On-treatment biomarkers can improve prediction of response to neoadjuvant chemotherapy in breast cancer

The NEO study consists of 50 breast cancer patients with sequentially sampled biopsies at four time points, pre-treatment (PT, 34 samples), 2 weeks on treatment (T2, 12 samples), mid-chemo (TM, 23 samples), and at surgical resection (TS, 24 samples) with three clinically defined response statuses: complete responders (pCR by resection), good responders (tumour volume reduction, but lack of pCR), and non-responders (progressive disease or small tumour volume changes on treatment). Patients were of mixed histological grade and HER2 status; ages ranged from 29 to 76. Patients were primarily treated with 3 cycles of FEC and docetaxel with Herceptin where appropriate. Three patients received paclitaxel, one patient received additional carboplatin, one patient received Epi-cyclophosphamide and paclitaxel, and one patient received docetaxel and cyclophosphamide. Eligible patients were women with histologically confirmed invasive breast tumours and with no evidence of distant metastatic disease, no prior history of malignancy, and fit enough to receive chemotherapy in the opinion of the responsible clinician irrespective of age. All cases were discussed at the breast MDM in Edinburgh Breast Unit at the Western General Hospital, and consensus from this meeting was to be treated with neoadjuvant chemotherapy.

Core needle (16-gauge) biopsies were taken from the primary breast tumours before treatment (PT) and between 10 and 14 days after the first dose (T2) of chemotherapy. A third sample was taken at the mid-chemotherapy point day 20–21 (TM), and finally, a core biopsy was taken from the excision specimen (TS) after it has been removed prior to submission to pathology. Fixed and frozen samples of normal and tumour tissue were collected from all specimens.

RNA extraction was performed via Ribo0-RNAseq, and whole transcriptome sequencing was performed with Life Sciences Ion AmpliSeq™ Transcriptome Human Gene Expression Kit. This generated greater than 8 M reads per sample with an average of more than 90% valid reads for 12,365 targeted genes. Most analyses were performed in R (http://​www.​r-project.​org) using packages available through CRAN (http://​cran.​r-project.​org/​) and Bioconductor (http://​www.​bioconductor.​org/​). Outside of the R environment, the stand-alone application Multiple Experiment Viewer (http://​mev.​tm4.​org/​) was utilised for pairwise ranked product feature selection, and DAVID (https://​david.​ncifcrf.​gov/​) was used for pathway identification. Additionally, the python package scikit-learn [14] was used for unsupervised clustering analysis. Ninety-seven samples were analysed over 13 AmpliSeq chips, but no systematic batch effects were evident and no batch correction was performed within the training data. Gene expression data for the NEO study has been made publicly available at the NCBI GEO data repository under accession GSE122630.

The I-SPY 1 Trial is composed of patients with invasive breast cancer > 3 cm, or at least one tumour-positive axillary lymph node [11]. Patients were treated with an anthracycline-based chemotherapy followed by taxanes [11]. Samples were normalised and corrected for background red/green signal; Bioconductor R packages marray and limma [15] were used to this end. From the original 221 patients, only 36 had matching pre- and on-treatment samples, and 39 had matching biopsy and excision samples; pathological complete response was used for response criteria. Pairwise gene expression was handled with SAM and follow-up analysis with Ingenuity Pathway Analysis from QIAGEN Bioinformatics. I-SPY 1 Trial data is hosted at NCBI GEO under accession GSE32603 [11].

