Construction of a risk stratification model integrating ctDNA to predict response and survival in neoadjuvant-treated breast cancer

Miller-Payne (MP) grading and residual cancer burden (RCB) index were used to validate the response to NAC. The RCB evaluation system (www.​mdanderson.​org/​breastCancer_​RCB) was used to calculate the RCB index [18]. pCR (stage yp-T0/is, ypN0) was defined as RCB = 0, and residual disease (no-pCR) was placed into three predefined subgroups (RCB-I, RCB-II, and RCB-III) on the basis of predefined cut-off points of 1.36 and 3.28 index scores. RCB and pCR status were evaluated by two independent pathologists. If their conclusions were inconsistent, a third pathologist reassessed the situation.

A total of 243 tissue puncture samples from formalin fixation and paraffin embedding (FFPE) were obtained before the initiation of chemotherapy by core needle biopsy from 246 patients. Among the 246 patients, 56 patients underwent ctDNA testing of blood samples. The blood samples were collected from each patient dynamically over the course of NAC at four time points as follows: before NAC (T₀), after the 1^st NAC and before the 2^nd NAC cycle (T₁), during intermediate evaluation (T₂), and after the end of NAC but before surgery (T₃) (Fig. 1A). The baseline plasma samples of all 56 patients were collected, while six patients failed to complete the sample collection at all four time points, and the remaining 50 patients completed the entire plasma sample collection process (Fig. 1B).

A GeneRead DNA FFPE kit (Qiagen, USA) was used to extract genomic DNA (gDNA) from FFPE and fresh tissue samples, and a DNA blood Midi/Mini kit (Qiagen, USA) was used to extract genomic DNA from white blood cell samples according to the manufacturer's instructions. The MagMAX cell-free DNA Isolation Kit (Thermo Company) was used to separate plasma cell-free DNA (cfDNA). The quality of purified DNA was quantified by gel electrophoresis and using a Qubit® 4.0 fluorometer (Life Technologies, USA).

First, the purified gDNA was fragmented to approximately 200 bp by enzymatic hydrolysis (5X WGS Fragmentation Mix, Qiagen, USA). After end repair and A-tail connection, the two ends were connected with T-adaptors and then PCR amplified to form a prelibrary. The purified prelibrary was hybridized with a customized biotin probe pool (457 gene panel, Berry Oncology, Beijing, China) to capture the target clip. According to the manufacturer’s instructions, the 96 rxn xGen Exome Research Group v1.0 (Integrated DNA Technologies, USA) was used to prepare the final sequencing library.

For targeted sequencing of cfDNA, a prelibrary was prepared according to the method described above. Internally designed panels were used to capture cfDNA fragments and generate sequencing libraries. The sequencing library was applied to the NovaSeq 6000 platform (Illumina, San Diego, USA) in 150PE mode.

The generated sequence was trimmed, low-quality filtered, and subjected to variant calls. Variations were filtered into nonsynonymous SNP, indel, and splicing variations. For gDNA, somatic mutations were left with allele frequencies (VAF) ≥ 3%, and cancer hotspots were retained with VAF ≥ 1%. For cfDNA, the frequency of variant alleles (cut-off value ≥ 0.5%) was used to identify somatic mutations, the frequency of variant alleles (cut-off value ≤ 0.1%) was used to screen for cancer hot spots, and at least 20 high-quality reads were screened.

FASTP [19] was used to trim adapters and delete low-quality sequences to obtain clean reads. The clean reads were compared with the Ensemble GRCh37/hg19 reference genome executed by BWA. The PCR repetitive sequence was processed to consensus sequence by GenCore [20], then SAMtools [21] was used to detect somatic single nucleotide variants (SNVs), insertions and deletions (InDels), and then false variants were filtered by a series of methods such as vaf cutoff, paired control sample, negative background database. HGVS variant description was annotated with ANNOVAR software [22]. After annotation, we used PopFreqMax > 0.05 to eliminate SNVs and InDels, then retained nonsynonymous SNVs and InDels with VAF > 0.5% or VAF > 0.1% among the cancer hot spots in the patient database for further analysis. Somatic copy number variants (CNVs) were called by CNVkit [23] through several steps, such as depth normalization and GC correction. If the copy number > 3, we marked the target as a target gain, and if the copy number < 1, we marked the target as a target loss. The tumor mutation burden (TMB) was defined as the number of all nonsynonymous mutations and indel variants per megabase of coding regions.

Baseline characteristics analysis was performed on all 246 patients, including the distribution of patients with baseline genomic or clinical characteristics in the pCR/non-pCR group and the correlation between baseline characteristics and pCR status or patient prognosis. The association analysis between mutation detection and prognosis was performed on 56 patients with continuous ctDNA test data (that is, the completion of the entire study).

Fisher’s exact test (two-sided test) was used to analyze the impact of baseline pathological characteristics (such as breast cancer type, Ki67, age, sex, and disease stage) and mutation characteristics on non-pCR. Those features related to the pCR/non-pCR state were selected to construct a predictive model for pCR/non-pCR prediction. Usually, we chose to perform the next step of data analysis with Fisher's exact test (P value < 0.05) for the feature. A total of 192 patients who had tissue samples from Shandong Cancer Hospital were randomized into a training cohort (n = 128) and a testing cohort (n = 64) using the “caret” R package. The training cohort was used to find a meaningful signature, and the testing cohort and patients of Liaocheng People’s Hospital (n = 51) were used to validate its efficiency.

In the training cohort, significant SNVs and CNVs with a mutation frequency greater than 10% including five SNVs and 10 CNVs mutations were subjected to multivariate stepwise logistic regression analysis. The nine mutated genes including five SNV mutations and four CNV mutations were identified by step-by-step logistic regression analysis. Then, a multi-response model was built including only clinical characteristics, only SNV characteristics, only CNV characteristics, both SNV and CNV characteristics, both SNV and clinical characteristics, both CNV and clinical characteristics, and SNV, CNV, and clinical characteristics for a total of 6 response models. All models were based on the random forest method and developed in a tenfold cross-validation (CV) schema. Performance was assessed in terms of accuracy (ACC), sensitivity, and specificity by receiver operating characteristic (ROC) curve analysis. A nomogram was also applied to estimate the performance of the signature. ROC analysis was performed using the “caret” and “ROCR” R packages and a nomogram was generated using the “rms” R package.

We defined disease-free survival (DFS) as the time from the breast surgery until disease progression (including local or distant recurrences) or death. We constructed a DFS prediction model based on the signature of the best model among the multiresponse models for 192 patients from Shangdong Cancer Hospital in the training set and test set. Then, a risk score was calculated for each patient, and patients were divided into high- and low-risk groups based on the median risk score. The ROC analysis was performed using the “timeROC” R package. Kaplan–Meier curves were generated for survival analysis, and the log-rank test was used for comparisons.

The threshold for ctDNA positivity was determined by the mutation numbers of nine genes included in the conducted tumor prediction model. If there were more than two mutations among these nine genes, then this sample was recognized as ctDNA positive [24]. ctDNA fractions were calculated as 2/(1/Max(VAF) + 1) [25]. One-way analysis of variance (ANOVA) was applied for comparison of ctDNA fractions in different groups. Use the random forest model to obtain the probability of treatment response or non-response for each sample, and combine the ctDNA status at different time points with the random forest model to construct a combined tissue and blood efficacy prediction model.

Conduct multivariable Cox regression analysis using the mutation levels of the nine genes at the tissue level, clinical features, and the ctDNA status at each time point to obtain the risk scores for each sample, and construct a DFS prediction model. Similarly, divide the samples into high and low-risk groups based on the median risk score.

A flowchart for the algorithm is shown in sFig. 4. All statistical analyses were performed using R software, version 3.6.3 (www.​r-project.​org). For all statistical analyses, p < 0.05 was considered statistically significant.

A total of 192 patients who had tissue samples in Shandong Cancer Hospital and 51 patients in Liaocheng People’s Hospital were used to predict the residual cancer burden status of patients receiving NAC. The specific clinical information is shown in Table 1. The median age was 50 years old, and 59.3% were younger than 50 years old. Most patients were in stage III (51.9%). The subtypes of Her2 positivity, luminal A, luminal B (Her2-), luminal B (Her2 +), and TNBC accounted for 19.3% (47/243), 21.4% (52/243), 23.5% (57/243), 18.1% (44/243), and 17.7% (43/243), respectively. The pCR rate in different molecular subtypes was shown in Additional file 1: Table S1. Patients with Ki67 expression > 20% accounted for the main population (67.8%). Postoperative pathological examination showed that 61 patients (25.1%) had achieved pCR by RCB index. For those patients who showed non-pCR, 11.9% (29/243), 35.8% (87/243) and 27.2% (66/243) reached RCB-I, RCB-II and RCB-III, respectively. 87.2% of patients received the anthracyclines plus taxanes-based chemotherapy, 5.8% received the anthracycline-containing regimen only, and 7% of patients received taxanes-containing regimen alone (Additional file 1: Table S2 and Table S3). An overview of the study design was shown in Fig. 1.

To discover somatic variations in tissue samples used for model construction, we extracted DNA from FFPE samples of punctured tissue samples and performed NGS-based 457 gene panel testing. We analyzed and summarized the somatic mutations of 192 samples with high-frequency mutation ≥ 10%. The 425 unique genes were identified, and the top five highly mutated genes were TP53, KMT2C, PIK3CA, EPHA1 and EPPK1 with mutation frequencies of 64% (123/192), 47% (90/192), 44% (84/192), 42% (80/192), and 36% (69/192) (Fig. 2A and Additional file 1: Table S4). The results of somatic CNV showed that 114 genes in tumor tissues were amplified with at least three times of normal tissues. There were 43 genes that were deleted, and 200 genes were both amplified and deleted (Additional file 1: Table S5).

We investigated the relationship with pCR and gene mutation characteristics and clinical phenotypes. All 192 patients were randomly divided into two groups in a 2:1 ratio into a training set (n = 128) and a test set (n = 64). In the training set, we first performed differential mutation gene analysis on pCR and no-pCR (RCB-I, RCB-II, and RCB-III) patients. Using Fisher’s exact test, we found five different SNV mutated genes (TP53, SETBP1, PIK3CA, NOTCH4, and MSH2) and ten CNV mutated genes (B4GALT3-gain, CDK12-gain, EGFR-gain, PRDM1-gain, GATA2-loss, GNA11-loss, NOTCH3-loss, FOXP1-gain, IL7R-gain, and NFKB1A-gain) with a mutation frequency greater than 10% (Additional file 1: Table S6). To further screen for the factors used to predict pCR status, stepwise logistic regression was used to screen for the differentially mutated genes, and nine differentially mutated genes that were significantly related to the pCR status (MSH2, NOTCH4, PIK3CA, SETBP1, TP53, EGFR, FOXP1, IL7R, NFKB1A) were screened (Fig. 2B, C). Given the potential importance of clinical factors in NAC, we screened the clinically relevant factors of luminal A, Her2⁺, and Ki67 related to NAC using Fisher’s exact test. Patients with luminal A were less sensitive to NAC than patients with other subtypes (P < 0.001) and no luminal A patients achieved pCR. In Her2⁺ patients, the proportion of patients who achieved pCR after NAC was significantly higher than that of non-Her2⁺ patients (43.2% vs 20.8%, P = 0.01). In addition, patients with higher Ki67 expression pretherapy had a higher proportion of attained pCR patients (31.39% vs 9.26%, P = 0.001) (Fig. 2D).

Above, the nine mutant genes (five SNV mutations and four CNV mutations) and three clinical factors were identified that were associated with response to NAC. This motivated the use of a machine learning framework to integrate these factors into a predictive model of pCR. We investigated a number of prediction models including the gene mutation information and clinical factors alone and the combination of mutation information and clinical factors. The results found that a combination of nine mutant genes and three clinical factor models had the higher sensitivity and specificity than other combinations (Additional file 2: Fig. S1) in the training test (AUC: 0.871, 95% CI: 0.797–0.927), in the verification set (AUC: 0.771, 95% CI: 0.649–0.883) and in the extra test (AUC: 0.726, 95% CI: 0.556–0.865) (Fig. 3A). We performed a multivariate logistic analysis of NAC response in the training cohort to generate a nomogram to predict the results of pCR according to RCB index after NAC. Among the nine mutant genes and three clinical factors, MSH2, FOXP1 and luminal A were the three most important factors (Fig. 3B). Given the role of MP scoring, which has universal application in the clinic, we also tested the applicability of the model to MP scores and found that our model also predicted pCR and non-pCR well in MP score classification (Fig. 3C), and the nomogram showed that MSH2 still had the greatest importance (Fig. 3D).

Given that patients who achieved a pCR had better OS than patients who did not, we tested the predictive ability of the model on the prognosis. In total, 23 patients were lost to follow-up in the 192 enrolled patients from Shandong Cancer Hospital. The median follow-up time after surgery was 898 days (97 to 1765 days), and 28 of the 169 patients (17.6%) progressed. The results showed that the model also has a good predictive effect on the prognosis of NAC patients, and the predictive effect of the long-term prognosis was better than the short-term predictive effect (AUC at 1 year = 0.749 vs. AUC at 3 years = 0.830) (Fig. 3E). According to the median risk score, patients were divided into high-risk groups and low-risk groups. The DFS of the low-risk group was significantly higher than that of the high-risk group (P < 0.0001) (Fig. 3F).

A personalized panel consisting of nine somatic mutation genes was selected to detect and analyze ctDNA from the tumor model factors for predicting NAC response. Fifty-six patients among all the 246 patients underwent ctDNA testing. two hundred sixteen blood samples were collected dynamically over the course of NAC (plasma samples were collected from 56 patients of T₀ and T₁, 54 patients of T₂, and 50 patients of T₃). A sample with at least two detectable somatic variations was considered positive for ctDNA (Additional file 1: Table S7). Before treatment (T₀), 46% of patients were ctDNA positive (Fig. 4A). Patients with TNBC had a higher expression of positive ctDNA (80%) compared with other subtypes while luminal A and luminal B patients mainly had negative ctDNA (Fig. 4B). In addition, patients with low Ki67 status expressed negative ctDNA (70%) (Fig. 4B).

The ctDNA positivity rate decreased with the passage of time during NAC. In the entire population, ctDNA positivity gradually declined during NAC, from 46% before treatment (T₀) to 14% before the 2^nd NAC cycle (T₁), and it was 13% during intermediate evaluation (T₂), and after NAC (T₃), it dropped to 10% (Fig. 4C). Similarly, the ctDNA fraction also decreased with the passage of NAC time (Fig. 4D).

To investigate whether the dynamic change of ctDNA could related to the NAC response, we constructed five patterns based on the clearance dynamics of ctDNA expression in 50 patients who had complete data at all four time points, with cleared at all time points (T₀, n = 26, 52%), cleared at T₁ (n = 13, 26%), cleared at T₂ (n = 4, 8%), cleared at T₃ (n = 2, 4%), and patients who remained ctDNA positive after NAC (T₃) (n = 5, 10%) (Fig. 5A). We identified 45 patients who had both survival data and ctDNA status at all four time points. The rate for positive detection of ctDNA decreased during the NAC, and the positive rate dropped from 44.4% (T₀, 20/45) to 11.1% (T₃, 5/45) (Fig. 5B). All the patients with pCR had undetectable ctDNA at T₂ and T₃ with no disease progression (Additional file 2: Fig. S2A, Fig. 5B). In contrast, patients with ctDNA-positive at T₂ and T₃ not achieved pCR (Fig. 5B). In addition, the patients with disease progression were mainly RCB-III (75%, 6/8) and RCB-II (25%, 2/8) (Fig. 5B).

To detect ctDNA dynamically for patients whose ctDNA is detected before the NAC (T₀), we performed monitoring of the ctDNA expression at T₁, T₂, and T₃. The positive rate of ctDNA gradually decreased as the NAC treatment with 35% (7/20) at T₁ and 20% at T₂ and T₃ (4/20) (Fig. 5C). Among patients who did not clear ctDNA at T₁, as many as 85.8% had residual disease at the time of surgery (6/7 non-pCR), while 69% of patients who cleared ctDNA at T₁ had residual disease (9/13 non-pCR). At T₂ and T₃, in patients who did not have clear ctDNA, 100% had residual disease during surgery (4/4 non-pCR), while in patients who cleared ctDNA, 69% (11/16 non-pCR) had residual disease. The positive predictive value of ctDNA increased with treatment time (Additional file 2: Fig. S2B).

To assess whether ctDNA status was related to metastasis and recurrence, we analyzed the association with ctDNA dynamic pattern and DFS. Patients who did not clear ctDNA at T₃ (n = 5) had a significantly higher risk of metastasis and recurrence than patients who cleared ctDNA at T₀, T₁, T₂, and T₃ (Fig. 5D, HR 4.61; 95% CI, 1.05–20.19, P = 0.027). Compared with patients who were ctDNA negative at T₀ (n = 22), patients who were ctDNA negative at T₁, T₂, or T₃ (n = 18) had a similar risk of metastasis and recurrence (Additional file 2: Fig. S3). Patients who had cleared ctDNA at T₁, T₂, T₃ had longer DFS than patients who had not cleared ctDNA at T₃ (Additional file 2: Fig. S3).

The clearance of ctDNA after NAC (T₃) is related to the improvement of the survival rate. After NAC, patients were stratified according to pCR and ctDNA status (n = 45). Seven patients with pCR (100%) (all ctDNA negative) showed good DFS. Among patients who did not achieve pCR (n = 38), ctDNA positivity (n = 5) was related to worse DFS. The probability of recurrence differed between patients who failed to achieve pCR, being greater in RCB/ctDNA + groups compared with the RCB/ctDNA- group (Fig. 5E, HR 3.92; 95% CI, 0.9–17.02, P = 0.061).

To further improve the accuracy in predicting NAC response to further predict the pCR status of breast cancer patients after NAC, we calculated the probability of pCR and non-pCR of the sample through the established tumor prediction model, and then we combined it with the negative and positive status of ctDNA at different time points and used random forest to construct a chemotherapy model. Firstly, the status of ctDNA before NAC combined with the tumor prediction model has a better prediction effect of pCR with the AUC of 0.961 (Fig. 6A). We constructed a chemotherapy prediction model by combining tumor prediction models with ctDNA status at different time points (T₀, T₁, T₂, T₀ and T₁, T₀ and T₁ and T₂) in order to assess the impact of dynamic changes in ctDNA on prediction. It was found that the prediction model was compatible with ctDNA, and combined with ctDNA status at different time points had similar results for AUC (Fig. 6B, T₀ = 0.961, T₁ = 0.951, T₂ = 0.92, T₀/T₁ = 0.961, T₀/T₁/T₂ = 0.961). To clarify the impact of the chemotherapy prediction model on patients' prognosis, we analyzed the predictive effect of the DFS. It showed a better predictive effect (AUC at 1 year = 1.000, AUC at 2 years = 0.941) on DFS (Fig. 6C). The patients were divided into high-risk groups and low-risk groups according to the median risk score. The DFS of the low-risk group was significantly higher than that of the high-risk group (P = 0.0031) (Fig. 6D).