lancet Oncol. 2021;22(1):132\u0026ndash;41.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eHinata M, Ushiku T. Detecting immunotherapy-sensitive subtype in gastric cancer using histologic image-based deep learning. Sci Rep. 2021;11(1):22636.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eEchle A, Grabsch HI, Quirke P, van den Brandt PA, West NP, Hutchins GGA, et al. Clinical-Grade Detection of Microsatellite Instability in Colorectal Tumors by Deep Learning. Gastroenterology. 2020;159(4):1406-16.e11.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eComprehensive molecular characterization of human colon and rectal cancer. Nature. 2012;487(7407):330\u0026ndash;7.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eBandi P, Geessink O, Manson Q, Van Dijk M, Balkenhol M, Hermsen M, et al. From Detection of Individual Metastases to Classification of Lymph Node Status at the Patient Level: The CAMELYON17 Challenge. IEEE Trans Med Imaging. 2019;38(2):550\u0026ndash;60.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eVahadane A, Peng T, Sethi A, Albarqouni S, Wang L, Baust M, et al. Structure-Preserving Color Normalization and Sparse Stain Separation for Histological Images. IEEE Trans Med Imaging. 2016;35(8):1962\u0026ndash;71.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eWang X, Yang S, Zhang J, Wang M, Zhang J, Yang W, et al. Transformer-based unsupervised contrastive learning for histopathological image classification. Med Image Anal. 2022;81:102559.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eLu MY, Williamson DFK, Chen TY, Chen RJ, Barbieri M, Mahmood F. Data-efficient and weakly supervised computational pathology on whole-slide images. Nat Biomed Eng. 2021;5(6):555\u0026ndash;70.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eSun X, Xu W. Fast Implementation of DeLong's Algorithm for Comparing the Areas Under Correlated Receiver Operating Characteristic Curves. IEEE Signal Processing Letters. 2014;21(11):1389\u0026ndash;93.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eChalabi M, Fanchi LF, Dijkstra KK, Van den Berg JG, Aalbers AG, Sikorska K, et al. Neoadjuvant immunotherapy leads to pathological responses in MMR-proficient and MMR-deficient early-stage colon cancers. Nat Med. 2020;26(4):566\u0026ndash;76.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eHu H, Kang L, Zhang J, Wu Z, Wang H, Huang M, et al. Neoadjuvant PD-1 blockade with toripalimab, with or without celecoxib, in mismatch repair-deficient or microsatellite instability-high, locally advanced, colorectal cancer (PICC): a single-centre, parallel-group, non-comparative, randomised, phase 2 trial. Lancet Gastroenterol Hepatol. 2022;7(1):38\u0026ndash;48.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eCohen R, Taieb J, Fiskum J, Yothers G, Goldberg R, Yoshino T, et al. Microsatellite Instability in Patients With Stage III Colon Cancer Receiving Fluoropyrimidine With or Without Oxaliplatin: An ACCENT Pooled Analysis of 12 Adjuvant Trials. J Clin Oncol. 2021;39(6):642\u0026ndash;51.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eBaxter NN, Kennedy EB, Bergsland E, Berlin J, George TJ, Gill S, et al. Adjuvant Therapy for Stage II Colon Cancer: ASCO Guideline Update. J Clin Oncol. 2022;40(8):892\u0026ndash;910.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eSchirris Y, Gavves E, Nederlof I, Horlings HM, Teuwen J. DeepSMILE: Contrastive self-supervised pre-training benefits MSI and HRD classification directly from H\u0026amp;E whole-slide images in colorectal and breast cancer. Med Image Anal. 2022;79:102464.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eHoward FM, Dolezal J, Kochanny S, Schulte J, Chen H, Heij L, et al. The impact of site-specific digital histology signatures on deep learning model accuracy and bias. Nat Commun. 2021;12(1):4423.\u003c/span\u003e\u003c/li\u003e\u003c/ol\u003e"}],"fulltextSource":"","fullText":"","funders":[],"hasAdminPriorityOnWorkflow":false,"hasManuscriptDocX":true,"hasOptedInToPreprint":true,"hasPassedJournalQc":"","hasAnyPriority":false,"hideJournal":false,"highlight":"","institution":"","isAcceptedByJournal":false,"isAuthorSuppliedPdf":false,"isDeskRejected":"","isHiddenFromSearch":false,"isInQc":false,"isInWorkflow":false,"isPdf":false,"isPdfUpToDate":true,"isWithdrawnOrRetracted":false,"journal":{"display":true,"email":"[email protected]","identity":"npj-precision-oncology","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":false,"externalIdentity":"npjprecisiononcology","sideBox":"Learn more about [npj Precision Oncology](http://www.nature.com/npjprecisiononcology/)","snPcode":"41698","submissionUrl":"https://submission.springernature.com/new-submission/41698/3","title":"npj