Principal component analysis (PCA) was performed on unsupervised gene lists to reduce dimensionality and visualise differences in response at all times and to identify present differences between patient treatment statuses. Local Fisher discriminant analysis (LFDA) [16] was used at each time point to determine if the response groups could be distinguished with treatment time with a semi-supervised clustering approach, concurrently with class advised K-means clustering. LFDA is a form of supervised dimensionality reduction that maximises between-class scattering and minimises within class scatter, and is a refined version of normal Fisher discriminant analysis [16]; this exploratory analysis was used in order to visualise comparative differences in treatment time, not as a means of feature selection. Pair-wise significance analysis of microarrays [17] using the siggenes package in R was used to consider the consistency of differentially expressed genes due to treatment in the sequential patient-matched samples. Rank Product analysis was used to identify differentially expressed genes between response classes at each time point. Successive levels of standard p value (0.05, 0.01, 0.001), without correction for multiple testing, were used in order to determine the number of differentially expressed genes, and at lower p values which the time points had the most strongly differentiating genes. Significance analysis of microarrays was also performed using varying false discovery rates (1%, 5%, 10%) to try to identify common differentially expressed genes between responders and non-responders across both datasets at each time point. Gene score enrichment analysis was used to validate the time point selection by looking for the highest number of enriched pathways. The gene list from the most differential time point (TM) using the NEO dataset was extracted and used in a random forest model (10,000 trees, m-try as the square root of the feature number) using pCR status as the class label (clinician-identified pCR and non-pCR). The most deterministic genes for class prediction were fed into a classification and regression tree in order to produce a maximally reduced and repeatable model; this methodology is further described by Turnbull et al. [7]. The CART decision tree was applied to the NEO dataset for training and tested in the independent I-SPY 1 dataset using the same cut-points determined by mean-centring the datasets. This protocol was repeated using the gene list from the pre-treatment only samples, using the same p values and tree configurations for selection. Survival analysis was performed at different time points using the log-rank test. Intrinsic subtypes, Mammaprint, and risk or relapse scores were estimated from the gene expression data using the GeneFu R package [18].

Unsupervised principal component analysis was first used to assess whether sequential patient-matched samples from patients receiving chemotherapy (Fig. 1b) would cluster by time point or response status. There was no significant grouping of patients according to sampling time: pre, early, or later after chemotherapy in either the NEO or I-SPY 1 studies (Fig. 1b). There were no significant differences between the two cohorts in terms of age, grade hormone receptor, and HER2 status, and the subset of patients with mid-chemo samples was not significantly different from the whole NEO cohort (Table 1). Patient-matched samples enable the pairwise analysis to look for consistent changes in the gene expression during treatment. Pairwise significance analysis of microarray analysis using a 10% false discovery rate (FDR) identified a relatively small proportion of overlapping upregulated (5%) and downregulated (4%) genes between the two studies. However, genes that were increased or decreased in response to treatment in one study were also clearly and consistently increased or decreased in the other study (Additional file 1: Figure S1A), further suggesting it would be difficult to discriminate responders from non-responders. Indeed, there was no clustering by response status before or during treatment (Additional file 1: Figure S1B). These results likely reflect the considerable inter-patient differences being substantially larger and more significant than the subtler commonalities in gene expression of a particular time point or response class of each tumour. More encouragingly, semi-supervised LFDA of each time point revealed significant separation on-treatment that was not apparent in pre-treatment samples; this indicated that there are meaningful differences between the classes, as early as 2 weeks on-treatment (Fig. 2a). Complete responders and non-responsive patients were more clearly separated than partially responding patients. These results suggest that there is a potentially greater predictive value looking at on-treatment than pre-treatment biomarkers.

In an attempt to quantify the molecular differences between the response groups at each time point, rank product analysis was performed at different standard p values (0.05, 0.01, and 0.001). This approach was hampered by different numbers of samples at each time point (with T2 having very few samples); however, the number of genes differentially expressed at all p values tended to be greater during rather than before treatment (Fig. 2b). Similar results were also seen using 1%, 5%, and 10% FDR (Fig. 2b). The biggest differences between the response classes were at TM (mid-chemo), which agrees with the LDFA results, which showed the least amount of overlap of the response classes at TM. Gene set enrichment analysis across the response classes at each time point also demonstrated more enriched pathways after 2 weeks of treatment (29), mid-chemo (30), and resection (29), compared to pre-treatment (18) (Additional file 2: Figure S2A). Next, we sought to examine common differentially expressed genes between responders and non-responders across the two datasets. Far more genes were commonly significantly differentially expressed (FDR = 10% between responders and non-responders on-treatment in the NEO and I-SPY 1 datasets compared with pre-treatment. In accordance with the LFDA results, more significantly differentially expressed genes (1814) were observed between on-treatment samples, with 6% (197), but only one was common between NEO and I-SPY pre-treatment (Additional file 2: Figure S2B and Additional file 4). Examination of the 468 most significantly differentially expressed genes (p < 0.001) between responders and non-responders in the NEO dataset at mid-chemo did not clearly distinguish between response groups or time points illustrated by the heatmap in Additional file 3: Figure S3, further demonstrating that identifying biomarkers of response to chemotherapy is very difficult.

We were also keen to evaluate whether the intrinsic subtype assigned to tumours would alter upon treatment. Looking at the NEO and I-SPY datasets, together we found that basal tumours were relatively stable with only 2/19 (11%) tumours changing. More tumours were classified as Luminal A or normal-like on-treatment, which likely reflects a reduction in the expression of proliferation genes during chemotherapy (Fig. 2c).

The mid-chemo gene list from the NEO dataset (1102 genes, unadjusted p value = 0.01) was fed to a random forest model for further feature selection and classification and regression tree (CART) model, which reported AAGAB as the most predictive gene for response prediction in the NEO training dataset with 100% accuracy for pCR prediction on the mid-chemo samples (Fig. 3a). Validation was conducted completely independently on publicly available sequentially sampled chemotherapy data from the I-SPY 1 Trial [10] and reported 76% accuracy using AAGAB at the same expression level on the scaled and centred expression data at the on-treatment time point prior to resection (T2). For comparison, the pre-treatment only sample gene lists were put through the same protocol in order to consider whether highly predictive models could be generated before chemotherapy. IGF1R was the most predictive pre-treatment marker with an accuracy of 74% and 63% in the NEO and I-SPY datasets, respectively (Table 2). AAGAB was the sixth most accurate predictor (65%, 57%); receiver operator curves show the relative specificity and sensitivity of this marker pre- and on-treatment (Fig. 3b). Gene expression levels of AAGAB were lower in responders across all time points in the NEO cohort but were most significantly different at mid-chemo. In the I-SPY dataset, AAGAB was significantly lower before treatment and at excision (Fig. 3c). We wondered whether AAGAB was lower in responders due to a reduction in proliferation, but Pearson correlation analysis with common proliferation-associated genes (TOP2A, BUB1, MKI67, MCM2, FOXM1, and PCNA) demonstrated no significant correlation to any of these genes (Fig. 3d), suggesting that AAGAB is independent of proliferation. Survival analysis demonstrated that response status predicted by AAGAB level, at mid chemo in the NEO study and at 2 weeks in the I-SPY 1, was significantly associated with the outcome (NEO p = 0.048, I-SPY 1 p = 0.0036) (Fig. 3e). Interestingly, the level of AAGAB before treatment was not associated with the outcome in either cohort (p = 0.71 and p = 0.2, Fig. 3e). None of the other top 10 pre- or on-treatment markers was significantly associated with the outcome in both datasets (Table 2); only one gene (ARF5) was associated with the outcome in the NEO dataset (p = 0.004). Taken together, the single gene on-treatment biomarker AAGAB appears to outperform novel pre-treatment markers and established prognostic tests in predicting pCR and long-term outcome to chemotherapy.

We were also keen to assess whether estimations of established prognostic signatures might be different upon treatment and if on-treatment might be more accurate. All and almost all responding patients were predicted to have poor outcomes with the estimated Mammaprint [19], PAM50 [20], or rorS [21] signatures in pre-treatment samples of the NEO cohort, whereas around half of the responding patients were predicted as good outcome using on-treatment data (Fig. 4a). Overall accuracy improved by 2–8% using on- rather than pre-treatment data; however, improvement in the predictive power of these tests was not uniform between response classes. Good outcome predictions for responders to neoadjuvant chemotherapy saw an aggregate increase in predictive power from 11 to 44.4%, whilst poor outcome predictions for non-responders saw a moderate decrease in accuracy, 75 to 63%. None of the gene expression signatures either pre- or on-treatment or established prognostic markers (NPI, Grade, Her2 status) was significantly associated with the outcome in contrast to the remarkable performance of on-treatment measurement of AAGAB (Fig. 4b).