Patients were recruited and treated at the Royal Marsden Hospital (RMH), London, UK. Eligible patients were those undergoing neoadjuvant AC chemotherapy treatment at doses of 60 and 600 mg/m², respectively, three times a week, for a clinically measurable breast tumour. The study was approved by the RMH Clinical Research and Ethics Committees (study number. 1,947), and written consent was obtained in all cases. Patients had been allocated neoadjuvant treatment for one of several standard indications, including locally advanced or inflammatory breast cancer, high tumour-to-breast size ratio, and tumours located close to the nipple.

Diagnosis was confirmed histologically by core-cut biopsy. All patients on hormone replacement therapy at diagnosis were advised to discontinue this treatment. Patients who demonstrated at least a partial clinical response received six cycles of treatment prior to local treatment. Patients in whom there was no or only marginal response after three or four cycles proceeded directly to local treatment or were commenced on alternative systemic treatment (docetaxel).

Clinical size of tumour (largest diameter and a diameter perpendicular to this) was recorded prior to commencement and at completion of treatment. Clinical response was categorised as follows: no palpable abnormality after treatment, complete clinical response (cCR); more than 50% reduction in the product of the bidimensional measurements, partial response (PR); less than 50% reduction in the product of bidimensional measurements was recorded as no change (NC); and residual ill-defined thickening after a good response, minimal residual disease (MRD). Those cases in which there was no residual invasive carcinoma at surgery were classified as a complete pathological response (pCR). Good responders were defined as pCR, cCR, or MRD; poor responders were defined as PR or NC. These categories were chosen on the basis of our previous study, which showed that patients with 'good' response had superior overall survival relative to those with 'poor' response 18. A proportion of patients undergoing a complete clinical and radiological (on ultrasound) response received radiation only as local treatment. Therefore, some of the cCRs may represent undocumented pCRs.

Research 14-gauge core biopsies were collected prior to commencing treatment and snap-frozen in liquid nitrogen. When consented to, a repeat sample was taken at 21 days after the first cycle of chemotherapy. All samples were thereafter coded using a study number as an identifier. Frozen cores were embedded in OCT (optimum cutting temperature embedding compound) and sectioned at -20°C in a cryostat. Sections (5 μm in thickness) were taken for haematoxylin and eosin staining to assess histological character superficially from the core as soon as 'full-face' was reached. The percentage of cells comprising invasive malignant disease and non-malignant components (that is, in situ disease, inflammatory infiltrate, non-malignant ductal/lobular structures, and fibroblastic involvement) were recorded by consensus between two breast pathologists. For patients in whom multiple biopsies were available, the biopsy with the highest invasive content was used for microarray analysis. Biopsies with less than 20% invasive cancer content were excluded from the study.

Cores were extracted from OCT as described by Ellis and colleages 19 and pulverised using a pestle and mortar on a bed of dry ice and subsequently in 1 ml of Trizol reagent (Invitrogen, Carlsbad, CA, USA) with a 'