The prediction of pCR and chemosensitivity for breast cancer patients using DLG3, RADL and Pathomics signatures based on machine learning and deep learning

Highlights

• •The nomogram based on multiomics signatures demonstrated the better predictive ability than the clinical TNM staging system and single RADL signature in predicting pCR.
• •DLG3 was proved to be related with the sensitivity of docetaxel and epirubicin.
• •Patients in the low RADL_Signature_DLG3 and PATHO_Signature_DLG3 groups had lower DLG3 IHC scores, indicating that these patients are more suitable for the TEC regimen.
• •The RADL and PATHO signatures were also been proved to be the independent prognostic factors for DFS and OS in breast cancer patients who underwent NAC.

Abstract

Background

Limited studies have investigated the predictive value of multiomics signatures (radiomics, deep learning features, pathological features and DLG3) in breast cancer patients who underwent neoadjuvant chemotherapy (NAC). However, no study has explored the relationships among radiomic, pathomic signatures and chemosensitivity. This study aimed to predict pathological complete response (pCR) using multiomics signatures, and to evaluate the predictive utility of radiomic and pathomic signatures for guiding chemotherapy selection.

Methods

The oncogenic function of DLG3 was explored in breast cancer cells via DLG3 knockdown. Immunohistochemistry (IHC) was used to evaluate the relationship between DLG3 expression and docetaxel/epirubin sensitivity. Machine learning (ML) and deep learning (DL) algorithms were used to develop multiomics signatures. Survival analysis was conducted by K-M curves and log-rank. Multivariate logistic regression analysis was used to develop nomograms.

Results

A total of 311 patients with malignant breast tumours who underwent NAC were retrospectively included in this multicentre study. Multiomics (DLG3, RADL and PATHO) signatures could accurately predict pCR (AUC: training: 0.900; testing: 0.814; external validation: 0.792). Its performance is also superior to that of clinical TNM staging and the single RADL signature in different cohorts. Patients in the low DLG3 group more easily achieved pCR, and those in the high RADL Signature_pCR and PATHO_Signature_pCR (OR = 7.93, 95 % CI: 3.49–18, P < 0.001) groups more easily achieved pCR. In the TEC regimen NAC group, patients who achieved pCR had a lower DLG3 score (4.00 ± 2.33 vs. 6.43 ± 3.01, P < 0.05). Patients in the low RADL_Signature_DLG3 and PATHO_Signature_DLG3 groups had lower DLG3 IHC scores (P < 0.05). Patients in the high RADL signature, PATHO signature and DLG3 signature groups had worse DFS and OS.

Conclusions

Multiomics signatures (RADL, PATHO and DLG3) demonstrated great potential in predicting the pCR of breast cancer patients who underwent NAC. The RADL and PATHO signatures are associated with DLG3 status and could help doctors or patients choose proper neoadjuvant chemotherapy regimens (TEC regimens). This simple, structured, convenient and inexpensive multiomics model could help clinicians and patients make treatment decisions.

Introduction

Breast cancer is the most common malignant tumour in females worldwide[1]. Although the 5-year relative survival rate has improved to 90 %, patients diagnosed with stage IV breast malignant tumours in the United States still have a lower survival rate of 28 %[2]. For locally advanced or advanced breast cancer patients, neoadjuvant chemotherapy (NAC) could be a good choice, and patients who achieve pathological complete response (pCR) have an improved survival rate[3]. However, for different reasons, such as tumour heterogeneity that can be caused by distinct genetic alterations, which leads to different subtypes of patients, making breast cancer patients with similar clinical characteristics response differently to NAC [4,5].Therefore, accurate individualized predictions of treatment efficacy and survival still need to be improved.

Radiomics is the process of extracting quantitative attributes from a specific region of interest (ROI) and has been widely applied to predict molecular subtypes or distinguish malignant breast lesions through machine learning (ML) or deep learning (DL) algorithms[6]. However, the features that been extracted by radiomics were from macroscopic perspective images, lacking of cytological information. In contrast, pathomics could provide microstructure information through identifying tumor areas. It uses pathological images to obtain quantitative features that are used to characterize different phenotypes of tissues, and these signatures are subsequently applied to determine diagnosis or predict outcome events[7]. Therefore, combing radiomics and pathomics signature may be a good choice to comprehensively predict outcome. Although radiomics or pathomics have been applied in the diagnosis, classification and prognosis of breast cancer[[8], [9], [10], [11]], most of them have focused on early-stage patients or disease-free survival (DFS). No study has explored pCR, chemosensitivity or other prognostic information in breast cancer patients receiving NAC using simultaneous radiomic and pathomic signatures. And there is no study to select suitable regimens for patients using radiomics or pathomics. Therefore, we designed a multi-omics model including radiomcis and pathomics signatures to predict pCR.

Discs large homologue 3 (DLG3) is a member of the membrane-associated guanylate kinase (MAGUK) family[12]. Compared with those in normal tissues, DLG3 mRNA levels are greater in many cancer tissues, including breast cancer tissues[13]. A previous study demonstrated that patients with high DLG3 expression in the breast showed worse survival[14]. In gastric cancer, Rui Jiang et al.[15] found that DLG3 was correlated with sensitivity of some chemotherapeutic drugs (such as paclitaxel). However, its value in predicting pCR is not clear. And its prognostic ability and relationship with chemosensitivity have not been explored in breast cancer patients receiving NAC.

Therefore, after proving the relationship between DLG3 and chemotherapy drugs, using radiomics and pathomics to predict DLG3 status may be helpful in selecting the appropriate chemotherapy regimen.

This study aimed to predict pCR using DLG3, RADL and PATHO signatures, as well as to compare their performance with single RADL signature and the clinical TNM staging system. And to investigate the biological function and relationship among RADL, PATHO, DLG3 and chemosensitivity in breast cancer. Furthermore, it intends to explore the underlying prognostic ability of the DLG3, RADL and PATHO signatures.

Methods

Research design and included patients

This study developed three signatures (DLG3, radiomics and pathomics) to construct a multiomics model for the prediction of pCR, chemosensitivity, DFS and OS. A total of 311 breast cancer patients who underwent NAC or surgery were included in this study. Of 215 patients, between 2012 and 2017, were from Harbin Medical University Cancer Hospital (HMUCH), 60 patients, between 2018 and 2023, were from the second cancer hospital of Heilongjiang province (NKYY), and 36 patients, between 2021 and 2023 were form the first people's hospital of Xiangtan city (XTYY). The inclusion criteria were as follows: 1) invasive breast cancer, 2) complete radiological and pathological images, 3) complete follow-up data, and 4) underwent NAC and surgery.

This multicentre study was conducted in accordance with the Declaration of Helsinki and subsequently amended versions and was approved by the ethics committees of the participating hospitals. Each patient signed an informed consent (including secondary data utilization) before undergoing any treatment. All the images, electronic data and formalin-fixed paraffin-embedded (FFPE) tissues were retrospectively collected and were noninvasive.

The details are shown in Fig. 1.

Fig. 1.

The flow chart of patient selection and study design.

Neoadjuvant chemotherapy regimens

Based on immunohistochemistry (IHC) and patients' preferences, NAC regimens were predominantly anthracycline- or taxane-based. AC-T: A 60 mg/m², C 600 mg/m², and docetaxel (T) 90 mg/m²; AC-TH: A 60 mg/m², C 600 mg/m², T 75 mg/m², and Herceptin (H) first dose 8 mg/kg, then 6 mg/kg; and TAC: T 75 mg/m², A 50 mg/m², and C 500 mg/m²; EC: epirubicin (E) 100 mg/m² and C 600 mg/m²; TEC: T 80–100 mg/m², E 90–100 mg/m² and C 600 mg/m². There are 21 days in every cycle. Only a few of patients had used Herceptin because of financial burden.

Follow‑up and outcomes

After surgery, all patients received postoperative follow-up in outpatient or inpatient care every 3 months for the first two years, every 6 months for the next three years, and annually thereafter. The follow-up period lasted until November 2023 or until the date of death from any cause. pCR was defined as the absence of invasive cancer but the presence of carcinoma in situ in the breast. OS was defined as the time from surgery to death from any cause or the date of the last follow-up visit. DFS was defined as the time from the date of surgery to the date of local recurrence or distant metastases, death from any cause, or the last follow-up.

The primary outcome was the prediction of pCR and chemosensitivity in breast cancer patients. The secondary outcomes were DFS and OS.

Public data access and analysis, in vitro experiment, RADL and PATHO feature extraction, signature development

The details are shown in supplement materials.

Statistical analysis

All analyses were conducted with GraphPad Prism (version 9.0), R software (version 4.1.2) and Python (version v.3.7.12). The thresholds of the RADL and PATHO signatures were determined by the Youden index. Student's t-test was applied to assess the differences between groups. All experiments were performed independently in triplicate. Survival analysis was conducted by Kaplan–Meier curves and the log-rank test. Based on the multivariate Cox or logistic regression analysis, nomogram models were developed. A graphic analysis was conducted on the differences between the actual and predicted probabilities obtained from the nomograms. Additionally, the predictive ability of the nomogram and clinical TNM stage were compared by area under the curve (AUC) analysis and decision curve analysis (DCA). All the statistical tests were two-sided, and P < 0.05 was considered to indicate statistical significance.

Results

The baseline of different cohorts

The mean age was 50.84 years in the HMUCH cohort and 54.7 years in the external validation cohort. In both cohorts, similar cT, cN, ER, PR, HER2, Ki-67, RADL and PATHO signature statuses were observed. Due to the limited follow-up time of patients in the external validation cohort, RADL and PATHO signatures related to DFS and OS were not developed. The pCR rate was 14.0 % in the HMUCH cohort and 22.9 % in the external validation cohort. A total of 56.3 % and 42.7 % of patients in the HMUCH cohort and the external validation cohort, respectively, exhibited increased DLG3 expression (Table 1).

Table 1.

The baseline of included breast cancer patients.

Characteristics	Overall	HMUCH Cohort	External Validation Cohort
	N = 311 (%)	n = 215 (%)	n = 96 (%)
Age (mean (SD))	50.84 (9.13)	49.12 (9.07)	54.70 (8.07)
Age(%)
≤51	169 (54.3)	135 (62.8)	34 (35.4)
>51	142 (45.7)	80 (37.2)	62 (64.6)
cT (%)
1/2	238 (76.5)	161 (74.9)	77 (80.2)
3/4	73 (23.5)	54 (25.1)	19 (19.8)
cN (%)
0	31 (10.0)	16 (7.4)	15 (15.6)
1–3	280 (90.0)	199 (92.6)	81 (84.4)
ER (%)
negative	117 (37.6)	79 (36.7)	38 (39.6)
positive	194 (62.4)	136 (63.3)	58 (60.4)
PR (%)
negative	170 (54.7)	116 (54.0)	54 (56.2)
positive	141 (45.3)	99 (46.0)	42 (43.8)
HER2 (%)
negative	188 (60.5)	126 (58.6)	62 (64.6)
positive	123 (39.5)	89 (41.4)	34 (35.4)
Ki-67 (%)
≤14 %	69 (22.2)	59 (27.4)	10 (10.4)
>14 %	242 (77.8)	156 (72.6)	86 (89.6)
pCR (%)
no	259 (83.3)	185 (86.0)	74 (77.1)
yes	52 (16.7)	30 (14.0)	22 (22.9)
DLG3 (%)
low expression	149 (47.9)	94 (43.7)	55 (57.3)
high expression	162 (52.1)	121 (56.3)	41 (42.7)
RADL_Signature_DFS (%)
low	131 (60.9)	131 (60.9)	none
high	84 (39.1)	84 (39.1)	none
Patho_Signature _DFS (%)
low	103 (47.9)	103 (47.9)	none
high	112 (52.1)	112 (52.1)	none
RADL_Signature _OS (%)
low	134 (62.3)	134 (62.3)	none
high	81 (37.7)	81 (37.7)	none
Patho_Signature _OS (%)
low	114 (53.0)	114 (53.0)	none
high	101 (47.0)	101 (47.0)	none
RADL_Signature _pCR (%)
low	220 (70.7)	157 (73.0)	63 (65.6)
high	91 (29.3)	58 (27.0)	33 (34.4)
Patho_Signature _pCR (%)
low	192 (61.7)	138 (64.2)	54 (56.2)
high	119 (38.3)	77 (35.8)	42 (43.8)
RADL_Signature _DLG3 (%)
low	100 (32.2)	65 (30.2)	35 (36.5)
high	211 (67.8)	150 (69.8)	61 (63.5)
Patho_Signature _DLG3 (%)
low	277 (89.1)	191 (88.8)	86 (89.6)
high	34 (10.9)	24 (11.2)	10 (10.4)

Univariate and multivariate logistic regression analyses for pCR

According to the univariate analysis, ER, PR, HER2, DLG3, RADL_Signature_pCR and PATHO_Signature_pCR were significantly related to pCR (P < 0.05). These characteristics were included in the multivariate analysis, and the results demonstrated that patients in the low DLG3 subgroup more easily achieved pCR (OR=0.33, 95 % CI: 0.15–0.71; P = 0.004). Patients in the high RADL Signature_pCR (OR=12, 95 % CI: 5.44–26.46, P < 0.001) and PATHO_Signature_pCR (OR=7.93, 95 % CI: 3.49–18, P < 0.001) groups achieved pCR more easily (Table 2).

Table 2.

Univariate and multivariate analysis for pCR.

Characteristics	Univariate analysis		Multivariate analysis
	OR (95 % CI)	P value	OR (95 % CI)	P value
Age
≤51	1 (reference)
>51	0.93 (0.51–1.7)	0.821
cT
1/2	1 (reference)
3/4	1.11 (0.55–2.21)	0.776
cN
0	1 (reference)
1–3	0.54 (0.23–1.27)	0.158
ER
negative	1 (reference)
positive	0.49 (0.27–0.9)	0.021	0.73 (0.25–2.07)	0.55
PR
negative	1 (reference)
positive	0.48 (0.25–0.9)	0.023	0.39 (0.13–1.17)	0.093
HER2
negative	1 (reference)
positive	1.83 (1.01–3.34)	0.047	1.72 (0.76–3.92)	0.195
Ki-67
≤14 %	1 (reference)
>14 %	0.94 (0.46–1.91)	0.866
DLG3
low expression	1 (reference)
high expression	0.47 (0.25–0.86)	0.015	0.33 (0.15–0.71)	0.004
RADL_Signature_DFS
low	1 (reference)
high	1.05 (0.48–2.3)	0.91
Patho_Signature _DFS
low	1 (reference)
high	0.57 (0.26–1.24)	0.156
RADL_Signature _OS
low	1 (reference)
high	0.95 (0.43–2.12)	0.902
Patho_Signature _OS
low	1 (reference)
high	1.15 (0.53–2.49)	0.721
RADL_Signature _pCR
low	1 (reference)
high	9.36 (4.79–18.31)	<0.001	12 (5.44–26.46)	<0.001
Patho_Signature _pCR
low	1 (reference)
high	5.96 (3.06–11.62)	<0.001	7.93 (3.49–18)	<0.001
RADL_Signature _DLG3
low	1 (reference)
high	1.2 (0.63–2.32)	0.576
Patho_Signature _DLG3
low	1 (reference)
high	0.28 (0.07–1.22)	0.091

Nomogram based on different omics signatures for pCR and comparison of their predictive performance

Characteristics with P < 0.05 in the multivariate logistic regression analysis were included in the nomogram. Interestingly, the selected characteristics were the multiomics signatures (DLG3, RADL and PATHO). Then, a nomogram was established based on the multiomics signatures (Fig. 2A). The C-index of this model in the training set was 0.900, that in the test set was 0.814, and that in the external validation set was 0.792; similar results (0.886, 0.810, 0.792) were observed when bootstrapping was applied.

Fig. 2.

Nomogram to predict the pCR of breast cancer patients undergoing neoadjuvant chemotherapy. A nomogram was generated based on multiomics signature to predict pCR (A). The predictive performance comparison of multi-omics signature and clinical TNM staging in train (B), test (C) and external validation (D) sets. The pCR rate of the breast cancer patients predicted by the nomogram were consistent with the actual observed values in train (E), test (F) and external validation (G) sets. Decision curve analyses (DCA) for the prognostic models of the multiomics signatures and clinical TNM staging system in train (H), test (I) and external validation (J) sets. The nomogram maps predict the probabilities onto the points on a scale from 0 to 100 and can be interpreted by adding the points together that correspond to the predicted probability. The total points were converted into the probabilities of pCR for breast cancer patients.

In order to further understand the value of multiomics signature, we compared the predictive ability of multiomics signature, single RADL signature and AJCC TNM staging. The multiomics signature achieved good predictive ability and presented a better predictive performance than did the clinical AJCC TNM staging system and RADL signature in the training (AUC: 0.900 vs. 0.539, P < 0.0001; 0.900 vs. 0.816, P = 0.0019; Fig. 2B, Table 3), test (AUC: 0.814 vs. 0.505, P = 0.0055; 0.814 vs. 0.791, P = 0.5735; Fig. 2C, Table 3) and external validation (AUC: 0.792 vs. 0.432; P = 0.0110; 0.792 vs. 0.660, P = 0.0004; Fig. 2D, Table 3) cohorts for pCR. The pCR predictions of nomogram were almost consistent with the actual observations in the training cohort (Fig. 2E), test cohort (Fig. 2F), and external validation cohort (Fig. 2G). According to the DCA, both the multiomics signature and clinical AJCC TNM staging models showed a positive net benefit. The nomogram based on multiomics signatures demonstrated better clinical applicability for pCR prediction in the training (AUC: 0.041 vs. 0.002, Fig. 2H), test (AUC: 0.013 vs. 0.005, Fig. 2I) and external validation (AUC: 0.029 vs. 0.012, Fig. 2J) cohorts.

Table 3.

The predictive performance comparison between multi-omics signatures, clinical TNM staging and RADL signature for pCR.

Parameters	Difference of AUC	95 % CI	Z statistic	P value
Multi-omics vs. clinical TNM staging system
Train	0.361	0.215 - 0.508	4.834	< 0.0001
Test	0.309	0.0909 - 0.528	2.776	0.0055
External validation	0.360	0.0514 - 0.396	2.544	0.0110
Multi-omics vs. RADL signature
Train	0.0843	0.0312 - 0.138	3.109	0.0019
Test	0.0238	−0.0591 - 0.107	0.563	0.5735
External validation	0.132	0.0590 - 0.205	3.547	0.0004

Relationship between DLG3 and chemosensitivity

It is known that patients who achieve pCR are more sensitive to NAC. Multivariate analysis revealed that DLG3 was an independent predictor of pCR. Therefore, we further explored the relationship between DLG3 expression and chemosensitivity.

The results of bioinformatic analysis showed that patients with low DLG3 expression were more sensitive to docetaxel and epirubicin (Figure S1, P < 0.0001). Therefore, we further explored the function of DLG3 in vitro experiments. The results demonstrated that the expression of DLG3 was greater in breast cancer cells than in normal breast cells and that si-#2 and si-#3 presented greater knockdown efficiency (Fig. 3A). In addition, the results demonstrated that the proliferation and migration abilities of DLG3-knockdown cancer cells were lower than those of negative control cells (Fig. 3B-D, P < 0.05). To validate the relationship between DLG3 and chemotherapy drugs, we examined the half-maximal inhibitory concentrations (IC₅₀) of docetaxel and epirubincin in different cells. The IC₅₀ values of docetaxel (812: 27.9 vs. 7.30 vs. 4.30 nMol/L; 1954: 51.1 vs. 34.2 vs. 17.3 nMol/L; Fig. 3E, P < 0.05) and epirubincin (812: 742.2 vs. 583.6 vs. 492.8 nMol/L; 1954: 419.2 vs. 308.3 vs. 171.6 nMol/L; Fig. 3E, P < 0.05) were lower in the DLG3-knockdown cells than in the negative control cells.

Fig. 3.

DLG3 knockdown efficiency in different sequence (A). Knockdown of DLG3 expression inhibits proliferation (B), migration (C, D) of UACC-812 and HCC-1954 cells. Knockdown of DLG3 expression increased sensitivity of UACC-812 and HCC-1954 cells to docetaxel and epirubicin (E). The DLG3 protein expression of two breast cancer patients who underwent NAC in Harbin Medical University cancer hospital and patients in pCR group showed a lower DLG3 protein expression breast cancer patients who underwent NAC with TEC regimen(F). Scale bar, 100 and 250 μm. Data are presented as means from three independent experiments ± S.D. *P < 0.05, **P < 0.01, ***P < 0.001, ****P < 0.0001.

We also examined DLG3 expression in 48 breast cancer patients who underwent NAC with the TEC regimen. The IHC staining score of each FFPE tissue sample was calculated according to the intensity and extent of staining. DLG3 expression was then scored as low (≤4) or high (>4). The results indicated that patients who achieved pCR had lower DLG3 scores (4.00±2.33 vs. 6.43±3.01; Fig. 3F, P < 0.05).

Relationship between DLG3 and different characteristics

The above results showed that patients with low DLG3 expression were more sensitive to docetaxel and epirubicin. Therefore, we explored the characteristics related to DLG3. RADL_Signature_DLG3 and PATHO_Signature_DLG3 were significantly associated with DLG3 (P < 0.05, Table 4). Then, we compared the difference of IHC scores, patients in the low RADL_Signature_DLG3 and PATHO_Signature_DLG3 groups had lower DLG3 IHC scores (P < 0.05, Fig. 4).

Table 4.

The relationship between DLG3 and different characteristics.

Characteristics	Overall	low DLG3	high DLG3	P value
	N = 311 (%)	n = 149 (%)	n = 162 (%)
Age (%)
≤51	169 (54.3)	81 (54.4)	88 (54.3)	1
>51	142 (45.7)	68 (45.6)	74 (45.7)
cT (%)
1/2	238 (76.5)	115 (77.2)	123 (75.9)	0.899
3/4	73 (23.5)	34 (22.8)	39 (24.1)
cN (%)
0	31 (10.0)	18 (12.1)	13 (8.0)	0.316
1–3	280 (90.0)	131 (87.9)	149 (92.0)
ER (%)
negative	117 (37.6)	55 (36.9)	62 (38.3)	0.897
positive	194 (62.4)	94 (63.1)	100 (61.7)
PR (%)
negative	170 (54.7)	79 (53.0)	91 (56.2)	0.657
positive	141 (45.3)	70 (47.0)	71 (43.8)
HER2 (%)
negative	188 (60.5)	95 (63.8)	93 (57.4)	0.304
positive	123 (39.5)	54 (36.2)	69 (42.6)
Ki-67 (%)
≤14 %	69 (22.2)	36 (24.2)	33 (20.4)	0.505
>14 %	242 (77.8)	113 (75.8)	129 (79.6)
RADL_Signature_DFS (%)
low	131 (60.9)	57 (60.6)	74 (61.2)	1
high	84 (39.1)	37 (39.4)	47 (38.8)
Patho_Signature_DFS (%)
low	103 (47.9)	48 (51.1)	55 (45.5)	0.497
high	112 (52.1)	46 (48.9)	66 (54.5)
RADL_Signature_OS (%)
low	134 (62.3)	60 (63.8)	74 (61.2)	0.795
high	81 (37.7)	34 (36.2)	47 (38.8)
Patho_Signature_OS (%)
low	114 (53.0)	55 (58.5)	59 (48.8)	0.199
high	101 (47.0)	39 (41.5)	62 (51.2)
RADL_Signature_pCR (%)
low	220 (70.7)	104 (69.8)	116 (71.6)	0.822
high	91 (29.3)	45 (30.2)	46 (28.4)
Patho_Signature_pCR (%)
low	192 (61.7)	92 (61.7)	100 (61.7)	1
high	119 (38.3)	57 (38.3)	62 (38.3)
RADL_Signature_DLG3 (%)
low	100 (32.2)	68 (45.6)	32 (19.8)	<0.001
high	211 (67.8)	81 (54.4)	130 (80.2)
Patho_Signature_DLG3 (%)
low	277 (89.1)	142 (95.3)	135 (83.3)	0.001
high	34 (10.9)	7 (4.7)	27 (16.7)

Fig. 4.

The relationship between RADL (A), PATHO (B) signatures and DLG3 expression score.

These results indicated that patients with low RADL_Signature_DLG3 and PATHO_Signature_DLG3 expression are sensitive to docetaxel and epirubicin.

The predictive ability of different omics signatures and DLG3 for DFS and OS

Previous studies have shown that the radiomics signature, deep learning signature and DLG3 are correlated with DFS and OS[14,16]. However, their predictive value has not been explored in breast cancer patients who underwent NAC, and the predictive value of the PATHO signature is not clear. Therefore, we investigated the ability of the RADL signature, PATHO signature and DLG3 to predict DFS and OS in breast cancer patients who underwent NAC.

The cut-off values of the RADL and PATHO signatures for DFS were 0.364 and 0.5, respectively, and those for OS were 0.292 and 0.271, respectively. Patients were subsequently classified into high and low groups. Patients in the high RADL signature, PATHO signature and DLG3 groups demonstrated a shorter mean DFS (85.2 vs. 54.2 months, P < 0.0001; 81.2 vs. 65.5 months, P = 0.0015; 78.9 vs. 68.4 months, P = 0.035; Fig. 5A, B, C) and mean OS (98.8 vs. 67.6 months, P < 0.0001; 97.5 vs. 74.9 months, P < 0.0001; and 91.9 vs. 82.9 months, P = 0.043; Fig. 5D, E, F).

Fig. 5.

Kaplan‒Meier curves of different signature risk groups for DFS (RADL, A; Patho, B; DLG3, C) and OS (RADL, D; Patho, E; DLG3, F).

Moreover, we explored the prognostic value of these signatures. Multivariate analysis revealed that patients in the high DLG3 expression (HR=1.45, 95 % CI: 0.94–2.25, P = 0.0957), RADL_Signature_DFS (HR=2.54, 95 % CI: 1.56–4.15, P < 0.001), PATHO_Signature_DFS (HR=2.02, 95 % CI: 1.31–3.12, P = 0.0015) and PATHO_Signature_OS (HR=1.79, 95 % CI: 1.16–2.77, P = 0.0091) groups had a worse DFS (Table S2). Similarly, patients in the cT3/4 (HR=2.17, 95 % CI: 1.23–3.83, P = 0.0079), high DLG3 expression (HR=1.51, 95 % CI: 0.84–2.72, P = 0.1689), PATHO_Signature_DFS (HR=2.07, 95 % CI: 1.15–3.71, P = 0.015), RADL_Signature_OS (HR=4.52, 95 % CI: 2.19–9.35, P < 0.001) and PATHO_Signature_OS (HR=3.07, 95 % CI: 1.57–6.01, P = 0.001) groups had a worse OS (Table S3).

All the above results presented that RADL and PATHO signatures are independent predictors for DFS and OS.

Discussion

In this multicentre study, multiomics signatures (including DLG3, RADL and PATHO signatures) based on deep learning and machine learning models were firstly proved to be independent predictors of pCR. Then, the nomogram based on multiomics signatures demonstrated the better predictive ability than the clinical TNM staging system and single RADL signature in predicting pCR. Subsequently, DLG3 was proved to be related with the sensitivity of docetaxel and epirubicine. And patients in the low RADL_Signature_DLG3 and PATHO_Signature_DLG3 groups had lower DLG3 IHC scores, indicating that these patients are more suitable for the TEC regimen. At last, the RADL and PATHO signatures were also been proved to be the independent prognostic factors for DFS and OS in breast cancer patients who underwent NAC.

Many studies have used radiomics to predict pCR in breast cancer patients, but several of these studies are ultrasound-based. Compared with MRI, ultrasound is a convenient, fast and inexpensive examination that is more suitable for breast cancer screening. Some studies have used ultrasound images from different treatment phases to predict pCR[17,18]. However, each patient responds differently to NAC, and some post-NAC images are difficult to accurately identify and delineate. Yu Liu et al. proposed a new strategy to improve the performance of this multimask network[19]. However, this study only explored its performance in HER2-positive breast cancer patients. Jionghui Gu et al. developed deep learning radiomics to predict pCR and demonstrated good predictive ability[20]. However, this model relies on ultrasound images from two different NAC cycles and lacks external validation, which weakens its clinical application. In fact, ultrasound images can only display macroscopic information about tumours. However, the emergence of digital PATHOlogy provides us with a new way to expediently understand microstructural information[7,21]. Whole-slide imaging (WSI) based on machine learning or deep learning models has shown good ability for survival prediction, microsatellite instability prediction and chemotherapy response prediction[[22], [23], [24]]. However, few studies have explored pathomics in breast cancer. Huang Y et al. developed radio-pathomics to distinguish luminal subtypes[25], but information on breast cancer patients who underwent NAC is lacking. Jieqiu Zhang et al. used contrast-enhanced CT images and whole-slide images to predict the pCR of patients with malignant breast tumours[11], but an external validation cohort was lacking. This study is the first to integrate ultrasound images and the WSI to predict the pCR of patients with malignant breast cancer who underwent NAC. These results demonstrated that the multiomics (DLG3, RADL and PATHO) signatures are independent predictors and that the nomogram based on the multiomics signatures could accurately predict pCR (AUC: training: 0.900; test: 0.814; external validation: 0.792). In addition, the multiomics model displayed a better predictive performance than the clinical TNM staging system and single RADL signature (P < 0.05). All the above results showed that this multiomics model has great potential for the prediction of pCR in breast cancer patients who underwent NAC. Compared to the single RADL signature, multiomics, including the PATHO signature, is more comprehensive in predicting pCR.

Our results demonstrated that DLG3 was an independent predictor of pCR. Therefore, we further explored its relationship with chemosensitivity. Several studies have investigated the biological function of DLG3 in breast cancer cells[26,27]. However, these studies focused on TNBC and luminal cells, and the opposite results were obtained for the expression of DLG3. In fact, bioinformatics analysis based on the TCGA database revealed that the expression of DLG3 is greater in HER2-positive patients than in patients with other subtypes. Therefore, we explored the function of DLG3 in HER2-positive cells. The results indicated that DLG3 is highly expressed in breast cancer cells and that the proliferation and migration abilities of DLG3-knockdown cells were lower than those of negative control cells. In addition, we further explored the relationship between DLG3 expression and chemotherapeutic drug sensitivity. The results showed that the IC₅₀s of docetaxel and epirubicin were lower in the DLG3-knockdown cells than in the negative control cells. Among breast cancer patients who underwent NAC with TEC regimens, those who achieved pCR had lower DLG3 protein expression. All these results indicated that DLG3 could be helpful in selecting NAC drugs. However, the expression of DLG3 relies on IHC or gene detection. Therefore, we explored the ability of the RADL and PATHO signatures to predict DLG3 status. The results demonstrated that patients in the low RADL_Signature_DLG3 and PATHO_Signature_DLG3 groups had lower DLG3 IHC scores (P < 0.05).

Based on the above results, patients with low RADL and PATHO signatures are more likely to receive NAC via the TEC regimen.

We also further explored the prognostic ability of multiomics signatures in breast cancer patients who underwent NAC. A previous study including 155 breast cancer patients revealed that patients with high DLG3 expression had worse OS[14], but its prognostic value in malignant breast cancer patients undergoing NAC is unclear. This study, for the first time, explored the prognostic value of DLG3 in breast cancer patients undergoing NAC. The results also showed that high DLG3 expression was significantly correlated with worse DFS (78.9 vs. 68.4 months, P = 0.035) and OS (91.9 vs. 82.9 months, P = 0.043). While previous studies[[28], [29], [30]] have shown the potential of ultrasound-based radiomic nomograms for predicting breast cancer survival, their clinical value is limited because these nomograms include additional clinicopathological features, and the performance of a single ultrasound-based radiomic model is poor. In addition, these studies lacked radiology and pathological features based on deep learning or machine learning models. The authors also lacked proof about OS and breast cancer patients who underwent NAC. Other studies have analysed the relationship between MRI-based radiomics and the prognosis of breast cancer patients[16,31,32]. Similarly, these studies focused only on recurrence. The predictive ability, as well as the predictive survival time, is limited. This study developed RADL and PATHO signatures via deep learning and machine learning models. The results demonstrated that these signatures were independent prognostic factors for DFS and OS. Patients in the high RADL and PATHO signature groups had shorter DFS (85.2 vs. 54.2 months, P < 0.0001; 81.2 vs. 65.5 months, P = 0.0015; 78.9 vs. 68.4 months, P = 0.035) and OS (98.8 vs. 67.6 months, P < 0.0001; 97.5 vs. 74.9 months, P < 0.0001; and 91.9 vs. 82.9 months, P = 0.043) times.

These results indicate that multiomics signatures have great potential to assist in survival prediction.

In contrast to previous nomograms that combined clinicopathological features, this study used only ultrasound images and the WSI before the first NAC cycle. Fewer signatures showed better predictive performance, which made this nomogram easier to use and apply. Although this is the first study to explore prognosis and pCR in breast cancer patients undergoing NAC based on multiomics signatures (RADL, PATHO and DLG3), several limitations still need to be considered. First, the expression of DLG3 was obtained from postoperative tissues. Although there are too few biopsy tissues from most patients for IHC, the availability of pretreatment DLG3 information will help us further understand its prognostic function. Second, due to the short follow-up time of the external validation cohort patients, we did not conduct external validation of the prognostic (DFS and OS) value, and a longer follow-up time will help us further validate the performance of the model.

Conclusion

The multiomics model, integrating with DLG3, RADL and PATHO signature, can effectively predict the pCR of breast malignant tumor patients undergoing NAC. That RADL and PATHO signatures are associated with DLG3 status, and could help doctors or patients choose proper neoadjuvant chemotherapy regimen (TEC regimen). At the meanwhile, the prognostic value of multiomics signatures are worth being further explored.

Ethics statement

This multicenter study was conducted in accordance with the Declaration of Helsinki and subsequently amended versions, and was approved by the ethics committee of the respective participating hospitals. Each patient signed an informed consent (including data secondary utilization) before underwent any treatment.

Availability of data and materials

The datasets used and/or analyzed during the current study are available from the corresponding author on reasonable request. The R codes were shared in: https://github.com/greenm00d/123.git.

CRediT authorship contribution statement

Cong Jiang: Writing – original draft. XueFang Zhang: Resources. Tong Qu: Resources. Xinxin Yang: Resources. Yuting Xiu: Resources. Xiao Yu: Resources. Shiyuan Zhang: Resources. Kun Qiao: Methodology. Hongxue Meng: Resources. Xuelian Li: Resources. Yuanxi Huang: Supervision.

Declaration of competing interest

There is no conflict of interest among all authors.