Precision Oncology","twitterHandle":"","acdcEnabled":true,"dfaEnabled":true,"editorialSystem":"stoa","reportingPortfolio":"NPJ","inReviewEnabled":true,"inReviewRevisionsEnabled":true},"keywords":"deep learning, POLE mutation, colorectal cancer, pathology images","lastPublishedDoi":"10.21203/rs.3.rs-6692980/v1","lastPublishedDoiUrl":"https://doi.org/10.21203/rs.3.rs-6692980/v1","license":{"name":"CC BY 4.0","url":"https://creativecommons.org/licenses/by/4.0/"},"manuscriptAbstract":"\u003cp\u003ePOLE-mutant colorectal cancers (CRCs) exhibit high tumor mutational burden (TMB) and immunogenicity, yet their clinical detection remains challenging due to cost and complexity. We developed whole-slide image cohorts of POLE-mutant, MSI-H, and MSS\u0026amp;TMB-L CRCs and trained an attention-based deep learning model (CLAM) to identify MSI-H and POLE mutations. POLE-mutant CRCs showed distinct pathological features, including poor differentiation, lymphocytic infiltration, Crohn’s-like reaction, and solid growth patterns. In binary classification, CLAM achieved an AUC of 0.9568 (95% CI: 0.9404–0.9722) in internal validation but dropped to 0.8193 (0.7006–0.9381) externally due to data heterogeneity. To improve generalizability, we introduced cross-domain feature alignment and adversarial training, creating the CDA-CLAM model. In ternary classification, CDA-CLAM achieved macro-average AUCs of 0.9638 (0.9453–0.9823) in cross-validation and 0.9323 (0.8693–0.9932) in independent testing. External validation class-wise AUCs were 0.9674 (POLE), 0.9674 (MSI-H), and 0.9091 (MSS\u0026amp;TMB-L), demonstrating enhanced robustness. Our model leverages interpretable attention maps from H\u0026amp;E-stained slides to predict POLE and MSI-H status in CRC, offering a cost-effective diagnostic tool.\u003c/p\u003e","manuscriptTitle":"A deep learning model for the prediction of pathogenic POLE mutations and microsatellite instability in colorectal cancer from digital pathology images","msid":"","msnumber":"","nonDraftVersions":[{"code":1,"date":"2025-07-31 10:34:33","doi":"10.21203/rs.3.rs-6692980/v1","editorialEvents":[{"type":"communityComments","content":0},{"type":"decision","content":"Revision requested","date":"2025-08-17T10:28:24+00:00","index":"","fulltext":""},{"type":"editorInvitedReview","content":"","date":"2025-08-16T18:30:59+00:00","index":"hide","fulltext":""},{"type":"editorInvitedReview","content":"","date":"2025-08-15T15:37:37+00:00","index":"hide","fulltext":""},{"type":"reviewerAgreed","content":"3418719865689336340638367639715715377","date":"2025-08-01T12:18:05+00:00","index":"hide","fulltext":""},{"type":"reviewerAgreed","content":"78475284542018353015913100198600520387","date":"2025-07-31T07:10:32+00:00","index":"hide","fulltext":""},{"type":"reviewersInvited","content":"","date":"2025-07-29T05:48:31+00:00","index":"","fulltext":""},{"type":"editorAssigned","content":"","date":"2025-07-29T05:47:06+00:00","index":"","fulltext":""},{"type":"checksComplete","content":"","date":"2025-07-07T10:04:11+00:00","index":"","fulltext":""},{"type":"submitted","content":"npj Precision Oncology","date":"2025-05-18T16:54:55+00:00","index":"","fulltext":""}],"status":"published","journal":{"display":true,"email":"[email protected]","identity":"npj-precision-oncology","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":false,"externalIdentity":"npjprecisiononcology","sideBox":"Learn more about [npj Precision Oncology](http://www.nature.com/npjprecisiononcology/)","snPcode":"41698","submissionUrl":"https://submission.springernature.com/new-submission/41698/3","title":"npj Precision Oncology","twitterHandle":"","acdcEnabled":true,"dfaEnabled":true,"editorialSystem":"stoa","reportingPortfolio":"NPJ","inReviewEnabled":true,"inReviewRevisionsEnabled":true}}],"origin":"","ownerIdentity":"f63a69d1-3545-47bf-a95d-b3b7100138d2","owner":[],"postedDate":"July 31st, 2025","published":true,"recentEditorialEvents":[],"rejectedJournal":[],"revision":"","amendment":"","status":"in-revision","subjectAreas":[{"id":52307266,"name":"Biological sciences/Cancer"},{"id":52307267,"name":"Health sciences/Medical research"},{"id":52307268,"name":"Physical sciences/Mathematics and computing"}],"tags":[],"updatedAt":"2026-01-03T22:23:15+00:00","versionOfRecord":[],"versionCreatedAt":"2025-07-31 10:34:33","video":"","vorDoi":"","vorDoiUrl":"","workflowStages":[]},"version":"v1","identity":"rs-6692980","journalConfig":"researchsquare"},"__N_SSP":true},"page":"/article/[identity]/[[...version]]","query":{"redirect":"/article/rs-6692980","identity":"rs-6692980","version":["v1"]},"buildId":"8U1c8b4HqxoKbykW_rLl7","isFallback":false,"isExperimentalCompile":false,"dynamicIds":[84888],"gssp":true,"scriptLoader":[]}