Integration of radiogenomic features for early prediction of pathological complete response in patients with triple-negative breast cancer and identification of potential therapeutic targets

Triple-negative breast cancer (TNBC) is characterized by the negative expression of estrogen receptor, progesterone receptor, and human epidermal growth factor 2 [1]. Neoadjuvant chemotherapy (NAC) serves as the main therapy for TNBC patients. Patients who achieve pathological complete response (pCR) have a better prognosis than non-pCR patients [2, 3]. However, less than 40% of TNBC patients achieve pCR after the completion of a standard NAC regimen [4]. Therefore, early response prediction is valuable for clinical decisions regarding whether treatment modification is needed to obtain an improved response.

To date, neither traditional clinicopathological nor radiologic features can accurately predict the NAC response. The accuracy of ultrasonography for pCR is only approximately 70% [5]. The change in the maximum standardized uptake value of PET scans before the first two cycles of NAC has been used to predict treatment response, but the area under the curve (AUC) values were less than 0.70 [6]. MRI serves as the standard tool for the surveillance of breast cancer response, but none of the traditional radiologic methods can predict the treatment response with sufficient accuracy [7]. Thus, early and accurate predictions to NAC response remain challenging.

In recent years, the application of radiomics approaches, which are less invasive, has contributed to both cancer diagnosis [8] and response prediction [9‐11]. However, most studies have focused on baseline predictions; data on the predictive ability of radiomic features mid-treatment are still lacking. Radiogenomics links imaging features with the genetic profiles of tumors [12, 13]. The integration of complementary data generated from radiomics and genomics might further improve the performance of predictive models [13]. Studies have also expanded to include multiple genetic markers, which has helped to stratify TNBC patients for more tailored therapy [14, 15].

In this study, we attempted to establish a new model based on the light gradient boosting machine (LightGBM) method [16], which combines both radiomic and genomic data for the early prediction of the response to NAC in TNBC patients. Additionally, we explored the relationships between two potential mutation targets and drug resistance in vitro.

TNBC is characterized by its high rate of early recurrence and poor survival, with limited treatment options available [22]. In the neoadjuvant setting, pCR remains the most convincing substitute indicator of survival [2]. For potential non-pCR patients with an increased risk of chemoresistance, treatment modification (e.g., the addition of immunotherapy [23], nab-paclitaxel [24, 25] or nanoparticles for codelivery of antitumoral agents [26]) might be an option to increase the possibility of pCR. In this study, we aimed to establish a radiogenomic model using LightGBM to predict the treatment response as early as possible.

Quantitative imaging features that can be used to predict the response to treatment can expand the application of radiomics in routine clinical settings [11]. Braman et al. [27] proposed that both intratumoral and peritumoral features can contribute to response predictions, and that peritumoral features cannot be replaced by tumoral features. Therefore, we also included peritumoral radiomic features to avoid missing information from the surrounding microenvironment. The selected radiomic features, including first-order_variance, GLDM and GLCM parameters were consistent with those in other studies [10, 28‐30]. Here we newly identified a series of wavelet-related features that also contributed to pCR prediction. Wavelet features have been previously mentioned in the response prediction of colorectal cancer [31], but their application in breast cancer treatment response remain unexplored., which is worthy of further investigation. Moreover, the addition of clinicopathological data has been reported to increase the predictive ability compared to that of the model alone [32], but external validation has failed to confirm the superiority of the radiomic model over the clinical model [29]. In our study, we also failed to show a significant improvement in pCR prediction with the addition of clinical factors. Studies with larger sample sizes are needed to justify the addition of clinicopathological features into a radiomic/radiogenomic model. In the ACRIN 6657/I-SPY study [33], prediction based on changes in tumor volume showed the greatest relative benefit after 1 cycle of anthracycline-based NAC, with an AUC of 0.70. Another recent study [30] revealed excellent predictive power for 3-year recurrence (AUC = 0.93) in TNBC, but both pre- and post-NAC features were used. Our radiogenomic model enabled the prediction of pCR even before the start of NAC (at baseline), with a high AUC value of 0.87 in the validation set, thus potentially supporting earlier treatment modification. Patients with a poor response might not need to complete additional cycles prior to treatment alteration.

