KRAS^G12V/A146T mutations are associated with nCRT resistance via enhanced DNA double-strand break repair and support a deep learning prediction framework in LARC

Abstract

Background

Neoadjuvant chemoradiotherapy (nCRT) is the standard treatment for locally advanced rectal cancer (LARC), yet clinically validated biomarkers for predicting response remain lacking. This study aimed to identify candidate molecular events associated with nCRT response and to develop a pretreatment prediction framework integrating genomic and pathological information.

Methods

Whole-exome sequencing (WES) was performed on pretreatment tumors from 67 patients with LARC, and an additional 22 published WES cases were integrated to compare genomic differences between responders (R) and nonresponders (NR). Using histopathological whole-slide images (WSIs; n = 106) and genome-derived features, a weakly supervised, multimodal deep learning fusion model was developed to predict nCRT response. Multiomics profiling was used for exploratory pathway characterization, and functional assays were conducted in colorectal cancer cell lines and mouse models harboring KRAS^G12V or KRAS^A146T.

Results

WES identified 41 response-associated hotspot codon events. KRAS^G12V and KRAS^A146T were detected in the NR group in this cohort and were directionally aligned with poor response, indicating an association with nCRT resistance. Because these events are low-frequency alterations, and the study is a single-center retrospective cohort with limited numbers of carriers, multivariable adjustment for key covariates (including stage and T/N status) was not feasible; thus, these findings should be interpreted as exploratory candidate signals. The multimodal fusion model showed good discrimination within the cohort (AUC = 0.882). Mechanistically, exploratory multiomics analyses and orthogonal functional assays were consistent with KRAS variants being associated with altered DNA damage repair signaling and increased repair capacity, with the functional assays providing the main support for this interpretation.

Conclusions

The proposed genome–pathology fusion model provides a research-oriented framework for pretreatment prediction and risk stratification of nCRT response in LARC. KRAS^G12V and KRAS^A146T are presented as candidate molecular events aligned with poor response, but their independent predictive value and the clinical usability of the model require validation in larger, multicenter prospective cohorts that include external WSI data, together with systematic evaluation of thresholding and calibration before clinical translation.

Graphical abstract

Supplementary Information

The online version contains supplementary material available at 10.1186/s11658-026-00947-3.

Introduction

Colorectal cancer (CRC) is a leading gastrointestinal malignancy worldwide, and rectal cancer accounts for approximately one third of CRC cases [1, 2]. Locally advanced rectal cancer (LARC) is generally defined as stage II–III disease without distant metastasis. Over the past two decades, LARC management has shifted toward multidisciplinary care centered on neoadjuvant chemoradiotherapy (nCRT) and total mesorectal excision (TME), with systemic therapy and, in selected settings, intensified neoadjuvant strategies and immunotherapy increasingly incorporated [3–7]. nCRT can downstage tumors and improve resectability, and randomized trials have demonstrated superior local control for nCRT followed by TME compared with TME alone or postoperative chemoradiotherapy (CRT) [3, 8–11]. Relative to postoperative CRT, nCRT is also associated with better tolerability and compliance and may reduce distant failure in several contexts [10, 12, 13].

Despite these advances, response to nCRT remains highly heterogeneous. Pathological complete response (pCR) is achieved in only ~20% of patients [14, 15]. For nonresponders, nCRT may offer limited benefit while increasing toxicity and delaying definitive surgery; radiotherapy-related bowel edema and tissue fragility can further complicate perioperative management [16, 17]. Therefore, robust pretreatment prediction of nCRT response is essential to enable risk-adapted treatment selection and rational escalation or de-escalation.

Predictive studies have evaluated clinicopathological factors, imaging features, and molecular biomarkers [18–21]. With broader access to sequencing, genomic alterations and expression signatures have supported biomarker discovery by contrasting responders and nonresponders [22]. Candidate markers such as TP53, BRAF, and SMAD4, as well as pathways involving apoptosis and DNA repair, have been implicated, yet clinically deployable models remain limited [23–27]. In parallel, weakly supervised deep learning on whole-slide images (WSI) has emerged as an effective approach for extracting predictive signals from routine histopathology and for integrating multimodal data in CRC [28–32]. These developments provide a rationale for combining genomic and pathological features to improve pretreatment prediction.

Mechanistically, resistance to radiotherapy-based regimens is often linked to enhanced DNA damage repair. Because radiotherapy primarily induces DNA double-strand breaks (DSBs), efficient repair through nonhomologous end joining (NHEJ) and homologous recombination (HR) can reduce tumor sensitivity to nCRT [33–35]. KRAS mutations are common in CRC and are associated with aggressive phenotypes and therapeutic resistance [36]. However, in the distinct clinical context of neoadjuvant chemoradiotherapy for LARC, whether specific KRAS codon variants contribute to treatment response remains unclear, and reported associations between KRAS status and nCRT outcome have been inconsistent across cohorts [37–40]. Notably, distinct KRAS codon variants may exert different biological effects [41, 42], and whether specific variants such as KRAS^G12V and KRAS^A146T drive nCRT resistance by modulating DSB repair remains unclear.

Here, we integrate whole-exome sequencing of biopsies (with matched blood) and weakly supervised deep learning on WSI to predict nCRT response in LARC. We integrate whole-exome sequencing of pretreatment biopsies and weakly supervised deep learning on WSI to predict nCRT response in LARC. We identify KRAS^G12V and KRAS^A146T as resistance-associated variants in a LARC cohort and provide mechanistic evidence from colorectal cancer experimental models linking these mutations to increased DSB repair via NHEJ and HR, supported by exploratory multiomics analyses and orthogonal functional validation. Together, our findings support the further evaluation of a multimodal strategy for pretreatment stratification in LARC and generate mechanistic hypotheses regarding codon-specific KRAS-associated treatment tolerance.

Materials and methods

Sample collection and grouping

A total of 137 patients with LARC treated with nCRT followed by TME were included. The internal cohort comprised 115 patients recruited at the National Cancer Center. The external cohort comprised 22 patients with available WES data reported by Gu et al. and Lin et al. [23, 43] and was used for independent external genomic validation; no WSI data were available for the external cohort.

LARC was defined by pretreatment staging on computed tomography (CT) and/or magnetic resonance imaging (MRI) as T3–4 and/or node-positive (N⁺) disease. Rectal adenocarcinoma was confirmed in all cases by pretreatment colonoscopic biopsy and pathological review. Pretreatment tumor tissue and matched peripheral blood were available from 67 internal patients for WES. In addition, formalin-fixed, paraffin-embedded (FFPE)-processed tumor specimens from 106 internal patients were stained with hematoxylin and eosin (H&E) and digitized to generate a WSI for downstream analysis. As WES and WSI were derived from the same internal cohort but were not available for every patient, analyses were conducted using all eligible cases for each modality; multimodal fusion models were trained and evaluated using patients with the required modalities available.

Pathological response to nCRT was assessed using the Dworak tumor regression grade (TRG) system [44]. Patients with TRG 3–4 were classified as responders (R), whereas those with TRG 0–1 were classified as nonresponders (NR), yielding 65 R and 72 NR patients (Supplementary Table S1). For model development, 99 internal patients were randomly assigned to the training set and 38 to the internal test set (Supplementary Table S2), including 51 R/48 NR in the training set and 14 R/24 NR in the internal test set. The external cohort (n = 22) served as an independent test set. To minimize data leakage, dataset splitting was performed at the patient level; all information from the same patient (including WSI tiles, genomic features, and clinical variables) was restricted to a single dataset. Both the internal and external test sets were used solely for final evaluation and were not used for model training, hyperparameter selection, or classification-threshold determination.

Clinical variables included sex, age, body mass index (BMI), comorbidities, surgical history, lifestyle history, and clinical stage (Supplementary Table S1). MRI-derived imaging measurements were independently assessed by two radiologists using a standardized protocol. Pathology was recorded according to routine CAP/AJCC reporting; however, Dworak TRG was used for response grouping and model labels in this study.

Prespecified primary endpoint and analysis framework

This retrospective cohort study prespecified the primary endpoint as pathological response stratification after nCRT, defined using the Dworak TRG system: responders (R; TRG 3–4) and nonresponders (NR; TRG 0–1). Guided by this endpoint, three prespecified analyses were performed. First, WES was used to compare somatic mutational landscapes between R and NR, including recurrently mutated genes, tumor mutational burden (TMB), copy number variation (CNV), microsatellite instability (MSI), ploidy, mutational signatures, differential index (DI), hotspot codon events, and pathway-level alterations, to prioritize candidate molecular events whose directionality was consistent with the response phenotype. Second, we developed a weakly supervised multimodal prediction framework integrating WSI- and genome-derived features; model selection was completed within the training set using cross-validation and subsequently evaluated in the held-out internal test set and the external cohort. Third, mechanistic validation focused on biologically plausible candidates identified from WES, using multiomics profiling and functional assays in cell-based and animal models to interrogate links to DNA damage repair and chemoradiotherapy tolerance. Candidate prioritization and mechanistic studies were designed to generate and test biological hypotheses; the clinical predictive value of these candidates warrants confirmation in independent prospective cohorts.

DNA extraction and WES

Genomic DNA was extracted from FFPE tumor specimens using the GeneRead DNA FFPE Kit (180,134, Qiagen, Germany). Peripheral blood samples were centrifuged twice at 1700 rpm to collect leukocytes, and DNA was extracted using the QIAamp DNA Mini Kit (51304, Qiagen, Germany). DNA quantity and integrity were assessed, and qualified samples were used for library preparation. WES libraries were constructed using the Agilent SureSelect Human All Exon V6 kit (5190–8865, Agilent Technologies, USA) and sequenced on an Illumina HiSeq platform (Illumina, USA), with a mean depth of 300× for tumor samples and 100× for blood controls.

Raw data were quality-filtered to obtain clean reads, which were then aligned to the human reference genome (hg19). Low-quality reads and adaptor sequences were removed, and PCR duplicates were marked. Base processing and quality control were performed following GATK best practices (including local realignment and base quality score recalibration). Somatic variants were called in matched tumor–normal pairs and filtered to remove sequencing artifacts, low-support calls, and common germline variants; only high-confidence somatic single-nucleotide variants (SNVs)/insertions/deletions (indels) passing prespecified quality thresholds were retained for downstream analyses.

Somatic mutation calling

High-quality reads were aligned to the human reference genome (UCSC hg19) using Burrows-Wheeler Aligner (BWA v0.7.12) with default parameters. Base quality score recalibration (BQSR) was performed using GATK 4.0, and duplicate reads were removed with Picard v2.13. Somatic single-nucleotide variants (SNVs) and insertions/deletions (indels) were identified using Mutect2 from GATK 4.0 for each matched tumor–normal pair. Low-confidence variants (Tumor Log Odds, TLOD It represents the statistical confidence that a given variant actually exists in the tumor sample, as opposed to being an artifact or background sequencing noise. < 10) were excluded, and high-confidence mutations were annotated in MAF format using vcf2maf. A total of nine mutation types were retained for further analysis, including missense mutations, in-frame insertions, in-frame deletions (dels), nonsense mutations, nonstop mutations, splice site mutations, frameshift dels, frameshift insertions, and splice region variants.

CNV

CNVs in paired tumor and blood samples were assessed using the facet-suite, an R package implementing the Fraction and Allele-Specific Copy Number Estimates from Tumor Sequencing (FACETS) algorithm. Genomic segments with total copy numbers exceeding twice the ploidy level of the sample were classified as amplifications (Amp), while segments with a total copy number of zero were defined as dels. Gene-level copy number was determined on the basis of the segment containing each gene. A CNV-norm score was assigned to each sample to quantify overall CNV burden across the exome. This score was calculated by multiplying the length of each CNV-altered segment by its relative deviation.

TMB and MSI

TMB was defined as the mutation frequency in the Response (R) group (Fr) of nonsynonymous mutations per million base pairs. Nonsilent mutations annotated by vcf2maf included nine categories: missense mutations, in-frame insertions, in-frame dels, nonsense mutations, nonstop mutations, splice site mutations, frameshift dels, frameshift insertions, and splice region variants. These mutations were used to calculate the TMB score, normalized against the total coding sequence (CDS) region (36 Mb) covered by the Agilent V6 WES kit. MSI status for each tumor sample was evaluated using Msisensor. Tumors with Msisensor scores greater than 20 were classified as MSI-high (MSI-H).

Mutational signatures

The Fr of 96 mutation types for each tumor sample was determined on the basis of six types of base substitutions (C > A, C > G, C > T, T > A, T > C, T > G), considering the trinucleotide context by incorporating the adjacent 5′ and 3′ bases. All mutations across samples were aggregated to generate a 96 × N matrix, where N represented the total number of mutations. This matrix was then input into the ConstructSig tool to identify mutational signatures 1–30 from the Catalogue of Somatic Mutations in Cancer (COSMIC) database and quantify their relative contributions in each sample. On the basis of the distribution of mutational profiles in CRC, the signatures were categorized as signature 1, signature 3, signature 6, signature 10, and “others,” the latter representing low-Fr signatures.

Differential index (DI)

To evaluate genomic differences between R and NR, we introduced a novel metric termed the DI. This index was standardized using fold change (FC), which ranged from −1 to 1 and represented the relative Fr of a given genomic feature between the two groups. A positive FC indicated that the feature occurred more frequently in the NR group, whereas a negative FC indicated that Fr was higher in the R group. An FC of 1 or −1 reflected extreme differences, while an FC of 0 denoted no difference between the groups. The formula is defined as follows:

where Fn represents the mutation frequency in the NR group, Fr represents the mutation frequency in the R group, and ∂ is a small constant (set to 0.01 in this study) used to prevent division by zero.

The DI for altered genes was calculated using the same approach, on the basis of the Fr of specific gene alterations between the groups. The formula is as follows:

The DI for altered pathways was derived by summing the contributions of all constituent genes within each pathway. The formula is defined as:

Weakly supervised clustering-constrained attention-based MIL pathology image model

H&E-stained WSIs from 106 patients were collected at the National Cancer Center. To extract pathology phenotypes associated with nCRT response, we adopted the previously reported clustering-constrained attention multiple instance learning (CLAM) framework and built a slide-level classification model [28]. CLAM is an interpretable, weakly supervised MIL approach that relies solely on slide-level labels: an attention module assigns weights to informative instances (patches) for WSI-level prediction, and an instance-level clustering constraint further refines and regularizes the discriminative feature space.

For feature extraction, WSIs were tiled into patches and encoded using a ResNet50 backbone (WSI-level input, not MRI) to obtain instance-level deep features. CLAM then aggregated these features via attention-weighted pooling to generate a WSI-level representation and output the nCRT response prediction. Preprocessing included patching, normalization, and data augmentation with random rotations and flips. The model was implemented in Python 3.9 and TensorFlow 2.10 and trained with a batch size of 32, an initial learning rate of 0.001, the Adam optimizer, and cross-entropy loss for 200 epochs.

Performance was assessed using fivefold cross-validation within the training set for model selection and hyperparameter tuning. The internal and external test sets were reserved for a single final evaluation and were not used for model training, hyperparameter optimization, or threshold determination. All splits were performed at the patient level, and all patches from the same patient were confined to a single dataset to minimize overfitting and data leakage.

Cell lines and cell culture

CRC cell lines RKO and Caco-2 were obtained from the American Type Culture Collection (ATCC, Manassas, VA, USA) with catalog numbers CRL-2577 and HTB-37, respectively. Both cell lines were cultured in Dulbecco’s modified Eagle medium (DMEM)/F-12 medium (11320033, Gibco, USA) supplemented with 10% fetal bovine serum (A5670701, Gibco, USA), at 37 °C in a humidified atmosphere containing 5% CO₂. According to the COSMIC, both cell lines possess wild-type KRAS (KRAS^WT). Endogenous KRAS and other target proteins were detected by western blot (WB) analysis.

WB analysis

Cells were lysed in radioimmunoprecipitation assay (RIPA) buffer (P0013B, Beyotime, China) supplemented with a protease inhibitor cocktail. Lysates were clarified by centrifugation at 12,000g for 20 min. Protein concentrations were determined using the Bradford assay. Proteins were separated by SDS–polyacrylamide gel electrophoresis (PAGE) and transferred to polyvinylidene difluoride (PVDF) membranes. The membranes were probed with the following antibodies: KRAS (3339S, CST, USA), Tubulin (2146, CST, USA), Ligase IV (14649, CST, USA), Ku80 (2753, CST, USA), Ku70 (4588, CST, USA), GAPDH (2118, CST, USA), H3 (9715, CST, USA), anti-rabbit IgG HRP-conjugated antibody (7074P2, GENE-PROTEIN LINK, China), and anti-mouse IgG HRP-conjugated antibody (7076P2, GENE-PROTEIN LINK, China).

Ionizing radiation treatment

Briefly, cells were seeded into six-well plates at appropriate densities and exposed to varying doses of ionizing radiation 12–18 h later (Supplementary Table S3). Cells were then incubated for 10–14 days to allow colony formation. Colonies were fixed with 4% paraformaldehyde at room temperature for 10 min, washed three times with phosphate-buffered saline (PBS), and stained with 0.5% crystal violet (Y268091, Beyotime, China) for 15 min at room temperature. Excess dye was removed by rinsing with water until the background cleared, and plates were air-dried at room temperature. Each radiation dose was tested in triplicate. Colonies were counted using ImageJ (NIH), and survival curves were analyzed using the multitarget single-hit model (y = 1 − (1 − exp(−k × x))ⁿ) and the linear-quadratic model (y = exp(−(axx + bxx²))).

Construction of KRAS knockdown (KRAS^KD) and overexpression cell lines

Transcript information for KRAS was first retrieved from Ensembl (http://grch37.ensembl.org/), and two transcript variants (Transcript IDs: ENST00000311936.3 and ENST00000256078.4) were identified. Among them, Transcript-001 (ENST00000311936.3) was confirmed as the major isoform. Lentiviral vectors expressing shRNA were used to infect cells for KRAS^KD, and the most effective shRNA was selected for subsequent experiments. The sequences of the three human shRNAs are listed in Supplementary Table S4. The shRNA constructs were cloned into the GV654 vector, which contains the following elements: hU6-MCS-Ubiquitin-mCherry-IRES-Neomycin (neo). The vectors were transfected into RKO and Caco-2 cells using Lipofectamine 3000 (L3000001, Invitrogen, Carlsbad, CA, USA). Stable cell lines were selected with G418 (600 ng/mL, ST081, Beyotime, China). Using stable KRAS^KD cell lines, KRAS^G12V and KRAS^A146T overexpression models were generated along with KRAS^WT controls. To prevent shRNA interference with overexpression constructs, synonymous mutations were introduced into the overexpression sequences. All constructs were verified by Sanger sequencing, and empty vectors were used as controls.

Cell proliferation assay

Two KRAS mutant variants (KRAS^G12V, KRAS^A146T and KRAS^G13D) were introduced into two CRC cell lines (RKO and Caco-2), which were seeded into six-well plates at a density of 600 cells per well and irradiated with 2, 4, 6, or 8 Gy. Cells were maintained in culture for 2 weeks, with the medium refreshed every 3 days. Upon completion of the culture period, cell colonies were fixed with methanol and stained with 0.1% crystal violet (C0121, Beyotime, China) for 15 min. After rinsing with distilled water, colonies were photographed and quantified using Image-Pro Plus 6.0 software to calculate the number of visible colonies and determine the surviving fraction.

For proliferation assays, cells subjected to different interventions were seeded at a density of 2 × 10³ cells per well into 96-well plates. After exposure to 8-Gy irradiation, cell proliferation was assessed every 24 h for 96 h using the CCK-8 assay (C0038, Beyotime, China). Absorbance was measured at 450 nm using a microplate reader.

Cell cycle assay

The cell lines were incubated in fresh, serum-free medium and irradiated with 8 Gy. After 24 h of continued culture, the medium was replaced with complete growth medium. Cells were then digested with ethylenediaminetetraacetic acid (EDTA)-free trypsin, washed three times with PBS, and centrifuged to remove the supernatant. The resulting cell pellets were resuspended and fixed in 200 μL of prechilled 70% ethanol. Cell cycle data were acquired using a FACS Calibur flow cytometer (BD) and analyzed with FlowJo software.

Apoptosis assay

When the cells reached 70–80% confluency, they were treated with 10 μM cisplatin (DDP, S1166, Selleck, USA) and irradiated with 8 Gy. After treatment, cells were harvested, digested with 0.25% EDTA-free trypsin, and washed twice with PBS. Apoptosis was evaluated by flow cytometry using Annexin V-FITC/propidium iodide staining (40302ES60, YEASEN Biotech, China).

Immunofluorescence

To assess the effect of KRAS^G12V, KRAS^A146T and KRAS^G13D mutations on DNA DSBs, γ-H2AX foci were detected by immunofluorescence staining. Cells were seeded onto confocal culture dishes (801,001, Nest, USA) to allow adherence and then irradiated with 8 Gy for 1 or 6 h. The medium was then removed, and cells were washed twice with PBS. Cells were fixed with 4% paraformaldehyde (P0099, Beyotime, China) for 20 min, followed by permeabilization with 0.1% Triton X-100 (ST797, Beyotime, China) in PBS for 10 min. After blocking, cells were incubated with a γ-H2AX primary antibody (9718, CST, USA) at 4 °C for 12 h. Subsequently, cells were incubated with a fluorescent goat anti-rabbit secondary antibody (ab150077, Abcam, UK) at room temperature for 1 h, followed by nuclear staining with DAPI (4083S, CST, USA) for 20 min. Images were acquired using a fluorescence microscope.

Comet assay

Comet assays were performed using a DNA Damage Detection Kit (KGA240-100, KeyGEN, China). Following 8-Gy irradiation, cells were suspended in PBS and mixed with 0.7% low-melting-point agarose. The mixture was spread onto microscope slides pre-coated with solidified 1% normal-melting-point agarose. Slides were incubated in lysis buffer at 4 °C for 2 h to lyse the cells, then transferred to alkaline electrophoresis buffer for 30 min to unwind DNA. Electrophoresis was performed at 25 V for 30 min. Cells were stained with 20 μL of propidium iodide solution. Fluorescence images were captured using a fluorescence microscope, and data were analyzed using Fiji and GraphPad Prism 10.1.2.

Multiomics analysis

To explore the mechanistic link between codon-specific KRAS mutations and resistance to nCRT, RNA sequencing (RNA-seq), proteomic, and phosphoproteomic analyses were performed on KRAS^G12V-, KRAS^A146T- and KRAS^G13D-overexpressing cells, as well as KRAS^WT cells. Differentially expressed genes (DEGs), proteins, and phosphopeptides were identified. Selected phosphopeptides showing differential expression were further quantified using PRM analysis. Multiomics analyses were used primarily for exploratory characterization of pathways and signaling differences associated with codon-specific KRAS variants.

Preparation of cell-free extract (CFE)

Approximately 1 × 10⁸ cells were resuspended in twice the cell volume of hypotonic buffer (10 mM Tris–HCl, pH 8.0 [ST780, Beyotime, China]; 1 mM EDTA [ST1303, Beyotime, China]; 5 mM dithiothreitol (DTT) [ST043, Beyotime, China]) supplemented with Complete Mini EDTA-free protease inhibitor (04693159001, Roche, Switzerland), and incubated at 4 °C for 20 min to induce cell swelling. An equal volume of high-salt buffer (50 mM Tris–HCl, pH 7.5 [ST775, Beyotime, China]; 1 M KCl [Y000949, Beyotime, China]; 2 mM EDTA; 2 mM DTT) was then added, followed by further homogenization and incubation on ice for another 20 min. The mixture was centrifuged at 42,000 rpm for 3 h at 4 °C using a TLA-100 rotor in an ultracentrifuge. The supernatant was collected and dialyzed overnight in dialysis buffer (20 mM Tris–HCl, pH 8.0; 100 mM potassium acetate [ST338, Beyotime, China]; 20% v/v glycerol [ST1353, Beyotime, China]; 0.5 mM EDTA; 1 mM DTT; 0.1 mM PMSF [ST505, Beyotime, China]) using Slide-A-Lyzer™ dialysis cassettes (66380, Thermo Fisher Scientific, USA). The resulting CFE was aliquoted into 10-μL portions and stored at −80 °C until use.

5′ end radiolabeling and annealing of oligonucleotides

The 5′ ends of oligonucleotides were radiolabeled to generate the radioactive substrates required for in vitro NHEJ assays. The labeling reaction was performed in a 20 μL reaction mixture containing 20 mM Tris–acetate buffer (pH 7.9, T6025, Sigma-Aldrich, USA), 10 mM magnesium acetate (M5661, Sigma-Aldrich, USA), 50 mM potassium acetate (P5708, Sigma-Aldrich, USA), 1 mM DTT, and [γ-³²P]-ATP (BLU502H250UC, Revvity, USA). T4 polynucleotide kinase (M0201S, New England Biolabs, USA) was added, and the reaction was incubated at 37 °C for 1 h. Labeled oligonucleotides were purified using Micro Bio-Spin 6 chromatography columns (Bio-Rad, USA) and stored at −20 °C until use. For preparation of the double-stranded substrate, the radiolabeled upstream strand SCR19 was mixed with a threefold molar excess of the unlabeled downstream strand SCR20 in a buffer containing 100 mM NaCl (S5886, Sigma-Aldrich, USA) and 1 mM EDTA. The mixture was heated in a boiling water bath for 10 min, then cooled slowly to room temperature to facilitate stable duplex formation.

In vitro NHEJ assay

An in vitro NHEJ assay was performed using a radiolabeled oligonucleotide. The 10-μL reaction mixture contained 4 nM 5′-end radiolabeled double-stranded DNA substrate, CFE, 30 mM HEPES–KOH (pH 7.9, T16723, Shanghai Shangbao Biotechnology, China), 7.5 mM MgCl₂ (208337 Sigma-Aldrich, USA), 1 mM DTT, 2 mM ATP (11140965001, Sigma-Aldrich, USA), 50 μM deoxynucleoside triphosphate (dNTP) mix (71004-M, Sigma-Aldrich, USA), and 0.1 μg bovine serum albumin (B2064, Sigma-Aldrich, USA). The reaction was incubated at 30 °C for 1 h to facilitate end joining. Termination was achieved by adding 10 mM EDTA, followed by phenol/chloroform extraction and ethanol precipitation with glycogen (10901393001, Sigma-Aldrich, USA) as a carrier. The DNA pellet was resuspended in 10 μL of Tris–EDTA buffer (TE buffer) and separated on an 8% urea-denaturing polyacrylamide gel. After drying, the gel was exposed to a phosphor screen and scanned using a GE Typhoon FLA 9500 phosphorimager (GE Healthcare, USA). Signal intensities (PSL/mm²) for each lane were quantified using Multi Gauge software (V3.0, Fujifilm, Japan). Bar graphs were generated and statistical analyses performed using GraphPad Prism.

Intracellular NHEJ and HR reporter assays

The efficiency of NHEJ and HR was assessed using the pimEJ5–GFP and DR–GFP reporter systems, respectively, in combination with I-SceI endonuclease-induced DSBs. A total of 3 × 10⁵ cells were seeded into six-well plates and cotransfected using branched polyethylenimine (PEI, 408727, Sigma-Aldrich, USA). Each well received 20 μg of the I-SceI expression plasmid (26477, Addgene, USA) along with either 15 μg of the pimEJ5–GFP plasmid (44026, Addgene, USA) or 15 μg of the DR–GFP plasmid (113193, Addgene, USA). In all transfections, 5 μg of pcDNA3.1–RFP plasmid (206177, Addgene, USA) was included as an internal control. Following transfection, cells were incubated at 37 °C in a 5% CO₂ incubator for 36 h, after which the percentage of GFP-positive cells was measured using flow cytometry. pcDNA3.1-RFP was used to normalize transfection efficiency, and bar graphs were generated to represent the percentage of GFP-positive cells after normalization.

HR repair assay

Two modified neoresistance gene plasmids, pTO223 and pTO231, were first constructed as substrates for the HR assay. The plasmid pTO221 served as the backbone, derived from the pBluescript SK vector (VT1892, UBio, China), with a 1.5 kb neo-gene fragment from the pBR322::Tn5 plasmid inserted at the HindIII/SalI sites. Based on this construct, pTO223 (neoΔ2) was generated by removing a 245 bp fragment from the 5′ end of the neo gene using NarI digestion. Similarly, pTO231 (neoΔ1) was constructed by deleting a 282-bp fragment from the 3′ end of the neo gene using NaeI. Both plasmids retained a functional ampicillin resistance gene for bacterial selection. Notably, pTO223 contained EcoRI, SalI, and HindIII restriction sites at the 5′ end of the neo gene, whereas pTO231 lacked the HindIII site in the corresponding region.

The HR reaction was carried out in a 25-μL system containing 1 μg of CFE, 500 ng of donor plasmid pTO223, and 500 ng of recipient plasmid pTO231. The reaction buffer included 35 mM HEPES, 10 mM MgCl₂, 1 mM DTT, 1 mM ATP, 50 μM dNTP mix, 1 mM nicotinamide adenine dinucleotide (NAD) (N1511, Sigma-Aldrich, USA), and 100 μg/mL bovine serum albumin. The mixture was incubated at 37 °C for 30 min.

The reaction was terminated by adding EDTA (final concentration 20 mM), proteinase K (200 μg/mL, 1.24568, Sigma-Aldrich, USA), and 0.5% sodium dodecyl sulfate (SDS) (L3771, Sigma-Aldrich, USA). The products were extracted with phenol/chloroform and purified by ethanol precipitation. The DNA was resuspended in 20 μL of TE buffer, and 2 μL was used to transform electrocompetent Escherichia coli DH5α cells (cat. no. 9057, Takara, Japan). Transformed cells were plated onto LB agar containing either ampicillin (100 μg/mL) alone or both ampicillin and kanamycin (50 μg/mL). The number of ampicillin-resistant transformants and dual-resistant recombinant colonies was recorded. Recombination Fr was calculated as the ratio of recombinant colonies to total transformants per microgram of DNA.

Animal model construction

BALB/cA-nu immunodeficient mice, aged 6–8 weeks and weighing 18–20 g, were purchased from Beijing Vital River Laboratory Animal Technology Co., Ltd. (China). All mice were housed in a sterile animal facility under specific pathogen-free (SPF) conditions, maintained at a constant temperature (22 ± 2 °C) and humidity (50 ± 10%), with a 12-h light/dark cycle. Sterilized feed and autoclaved water were provided ad libitum. All procedures complied with ethical guidelines for animal experimentation and were approved by the Institutional Animal Care and Use Committee (IACUC) (approval number: NCC2023 C-248). Animals were housed and handled in accordance with humane care principles, and all efforts were made to minimize their suffering. At the end of the study, mice were euthanized under ether anesthesia.

Before implantation, RKO and Caco-2 cells were transfected with a firefly luciferase reporter plasmid (D2102, Beyotime, China), and stable transfectants were selected for further analysis. A total of 5 × 10⁶ luciferase-expressing cells, resuspended in 100 μL of sterile PBS, were injected subcutaneously into the right dorsal flank of each mouse to establish xenograft tumor models. Treatment was initiated once the tumors reached approximately 100 mm³ in volume. Experimental groups for both RKO and Caco-2 models included: NC group (subcutaneous injection of NC cells), KRAS^KD group (KRAS^KD cells), KRAS^G12V group (KRAS^G12V mutant cells), KRAS^A146T group (KRAS^A146T mutant cells), KRAS^G13D group (KRAS^G13D-mutant cells), and KRAS^WT group (KRAS^WT cells).

nCRT treatment was initiated on day 10 following cell transplantation. The chemotherapy regimen consisted of a single dose of localized 6 Gy irradiation on day 10, followed by intraperitoneal injections of oxaliplatin (5 mg/kg; HY-17371, MedChemExpress, USA) every 2 days starting the same day. Radiotherapy was delivered using an X-ray irradiator specifically designed for small animals (RS2000, Rad Source, USA), with a localized dose of 6 Gy to the tumor. Mice were closely monitored for general health status throughout the treatment period.

On days 0, 4, 8, and 12 of treatment, mice were anesthetized for in vivo imaging. d-luciferin (150 mg/kg, ST196, Beyotime, China) was administered via intraperitoneal injection. After 10 min, the mice were deeply anesthetized with 2% isoflurane in an oxygen mixture (1 L/min, R510-22-10, RWD, China), and tumor growth was imaged using an in vivo imaging system (IVIS Lumina Series, PerkinElmer, USA). Tumor volume was measured every 2 days using a vernier caliper. Length (L) and width (W) were recorded, and tumor volume was calculated using the formula: V = (L × W²)/2. Monitoring continued for 14 days. On day 14, all mice were euthanized, and tumors were excised and weighed using an electronic balance (Sartorius, Germany). The tumors were then documented via photography [45, 46].

TUNEL staining for apoptosis detection

Tumor tissues were fixed in 4% paraformaldehyde for 24 h, paraffin-embedded, and sectioned at 4 μm. TUNEL staining was performed using a commercial kit (C1088, Beyotime, China) according to the manufacturer’s instructions. Nuclei were counterstained with DAPI. TUNEL-positive cells were observed under a fluorescence microscope, and the positive rate was calculated on the basis of captured images.

Immunohistochemical staining

Following deparaffinization and rehydration through graded alcohols, antigen retrieval was performed. Endogenous peroxidase activity was blocked using 3% H₂O₂ (H1009, Sigma-Aldrich, USA). After blocking, sections were incubated overnight at 4 °C with primary antibodies against γ-H2AX (A700-053, Thermo Fisher, USA; 1:200) and Ki67 (9027, CST, USA; 1:200). The next day, sections were incubated for 30 min with goat anti-rabbit HRP-conjugated secondary antibody (ab6721, Abcam, UK). Signal development was performed with DAB (P0203, Beyotime, China), followed by hematoxylin counterstaining (H3136, Sigma-Aldrich, USA), dehydration, and mounting. Images were captured under a light microscope, and the positive-staining area was quantified using ImageJ.

Statistical analysis

Statistical analyses were performed using GraphPad Prism 10.1.2 and R 4.2.1. In vitro and in vivo data are presented as mean ± standard deviation (SD) from at least three independent experiments. Two-group comparisons were performed using two-sided unpaired t-tests, and comparisons among three or more groups were performed using one-way analysis of variance (ANOVA) with Tukey’s post hoc test. When assumptions of normality or homogeneity of variance were not met, the Mann–Whitney U test or Kruskal–Wallis test was used, respectively. A two-sided p-value < 0.05 was considered statistically significant. For high-dimensional omics and genomic analyses (including gene- and pathway-level WES analyses and differential analyses of transcriptomic, proteomic, and phosphoproteomic data), multiple testing was controlled using the Benjamini–Hochberg procedure to estimate the false discovery rate (FDR), with correction applied separately within each analysis family.

Results

Genomic profiling of patients with LARC resistant to nCRT

Patient cohort

We retrospectively collected data from 137 patients with LARC who underwent nCRT followed by TME. All patients were classified as having LARC based on CT and/or MRI, defined as clinical stage T3-T4 and/or N⁺. Preoperative colonoscopic biopsy and pathological examination confirmed that all cases were rectal adenocarcinoma. The internal cohort included 115 patients treated at the National Cancer Center, while an additional 22 patients with available WES data were sourced from previously published studies by Gu et al. and Lin et al., comprising the external cohort [23, 43]. All 137 patients met the inclusion criteria. Paired pretreatment tumor and peripheral blood samples were obtained from 67 patients in the internal cohort for WES analysis. Additionally, FFPE tumor samples from 106 patients in the internal cohort were subjected to H&E staining and WSI scanning for subsequent pathological analysis. The study design is illustrated in Fig. 1A.

Fig. 1.

Overview of the models and datasets used in this study and the genomic landscape of patients with LARC, stratified by sensitivity or resistance to nCRT. A Study cohorts and data modalities. patients with LARC treated with nCRT followed by TME were included (total n = 137). Internal cohort: WES (tumor–blood pairs, n = 67) and H&E-stained WSIs (n = 106). External cohort: WES only (n = 22; Gu et al. and Lin et al.) for independent genomic validation; not used for WSI analyses. Mechanistic assays were performed in KRAS^WT, KRAS^G12V and KRAS^A146T cell models. B Oncoprint of recurrent alterations in the LARC cohort. Alteration classes include missense, truncating, frameshift, splice-site, amplification (Amp), and deletion (Del). Top: altered genes per patient; right: alteration frequency per gene. Clinical annotations include sex, age, chemotherapy, cT stage, cN stage, and serum CEA/CA19–9

The pathological response to nCRT was evaluated using the TRG system described by Dworak [44]. Patients with TRG 3 or TRG 4 were classified as the R group, whereas those with TRG 0 or TRG 1 were categorized as the NR group. Among the 137 patients, 65 were assigned to the R group and 72 to the NR group (Supplementary Table S1). For model development, 99 patients were randomly assigned to the training set and 38 to the test set (Supplementary Table S2). The training set consisted of 51 R and 48 NR patients, while the test set comprised 14 R and 24 NR patients. All 22 patients in the external cohort were included in the test set (Supplementary Fig. S1A). The external cohort provided WES data only and was used for independent genomic-level validation; it was not included in WSI-based analyses. To minimize data leakage, model development and evaluation were performed with patient-level splitting, such that all data from the same patient were confined to a single dataset.

NR-specific detection of KRAS^G12V and KRAS^A146T

WES of 67 patient samples identified 13,576 putative pathogenic events, including 12,213 SNVs and 1363 indels. The most frequently mutated genes are summarized in Fig. 1B. APC showed the highest mutation frequency (74%), consistent with prior reports [47, 48]. Other established colorectal cancer driver genes were also detected, including TP53 (59%), KRAS (48%), and TTN (43%). The overall mutation frequencies of these highly mutated genes did not differ significantly between R and NR (Supplementary Fig. S1B). Genome-wide features, including TMB, CNV, MSI status, and ploidy, were also comparable between the two groups (Supplementary Fig. S1C).

Codon-specific hotspot events were then evaluated using the DI and event frequency. In total, 41 hotspot events met the criteria (|DI|> 0.6 and Fr > 0.08), with 26 events aligned with the poor-response direction and 15 aligned with the good-response direction (Fig. 2A). KRAS contributed the most hotspot sites and exhibited a high DI value (DI = 0.89), with major sites including KRAS^G12V, KRAS^A146T, and KRAS^G13D. KRAS^G12V was also observed in an external colorectal cancer cohort (TCGA CRC: 6/378), supporting detectability across datasets. In our cohort, KRAS^G12V and KRAS^A146T were detected in the NR group (Fig. 2A) and were directionally consistent with poor response. Given that hotspot codon events are typically low-frequency alterations and that this study is retrospective, these findings were considered exploratory association signals intended to support hypothesis generation and subsequent mechanistic validation. Because the number of event carriers was limited, we did not perform multivariable adjustment for key covariates (including clinical stage and T/N status); therefore, no confirmatory inference is made regarding an independent predictive effect of these codon-level events.

Fig. 2.

Differential genomic biomarkers predictive of response to nCRT. A In total, 41 hotspot mutations showing significant differences. The heat map displays the presence (blue cells) or absence (white cells) of specific mutations across individual patients in the R and NR groups. The sidebar indicates the total number of mutations observed in TCGA READ samples. B Comparison of C > G mutation frequencies between R and NR groups (Kruskal–Wallis rank-sum test). C Mutational signature composition for each sample in the R and NR groups. Bar plots with different colors represent the relative abundance of each mutational signature. The upper right histogram shows the proportion of R and NR patients when the abundance of signature 1 is > 0.75 (left bar), < 0.75 (middle bar), or unfiltered (right bar). The lower right histogram similarly shows the patient distribution when signature 3 is ≥ 0.2 (left bar), < 0.2 (middle bar), or overall (right bar). D Heat map of DI scores for specific signaling pathways between R and NR groups. Each row represents a pathway, and each column corresponds to an individual patient. Color intensity reflects the DI of the pathway in each patient. The sidebar shows the total SUM DI values for each pathway between the two groups

Base substitution and mutational signatures are associated with nCRT response

Somatic mutations arising from intrinsic and extrinsic cellular processes leave distinct mutational signatures across the genome, which have broad applications in cancer treatment and prevention. We first analyzed the Fr of six types of base substitutions (C > A, C > G, C > T, T > A, T > C, and T > G) and found that C > G substitutions were significantly more frequent in the R group (p = 0.019; Fig. 2B). Mutational signature analysis identified signatures 1, 3, and 6 as the predominant patterns (Fig. 2C). Next, we performed unsupervised clustering of 30 mutational signatures and identified six distinct clusters. A high Fr of signature 1 characterized cluster 1 and was predominantly composed of NR cases (11/13, 85%). In contrast, cluster 6 was enriched for signature 3 and consisted exclusively of R cases (5/5, 100%) (Supplementary Fig. S2A). Further analysis confirmed that signatures 1 and 3 were significantly associated with nCRT response. Stratification based on a signature 1 abundance threshold of 0.75 (high versus low) revealed a significant difference in treatment response (p = 0.004; Fig. 2C and S2B). Similarly, using a 0.25 threshold for signature 3, we observed a notable difference in response (p = 0.059; Fig. 2C and Supplementary Fig. S2C).

Genomic alterations associated with nCRT resistance

We further investigated mutations beyond well-established driver genes and identified 151 differentially altered genes on the basis of thresholds of |DI|> 0.6 and Fr > 0.1. To assess whether these somatic mutations were enriched in specific biological pathways, we conducted a differential pathway alteration analysis between the R and NR groups. Using a t-test (p < 0.01), we identified 21 candidate pathways that met the criteria of a mean DI > 0.33 (indicating at least a 1.5-FC) and an alteration Fr > 0.5 (altered in at least half of the patients). Among these, the ubiquitin-mediated proteolysis pathway showed the highest mean DI (Fig. 2D), suggesting a potential association with nCRT resistance. This pathway was subsequently selected as the foundation for model development. Previous studies have reported that ubiquitin-mediated proteolysis promotes TRIB3-induced radiotherapy resistance in esophageal squamous cell carcinoma [49]. Additionally, the ABC transporter pathway emerged as another candidate associated with treatment resistance (Fig. 2D). The NR group exhibited higher mutation frequencies in both the ubiquitin-mediated proteolysis and ABC transporter pathways (Supplementary Fig. S3A and S3B).

Development and validation of a genomic–pathologic fusion model to predict radiotherapy resistance

Development and evaluation of individual models

On the basis of the genomic and pathological data available in this study, we defined five information modalities: mutations (M), mutational signatures (S), pathology (P), base-substitution patterns (B), and ubiquitin-mediated proteolysis (U). Integrating these modalities can better capture the heterogeneity of nCRT response in LARC, but any single modality is informative only for a subset of patients. In addition, modalities differ markedly in feature dimensionality and sample availability, creating feature imbalance that can degrade conventional fusion approaches (for example, direct concatenation or pooling). To address this, we developed a cascaded integration framework to reduce the impact of feature imbalance.

We first trained separate unimodal predictors for each modality (M, S, P, U, and B) (Fig. 3A) and evaluated decision reliability using accuracy (ACC) in repeated fivefold cross-validation. The M model achieved the highest ACC, followed by S, P, U and B (Fig. 3B). We also quantified sample coverage, defined as the proportion of patients each model could evaluate. Although the M model had the best ACC, its coverage was relatively limited, whereas the P model showed moderate ACC but the broadest coverage because pathology was available for all patients (Fig. 3C). These results indicate that delineating the patient subgroups in which each modality is informative and integrating models accordingly is critical to improving overall performance and applicability.

Fig. 3.

Construction and evaluation of unimodal and multimodal fusion models for predicting patient response to nCRT. A Individual-level classification outputs of unimodal models (M, S, P, U and B) in the training and test sets. Correct, predicted class matches the ground truth; Wrong, predicted class is opposite to the ground truth; Uncertain, modality input is available, but the predicted confidence does not meet the prespecified decision criterion, and no definitive class is returned (safe abstention); None, the required modality input is missing, and no prediction is generated. B Bar plot summarizing the proportions of Correct, Wrong, and Uncertain predictions for each unimodal model in the training set; the line indicates the corresponding ACC. C Construction and evaluation of multimodal fusion models. Left, modality combinations used by each model; right, discrimination under the training/test evaluation framework (ROC curves and AUC), overall ACC, coverage, and cost index. Coverage is defined as the proportion of cases for which a definitive prediction (not Uncertain) is returned, conditional on modality availability. D ROC curve of the M/S/P fusion model in the test set

Development and evaluation of the fusion model

We integrated the unimodal models into a multimodal fusion model based on principles established through cross-validation. Specifically, for each patient, we first identified all applicable models—defined as those that encountered cases similar to the patient’s during training and were thus considered to “cover” the patient. For example, the M model applied only to patients with at least 41 somatic mutations, whereas the P model covered all patients with available WSI. Among the applicable models, we selected the one with the highest ACC for patient classification.

Subsequently, we evaluated various combinations of the M, S, P, U, and B models in the test set (Supplementary Table S5). To determine the optimal fusion model, we considered multiple criteria, including the area under the curve (AUC) of the receiver operating characteristic (ROC) curve, ACC, coverage, and relative healthcare cost. Based on these metrics, the combination of the M, S, and P models emerged as the optimal solution. This tri-modal model achieved the third-highest ACC (0.931), full patient coverage (100%), and a relatively low cost index (2) (Fig. 3C). The model’s AUC was 0.882 (Fig. 3D). Its superior performance over any unimodal model was attributed to the high coverage of pathological data and the predictive strength of omics-derived features. Additionally, the combination of M and P models offered a viable alternative, achieving full patient coverage (100%), a competitive ACC (ACC = 0.913, AUC = 0.815), and the same cost index (2).

KRAS^G12V and KRAS^A146T mutations confer NCRT resistance by enhancing DNA damage repair following irradiation

KRAS^G12V and KRAS^A146T mutations affect radiosensitivity

KRAS mutations are common in CRC and activate oncogenic signaling pathways, but whether codon-specific KRAS variants modulate sensitivity to nCRT remains unclear. In our cohort, KRAS^G12V, KRAS^A146T and KRAS^G13D were detected in nonresponders, suggesting a potential link between selected codon-specific KRAS variants and radioresistance. These findings suggest codon-dependent differences in KRAS-associated treatment tolerance, without establishing exclusivity to these variants. To test this, we generated stable KRAS^KD RKO and Caco-2 cell lines and, on this background, established RKO and Caco-2 cells stably overexpressing KRAS^WT, KRAS^G12V, KRAS^A146T or KRAS^G13D, together with an empty-vector control (Fig. 4A).

Fig. 4.

In vitro experiments assessing the impact of KRAS^G12V and KRAS^A146T mutations on radiosensitivity. A Immunoblotting of KRAS expression in RKO and Caco-2 CRC cells overexpressing KRAS^WT, KRAS^G12V, KRAS^A146T or KRAS^G13D. B Clonogenic survival after irradiation in the indicated RKO and Caco-2 cell lines. C Apoptosis after irradiation measured by flow cytometry in cells expressing KRAS^WT, KRAS^G12V, KRAS^A146T or KRAS^G13D. D Cell-cycle distribution after irradiation in the indicated cell lines. E Cell proliferation measured by CCK-8 in cells expressing KRAS^WT, KRAS^G12V, KRAS^A146T or KRAS^G13D. Data are mean ± SD from three independent experiments. Statistical significance is indicated as *p < 0.05, **p < 0.01, ***p < 0.001, ****p < 0.0001; ns, not significant

Clonogenic survival assays following irradiation showed that KRAS^G12V and KRAS^A146T overexpression reduced radiosensitivity, yielding significantly more colonies than KRAS^WT, whereas KRAS^G13D displayed an intermediate phenotype between KRAS^WT and KRAS^G12V/KRAS^A146T. In contrast, KRAS^KD cells were the most radiosensitive (Fig. 4B). Although both KRAS^G12V and KRAS^A146T increased resistance, the surviving fraction at 6 Gy was significantly higher in KRAS^G12V than in KRAS^A146T, indicating a stronger effect of G12V on clonogenic survival. Consistently, KRAS^G13D showed survival higher than KRAS^WT but lower than KRAS^G12V/KRAS^A146T, supporting a codon-dependent gradient of effect. Together, these results link KRAS^G12V and KRAS^A146T to reduced radiosensitivity in CRC cell models.

KRAS^G12V and KRAS^A146T mutations suppress radiation-induced apoptosis and modulate the cell cycle

To further define the effects of KRAS^G12V and KRAS^A146T on radiosensitivity, we assessed apoptosis, cell-cycle distribution and proliferative capacity after irradiation. Compared with KRAS^WT, apoptosis after irradiation was reduced in KRAS^G12V- and KRAS^A146T-expressing cells; KRAS^G13D showed an intermediate phenotype, whereas KRAS^KD cells exhibited the highest apoptosis (Fig. 4C). Cell-cycle analysis showed a decreased S-phase fraction in KRAS^KD cells, while KRAS^G12V and KRAS^A146T were associated with a higher S-phase fraction; KRAS^G13D again showed an intermediate pattern (Fig. 4D). In proliferation assays, KRAS^KD reduced post-irradiation survival and growth, whereas KRAS^G12V and KRAS^A146T increased post-irradiation survival and growth; KRAS^G13D remained intermediate (Fig. 4E).

Differences were also observed between KRAS^G12V and KRAS^A146T. Apoptosis was lower in KRAS^A146T than in KRAS^G12V, whereas the S-phase fraction was higher in KRAS^G12V and G2/M arrest was less pronounced in KRAS^A146T. Overall, KRAS^G12V and KRAS^A146T were associated with reduced irradiation-induced apoptosis, maintenance of S-phase cells and increased post-irradiation proliferative capacity, consistent with their enrichment in the NR group (Fig. 4C–E).

RNA-seq analysis

To investigate the mechanisms by which KRAS^G12V and KRAS^A146T mutations enhance radiotherapy resistance in CRC cell lines, we performed RNA-seq analysis on cells exogenously expressing KRAS^G12V, KRAS^A146T, or KRAS^WT proteins. These experiments were intended to characterize mechanism-relevant signaling changes in CRC models rather than to recapitulate rectal-site specificity. Differential expression analysis revealed that, compared to KRAS^WT cells, KRAS^G12V mutant cells exhibited 1,339 upregulated and 1,746 downregulated genes (|log₂FC|> 1.0, p < 0.05). Similarly, KRAS^A146T mutant cells showed 358 upregulated and 2,081 downregulated genes (Supplementary Fig. S4A-B, Supplementary Table S6–7). Selected DEGs are visualized in Fig. 5A, B as descriptive summaries. To characterize the distinct biological processes and pathways regulated by KRAS^G12V and KRAS^A146T mutations, we performed Gene Ontology (GO) (Supplementary Fig. S4C-D) and Kyoto Encyclopedia of Genes and Genomes (KEGG) (Fig. 5C, D) enrichment analyses. KEGG analysis of the differential-expression results indicated enrichment of several DNA repair-related pathways in KRAS^G12V mutant cells, including nucleotide excision repair, base excision repair, HR, and mismatch repair, indicating that the KRAS^G12V mutation promotes DNA repair (Fig. 5C). In addition, the type I interferon signaling pathway was upregulated in KRAS^G12V-overexpressing cells (Supplementary Fig. S4C). In KRAS^A146T cells, enrichment analysis identified upregulation of the MAPK signaling pathway, TNF signaling, and other oncogenic pathways (Fig. 5D). Notably, genes involved in DNA replication were differentially expressed in KRAS^A146T cells (Supplementary Fig. S4D).

Fig. 5.

Exploratory multiomics analyses of KRAS^G12V and KRAS^A146T mutant cells. Heat maps of differentially regulated genes in KRAS^G12V versus KRAS^WT cells (A) and in KRAS^A146T versus KRAS^WT cells (B). KEGG enrichment analyses based on differentially regulated genes in KRAS^G12V versus KRAS^WT cells (C) and KRAS^A146T versus KRAS^WT cells (D). Heat maps of differentially regulated phosphopeptides in KRAS^G12V E and KRAS^A146T (F) cells compared to KRAS^WT cells. KEGG (G) and GO (H) enrichment analyses of differentially regulated phosphopeptides in KRAS^G12V versus KRAS^WT cells. (I) Heat map of differentially regulated phosphopeptides between KRAS^A146T and KRAS^WT cells identified by PRM analysis. Heat maps in panels A, B, E, F and I are shown as descriptive visual summaries of selected differential features

Proteomic analysis

Next, we conducted exploratory global proteomic profiling using TMT-based quantitative analysis to identify differentially regulated proteins between KRAS^WT cells and the two KRAS mutant cell lines (Supplementary Fig. S5A, B and Supplementary Table S8–9). In KRAS^G12V mutant cells, compared with KRAS^WT cells, we identified 17 upregulated proteins and 1 downregulated protein (FC ≥ 1.2, |log₂FC|, p < 0.05) (Supplementary Fig. S5C). Similarly, in KRAS^A146T mutant cells, 16 proteins were upregulated, and 5 were downregulated (Supplementary Fig. S5D). Enrichment analysis revealed that, compared with KRAS^WT cells, the following pathways were significantly enriched in KRAS^G12V cells: negative regulation of protein localization, Salmonella infection, and the NOD-like receptor signaling pathway (Supplementary Fig. S5E–F). In KRAS^A146T cells, enriched pathways included peptidyl-glutamic acid modification, HTLV-1 infection, and thermogenesis (Supplementary Fig. S5G–H).

Phosphoproteomic analysis and PRM validation

Phosphoproteomic profiling was performed as an exploratory analysis to investigate phosphorylation-associated mechanisms and to identify differentially abundant phosphopeptides between KRAS-mutant and KRAS^WT cells (Fig. 5E, F and Supplementary Table S10–S11). Relative to KRAS^WT, KRAS^G12V cells showed 763 upregulated and 444 downregulated phosphopeptides (FC > 1.2, p < 0.05), whereas KRAS^A146T cells showed 2,049 upregulated and 1,716 downregulated phosphopeptides (Supplementary Fig. S6A, B). Enrichment analysis indicated over-representation of DNA repair, cellular response to DNA damage, DNA replication and cell-cycle processes in KRAS^G12V cells (Fig. 5G, H). No significantly enriched biological pathways were identified in KRAS^A146T cells (Supplementary Fig. S6C). These phosphoproteomic findings were used to support exploratory pathway interpretation and were not treated as stand-alone proof of mechanism. Differential phosphorylation events were predominantly mapped to DNA damage response and cell-cycle checkpoint pathways, consistent with downstream signaling changes associated with chemoradiotherapy resistance at the genomic level.

Fifty differential phosphopeptides were further assessed by PRM. Figure 5I and Supplementary Fig. S6D are shown as descriptive visual summaries of the quantified phosphopeptides, whereas the interpretive emphasis is placed on the quantified PRM results reported in Supplementary Table S12. PRM results were concordant with the discovery phosphoproteomics, showing increased phosphorylation of ZYX, TJAP1, and KLC3, and decreased phosphorylation of USP22, MYO9B, and DDX20 in KRAS^A146T cells.

Codon-specific KRAS variants were associated with reduced radiation-induced DNA damage

To assess whether KRAS variants contribute to radioresistance by attenuating DNA damage, γ-H2AX immunofluorescence was performed in RKO and Caco-2 cells after irradiation (Fig. 6A). γ-H2AX foci were most abundant in KRAS^KD cells and lowest in KRAS^G12V- and KRAS^A146T-expressing cells. Compared with KRAS^WT, KRAS^G12V and KRAS^A146T showed reduced γ-H2AX foci formation, whereas KRAS^G13D displayed an intermediate phenotype between KRAS^WT and KRAS^G12V/KRAS^A146T (Fig. 6A). Comet assays performed after irradiation showed a consistent pattern (Fig. 6B): KRAS^G12V and KRAS^A146T cells exhibited reduced comet tail formation relative to KRAS^WT, with KRAS^G13D again intermediate, indicating lower levels of irradiation-induced DNA damage.

Fig. 6.

KRAS^G12V, KRAS^A146T, and KRAS^G13D mutations confer radiotherapy resistance by suppressing DNA damage. A γ-H2AX immunofluorescence in parental RKO/Caco-2 cells, KRAS^KD cells, and cells expressing KRAS^WT, KRAS^G12V, KRAS^A146T or KRAS^G13D after irradiation; γ-H2AX foci were reduced in KRAS^G12V, KRAS^A146T and KRAS^G13D relative to KRAS^WT. B Comet assays after irradiation; comet tail formation was reduced in KRAS^G12V, KRAS^A146T and KRAS^G13D relative to KRAS^WT, consistent with lower levels of irradiation-induced DNA damage. Data are mean ± SD from three independent experiments. Statistical significance is indicated as **p < 0.01, ***p < 0.001, ****p < 0.0001; ns, not significant

Because DSB repair is mediated primarily by NHEJ and HR, these findings are consistent with enhanced DSB repair capacity in KRAS^G12V and KRAS^A146T cells, potentially involving NHEJ as the dominant DSB-resolution pathway. Across assays, both KRAS^G12V and KRAS^A146T were associated with decreased radiation-induced DSB signals, with KRAS^G12V showing a pattern consistent with faster damage clearance and KRAS^A146T showing a pattern consistent with stronger damage avoidance, suggesting codon-dependent differences in DNA damage response dynamics.

KRAS mutations enhance NHEJ and HR repair activity in tumor cells

Given the reduced γ-H2AX and comet tail signals, NHEJ activity was assessed using both cell-free and cellular reporter assays. CFE from RKO and Caco-2 cells were prepared with comparable protein input across conditions (Supplementary Fig. S7A). In an in vitro end-joining assay using a radiolabeled dsDNA substrate with 5′ overhangs, control CFE generated the expected ligation products, including a ~150-nt dimer and closed circular products migrating at ~75–100 nt, consistent with intact end-joining activity (Fig. 7A). CFE from KRAS^KD cells showed the weakest ligation activity. Relative to KRAS^WT, KRAS^G12V and KRAS^A146T increased end-joining product formation, whereas KRAS^G13D showed an intermediate pattern (Fig. 7A, B).

Fig. 7.

KRAS^G12V, KRAS^A146T, and KRAS^G13D mutations alter NHEJ repair activity. A Denaturing PAGE analysis showing end-joining efficiency of radiolabeled double-stranded DNA substrates incubated with CFE from different sources. M indicates the 50 bp DNA ladder. Lane 0 represents the substrate-only control without the addition of CFE. B Quantification of end-joining products from different CFE sources. C Schematic of the intracellular NHEJ repair assay. The reporter plasmid pimEJ5–GFP contains an inactivated GFP sequence flanked by two I-SceI recognition sites. Upon cotransfection with the I-SceI expression plasmid, DSBs are induced. If NHEJ repair occurs, GFP expression is restored. Schematic created in BioRender. D Quantification of GFP-positive cells in RKO and Caco-2 cell lines. Quantitative data are presented as mean ± SD. All experiments were independently repeated three times. **p < 0.01, ***p < 0.001, ****p < 0.0001

Endogenous NHEJ capacity was evaluated using the pimEJ5–GFP reporter, in which I-SceI–induced DSB repair restores GFP expression (Fig. 7C). After the cotransfection of I-SceI and pimEJ5–GFP, the GFP-positive fraction measured by flow cytometry was lowest in KRAS^KD cells and highest in KRAS^G12V and KRAS^A146T cells; KRAS^G13D again showed intermediate activity between KRAS^WT and KRAS^G12V/KRAS^A146T (Fig. 7D, Supplementary Fig. S8A–B). Across the in vitro ligation assay and the reporter system, KRAS^G12V yielded a higher end-joining output than KRAS^A146T (Fig. 7A–D).

Chromatin fractionation after irradiation showed that chromatin-associated Ku70, Ku80, and LIG4 were lowest in KRAS^KD cells and highest in KRAS^G12V and KRAS^A146T cells, with KRAS^G13D remaining intermediate relative to KRAS^WT (Supplementary Fig. S7B), supporting enhanced recruitment of core NHEJ factors in KRAS-mutant cells.

HR activity was then assessed using DR–GFP. The GFP-positive fraction was lowest in KRAS^KD cells and increased in KRAS^G12V and KRAS^A146T cells relative to KRAS^WT, with KRAS^G13D again intermediate (Supplementary Figs. S9A and S8C, D). A plasmid-based HR assay using neoΔ1/neoΔ2 substrates incubated with CFE showed concordant results: HR output was lowest in KRAS^KD cells and increased in KRAS^G12V and KRAS^A146T cells compared with KRAS^WT, with KRAS^G13D intermediate (Supplementary Fig. S9B–E).

Together, these assays indicate increased DSB repair capacity through both NHEJ and HR in KRAS-mutant cells. Relative to KRAS^WT, KRAS^G12V and KRAS^A146T were associated with higher repair activity and increased recruitment of DSB repair factors, whereas KRAS^G13D showed an intermediate phenotype and KRAS^KD showed the weakest repair capacity. Across assays, KRAS^G12V showed higher NHEJ output and stronger recruitment of NHEJ factors than KRAS^A146T, whereas KRAS^A146T showed a trend toward higher HR output, consistent with codon-dependent differences in pathway engagement.

KRAS variants promoted resistance to radiotherapy plus oxaliplatin in CRC xenografts

Subcutaneous CRC xenografts were established in BALB/cA-nu immunodeficient mice using RKO or Caco-2 cells stably expressing NC, KRAS^KD, KRAS^WT, KRAS^G13D, KRAS^G12V or KRAS^A146T (Fig. 7A). On day 10 after implantation, tumors received a single local 6-Gy irradiation, followed by intraperitoneal oxaliplatin (5 mg/kg) every 2 days starting on day 10 (Fig. 8A). Tumor burden was monitored longitudinally, with IVIS measurements on days 0, 4, 8, and 12 of treatment. KRAS^KD xenografts showed the slowest growth, whereas KRAS^G12V and KRAS^A146T xenografts grew fastest. Relative to KRAS^WT, tumor growth was increased in KRAS^G12V and KRAS^A146T, with KRAS^G13D consistently intermediate (Fig. 8B). Caliper-derived tumor volumes recapitulated the same pattern (Fig. 8C). At day 14, excised tumors from KRAS^KD groups were smallest, whereas KRAS^G12V and KRAS^A146T groups were largest; KRAS^G13D remained intermediate relative to KRAS^WT (Fig. 8D). Endpoint tumor weights were consistent with these measurements (Fig. 8E).

Fig. 8.

KRAS variants modulate response to radiotherapy plus oxaliplatin in CRC xenografts. A Schematic of subcutaneous xenograft establishment and treatment schedule. B Tumor burden measured by IVIS Lumina II on treatment days 0, 4, 8, and 12. C Tumor growth curves based on caliper-measured volumes. D Representative images of excised xenograft tumors. E Endpoint tumor weights. Data are mean ± SD; n = 6 mice per group. Statistical significance is indicated as ***p < 0.001, ****p < 0.0001

To assess treatment-associated cellular responses in vivo, tumor sections were analyzed by TUNEL staining. TUNEL positivity was highest in KRAS^KD tumors and lowest in KRAS^G12V and KRAS^A146T tumors, with KRAS^G13D intermediate between KRAS^WT and KRAS^G12V/KRAS^A146T (Supplementary Fig. S10A). Ki67 immunohistochemistry showed the lowest proliferative index in KRAS^KD tumors and higher Ki67 in KRAS^G12V and KRAS^A146T tumors compared with KRAS^WT, again with KRAS^G13D intermediate (Supplementary Fig. S10B). γ-H2AX immunohistochemistry, used as a marker of DNA damage signaling, was highest in KRAS^KD tumors and lower in KRAS^G12V and KRAS^A146T tumors relative to KRAS^WT, with KRAS^G13D intermediate (Supplementary Fig. S10C). Across readouts, codon-specific KRAS variants were associated with reduced DNA damage signaling and apoptosis and increased tumor growth under radiotherapy plus oxaliplatin in CRC xenograft models, with KRAS^G13D consistently showing an intermediate phenotype.

Discussion

This study addressed the heterogeneity of nCRT response in LARC by developing and evaluating a multimodal prediction framework in a clinical LARC cohort and by performing mechanistic validation in colorectal cancer experimental models. The fusion model showed overall discriminative performance for separating responders from nonresponders in LARC, whereas the mechanistic experiments, particularly the orthogonal functional assays, provided biological plausibility for KRAS-associated chemoradiotherapy tolerance at the level of CRC model systems.

Unimodal models were first built and evaluated across ACC, coverage and resource cost, followed by cascaded integration into fusion models. The M/S/P fusion model, integrating the M, S and P unimodal predictors, achieved strong discrimination (AUC = 0.882), full coverage and relatively controllable cost, reflecting a trade-off among performance, accessibility, and resource consumption. A previously reported WSI-based machine-learning model also predicted nCRT response in LARC [50] but showed lower discrimination (AUC = 0.779) [51]. Radiomics-based prediction has likewise been widely explored in LARC [52–54], and a radiopathomics model integrating pelvic MRI and WSI showed performance comparable to that observed here [55]. Collectively, these studies indicate that integrating multisource information can improve model robustness and clinical applicability.

For KRAS, no significant association was observed between overall KRAS mutation status and response to standard nCRT, whereas codon-specific variants were directionally aligned with resistance. Prior evidence is inconsistent: in settings combining nCRT with anti-EGFR therapy, KRAS mutations were associated with lower pathological response [37, 51]; in large-scale genomic/transcriptomic analyses and clinical studies, overall KRAS mutation status did not show a stable association with nCRT response [38, 39]. Meta-analyses likewise did not support an association between KRAS mutations and improvements in pCR or TRG [56]. Other studies reported associations between KRAS and nCRT resistance [40, 57]. Treating KRAS as a binary variable may obscure functional and clinical differences among codon-specific variants [41, 58].

Systematic screening of codon-specific KRAS events prioritized KRAS^G12V and KRAS^A146T. This prioritization was based not only on their differential indices and detection pattern in our cohort but also on external evidence that codon-specific KRAS mutations can confer distinct biological and clinical consequences in colorectal cancer. Prior studies have shown that specific KRAS variants are not functionally equivalent, and that codon 12 mutations, particularly c.35G > T (G12V), may be associated with inferior outcomes in CRC [59, 60]. In addition, KRAS A146-mutant CRC has been described as a distinct molecular subgroup with adverse clinical features [61]. We also considered the maturity of clinical KRAS testing workflows, which routinely cover exon 2/3/4 alterations including G12V and A146T [62, 63]. These considerations supported mechanistic prioritization of KRAS G12V and KRAS A146T over other low-frequency NR-enriched events. At the same time, KRAS-associated treatment tolerance was not limited to G12V and A146T. G13D showed partially overlapping phenotypes in several functional assays, although effect size and pathway emphasis differed across variants. Together with the nonidentical distribution of codon-specific events across responders and nonresponders, this pattern supports codon-dependent heterogeneity rather than strict allele exclusivity.

The predictive value of KRAS is treatment-pathway dependent: associations may vary with different neoadjuvant sequences (for example, induction chemotherapy before versus after chemoradiotherapy) [38]. Fractionation and duration of radiotherapy, the interval from nCRT to surgery and differences in concurrent chemotherapy regimens are additional potential confounders. Outcome definitions also vary across studies, with some using pCR and others using TRG [41, 58]. WES was performed on pretreatment biopsies to reduce potential bias from nCRT-induced shifts in mutation profiles [57], but prospective validation under standardized protocols and unified response criteria remains required.

From a clinical workflow perspective, the fusion framework is suited to pretreatment decision-making before nCRT and is positioned as an adjunctive tool. WSI derived from routine biopsy H&E slides can support initial risk assessment; when available, adding genomic information can improve identification of high-risk subgroups, particularly cases carrying KRAS^G12V or KRAS^A146T. Outputs are intended to inform MDT discussion rather than replace imaging-based staging or clinical judgement. Given differences in modality availability and feature dimensionality, coverage was defined to quantify the proportion of patients receiving a definitive prediction under given data conditions: pathology-only models approach near-universal coverage, whereas incorporating WES can improve accuracy in subsets, balancing performance with accessibility. An “uncertain” prediction option was implemented as a safe-abstention mechanism: when the discriminative signal is insufficient, and confidence does not meet prespecified thresholds, the model does not force a class label. Clinically, this corresponds to continuing standard management or considering additional testing, reducing the risk of overconfident outputs and improving interpretability.

At the mechanistic level, exploratory multiomics analyses and orthogonal functional assays in CRC model systems were consistent with KRAS^G12V and KRAS^A146T being associated with altered regulation of DNA damage repair pathways. WES and exploratory multiomics comparisons suggested repair-associated signals in KRAS-mutant contexts; across dual-reporter systems and in vitro/in vivo assays, both variants were associated with increased NHEJ and HR activity and reduced postirradiation DNA damage markers. Prior work has described endogenous genotoxic stress in KRAS-mutant cells and dependence on checkpoint signaling to maintain genome stability [64, 65]; the inhibition of metabolic pathways together with the ATR/CHK1 axis can induce DNA damage accumulation and apoptosis [66]. These mechanistic lines are directionally consistent with the repair-enhancement/tolerance phenotypes observed here, but experimental data do not directly translate into the magnitude of clinical benefit.

Codons 12 and 13 represent common KRAS hotspots [67, 68]. In this WES dataset, four mutations occurred at codon 12, six at codon 13 and two at codon 146; all cases carrying KRAS^G12V or KRAS^A146T were classified as nCRT nonresponders. Prior studies reported poorer nCRT response in codon 12/13 mutants [40], and KRAS^A146T in LARC has been reported to align with failure to achieve pCR [42, 58]. Exploratory multiomics comparisons (KRAS^G12V, KRAS^A146T and KRAS^WT) suggested codon-dependent differences in pathway emphasis. Transcriptomic analysis indicated enrichment of DNA repair-related pathways in KRAS^G12V, whereas KRAS^A146T showed a distinct enrichment pattern. Phosphoproteomic findings were interpreted cautiously and used mainly to support exploratory pathway-level hypotheses. Functional assays, rather than the omics heat maps themselves, provided the principal support for the biological interpretation, as both variants were associated with reduced post-irradiation DNA damage markers and increased repair capacity, and dual-reporter systems showed increased NHEJ and HR activity. Prior cell-model studies recorded codon-dependent differences in apoptosis sensitivity despite shared MAPK activation [69], and KRAS^G12V has been associated with stronger cetuximab tolerance in CRC cells [70]. In CRC xenograft models, KRAS-variant tumors grew faster and showed lower sensitivity to chemoradiotherapy than KRAS^KD controls, consistent with the in vitro patterns of enhanced repair and reduced damage signaling; effect sizes in clinical populations require evaluation in larger independent cohorts under standardized regimens.

Tissue biopsy in LARC is constrained by sampling risk, spatial heterogeneity, and turnaround time. Circulating free DNA (cfDNA)-based liquid biopsy enables the noninvasive and dynamic monitoring of KRAS variants [71, 72]. Incorporating cfDNA into pretreatment assessment and response monitoring could increase screening coverage and provide data for dynamically updated prediction models; incremental value requires prospective evaluation.

Several limitations apply. Model development and evaluation relied mainly on a single-center retrospective cohort of limited size; although an external WES cohort was included, the absence of external WSI data precluded independent validation of the full WES + WSI fusion model, and generalizability will require multicenter prospective confirmation. Current reporting emphasizes discrimination metrics (AUC/ACC) and does not include threshold-dependent performance measures or calibration analyses, which limits assessment of deployment thresholds and clinical risk interpretation. Incomplete overlap between WES and WSI within the internal cohort, together with the absence of WSI in the external cohort, may have introduced availability-driven selection bias, and weakly supervised MIL/ResNet architectures remain susceptible to overfitting in small-sample settings.

At the molecular level, KRAS^G12V and KRAS^A146T were low-frequency codon events with limited numbers of carriers, restricting robust multivariable adjustment and reliable estimation of effect sizes and confidence intervals; the clinical associations should therefore be regarded as exploratory. Certain omics heat map visualizations are included for descriptive purposes only and should be interpreted cautiously. Mechanistic experiments incorporated KRAS G13D as a comparator but did not systematically examine a broader panel of KRAS alleles, including variants observed in responder-associated contexts; accordingly, the present data do not definitively distinguish strict allele-specific resistance from broader KRAS-mediated treatment tolerance with allele-dependent magnitude. Finally, the clinical observations derive from patients with LARC, whereas much of the mechanistic support derives from CRC models. Subcutaneous CRC xenografts do not recapitulate rectal anatomy, the pelvic microenvironment, or immune effects associated with fractionated radiotherapy; these models inform biological plausibility but should not be interpreted as implying complete biological equivalence between colorectal cancer and rectal cancer.

Two types of evidence are distinguished. Clinically, KRAS^G12V/KRAS^A146T were observed only in nCRT nonresponders, without establishing causality or estimating effect magnitude. Mechanistically, cell and animal models showed enhanced DSB repair consistent with tolerance phenotypes, but experimental findings do not directly translate to clinical benefit or harm. Both lines of evidence are hypothesis-generating and require confirmation in larger prospective cohorts with standardized regimens and unified response endpoints.

Conclusions

In the LARC clinical cohort, KRAS^G12V and KRAS^A146T were observed to co-occur with nCRT nonresponse, and chemoradiotherapy-tolerance–related phenotypes were evaluated in cell-based and animal models. Exploratory multiomics analyses and functional assays were consistent with these codon variants being associated with increased activity of DSB repair-related pathways, accompanied by increased NHEJ and HR output, reduced postirradiation DNA damage markers and decreased apoptosis; xenograft models similarly showed tumor growth phenotypes consistent with attenuated treatment response. These findings provide a testable mechanistic explanation for the clinical association but do not constitute confirmatory evidence of clinical causality or effect size.

In parallel, a fusion deep-learning framework built on pretreatment WES and WSI enabled risk stratification of nCRT response in this cohort and offered an operational tradeoff among accuracy, coverage, and resource use. This framework may serve as a research-oriented reference for pretreatment risk stratification in research and exploratory settings; its generalizability, threshold deployment, and incremental clinical value require further validation in independent, multicenter prospective cohorts with paired WES + WSI data.

Abbreviations

ACC

Accuracy

Amp

Amplifications

ANOVA

Analysis of variance

AUC

Area under the curve

BQSR

Base quality score recalibration

B

Base substitution patterns

CDS

Coding sequence

CFE

Cell-free extract

cfDNA

Circulating free DNA

COSMIC

Catalog of somatic mutations in cancer

CRC

Colorectal cancer

CRT

Chemoradiotherapy

CT

Computed tomography

CNV

Copy number variation

DDP

Cisplatin

DEGs

Differentially expressed genes

DI

Differential index

DSB

Double-strand break

del

Deletions

FDR

False discovery rate

FC

Fold change

FFPE

Formalin-fixed, paraffin-embedded

Fr

The mutation frequency in the Response (R) group

GO

Gene ontology

H&E

Hematoxylin and eosin

HR

Homologous recombination

Indels

Insertions/deletions

IVIS

In vivo imaging system

KEGG

Kyoto Encyclopedia of Genes and Genomes

KRAS^WT

Wild-type KRAS

KRAS^KD

KRAS knockdown

LARC

Locally advanced rectal cancer

Mean ± SD

Mean ± standard deviation

M

Mutation

MRI

Magnetic resonance imaging

MSI

Microsatellite instability

MSI-H

MSI-high

N⁺

Node-positive

NAC

Neoadjuvant chemotherapy

nCRT

Neoadjuvant chemoradiotherapy

NHEJ

Nonhomologous end joining

neo

Neomycin

NR

Nonresponders

P

Pathology

pCR

Pathological complete response

PVDF

Polyvinylidene difluoride

R

Responders

ROC

Receiver operating characteristic

S

Mutational signatures

SNVs

Single nucleotide variants

TME

Total mesorectal excision

TMB

Tumor mutational burden

TRG

Tumor regression grade

U

Ubiquitin-mediated proteolysis pathway

WB

Western blot

WES

Whole-exome sequencing

WSI

Whole-slide images

Author contributions

Hengchang Liu, Dechao Bu, Guanhua Yu, and Ran Wei contributed equally to this work. Hengchang Liu, Guanhua Yu, and Zheng Jiang conceived and designed the clinical and translational aspects of the study. Dechao Bu, Hui Jin, Yi Zhao, and Yixiao Liu designed and implemented the deep learning models and performed computational analyses. Ran Wei, Xu Guan, Zhixun Zhao, and Haipeng Chen collected clinical samples and pathological data. Hengchang Liu, Guanhua Yu, Yixiao Liu, and Xu Guan performed whole-exome sequencing analysis and genomic interpretation. Hui Jin, Dechao Bu, and Yi Zhao conducted multimodal data integration and model evaluation. Hengchang Liu, Guanhua Yu, and Ran Wei carried out in vitro and in vivo functional experiments. Hengchang Liu, Dechao Bu, and Guanhua Yu drafted the manuscript. Yi Zhao and Zheng Jiang critically revised the manuscript and supervised the study. All authors reviewed and approved the final manuscript.

Funding

This study was supported by the National Key Research and Development Program of China (2024ZD0520303, 2016YFC0905303) and State Key Laboratory of Traditional Chinese Medicine Syndrome Projects (SKLKY2025C0011)State Key Laboratory of Systems Medicine for Cancer（KF2422-93）

Data availability

All data generated or analyzed during this study are included in this article and/or its supplementary material files. Further inquiries can be directed to the corresponding author.

Declarations

Ethics approval and consent to participate

This study involving human participants was conducted in accordance with the Declaration of Helsinki and was approved by the Ethics Committee of the National Cancer Center/Cancer Hospital, Chinese Academy of Medical Sciences (approval no. NCC2023 C-248). Written informed consent was obtained from all individual participants included in the study. All animal experiments were performed in compliance with the International Council for Laboratory Animal Science (ICLAS) guidelines and were approved by the Institutional Animal Care and Use Committee (IACUC) of the National Cancer Center/Cancer Hospital, Chinese Academy of Medical Sciences (approval no. NCC2023A179).

Competing interests

The authors declare no competing interests.