In addition to the macroscopic and high-throughput characteristics from radiology, we tried to add genomic data to enhance the predictive value, discover novel biomarkers and identify potential genomic mechanisms associated with the drug-resistant phenotype. For response predictions, our results revealed the excellent performance of the radiogenomic model after the addition of 5 VAF features, with an AUC of To identify potentially actionable targets, we tested whether the two high-frequency mutation spots (MED23 p.P394H and REL p.D268E) were associated with chemotherapy resistance. Mediator (MED) is a multisubunit complex that conveys signals to the basal transcription machinery [34]. Mediator complex subunit 23 (MED23) is a subunit that plays a crucial role in alternative mRNA processing [35], cell differentiation [36] and tumorigenesis [37]. Overexpression of MED23 was confirmed to promote tumorigenesis in NSCLC [38] and hepatocellular carcinoma [39]. Conversely, overexpression of MED23 in esophageal squamous cell carcinoma dramatically inhibited cell growth [40]. However, few studies have investigated the role of MED23 in breast cancer. We found that the cell viability and IC50 values of MED23 p.P394H-mutant cells increased after epirubicin treatment. This trend was opposite to that observed after MED23 was knocked down, indicating that P394H might be a pathogenic mutation.

Epirubicin forms complexes with DNA by intercalation between base pairs, which leads to the formation of free radicals and inhibits the catalytic activity of DNA topoisomerase II [41]. Conversely, DNA damage repair is an important contributor to epirubicin resistance [42]. Significantly reduced levels of p-ATM and γ-H2A.X and p-CHK2 were observed at multiple time points post-mutation, indicating that epirubicin resistance after P394H mutation might manifest through regulation of the p-ATM- γ-H2A.X- p-CHK2 pathway. HR is one of the most important mechanisms for the repair of anthracycline-DNA adducts [43]. Our results showed that the P394H mutation might affect HR repair, thus further inducing epirubicin resistance. Cancer cells with efficient DNA damage repair machinery might be able to overcome the cytotoxicity of anthracyclines. Further experiments are needed to explain the underlying mechanisms and verify whether MED23 p.P394H or HR inhibitors could help to reverse epirubicin resistance. Another high-frequency mutation, REL p.D268E, failed to elicit any significant effect on drug resistance, possibly because of the limited sample size. The relatively high-frequency (2 patients) of this mutation in this cohort might not reflect its true frequency in the general TNBC population.

Despite the excellent predictive value of the radiogenomic model, several limitations still need to be considered. First, actionable somatic mutations are not frequent events in TNBC patients, so bias existed in this relatively small cohort. Further validation in external populations is needed. Additionally, the epirubicin-resistant phenotype caused by MED23 P394H and its underlying mechanisms need to be fully explored in clinical samples in the future. Second, this was a single-institution study, the chemotherapy regimens were not unified, and heterogeneity existed due to the use of different MRI machines. Third, the genomic data were sequenced using core-needle biopsy tissue, which might be affected by paracancerous tissue. Finally, the genomic analyses only covered mutation data; other aspects, including copy number variation and quantitative expression, were not included.

Despite these limitations, in this work, an excellent radiogenomic model was constructed, with an AUC as high as 0.87, that could predict pCR in the TNBC population prior to NAC administration. Furthermore, we are the first to propose that the MED23 p.P394H mutation might cause epirubicin resistance and to explore this phenotype. The underlying mechanisms will be investigated and validated in future experiments.

The proposed radiogenomic model has the potential to accurately predict pCR before NAC in TNBC patients. The resistance to epirubicin observed after MED23 p.P394H mutation occurs might be associated with HR repair through regulation of the p-ATM-γ-H2A.X-p-CHK2 pathway.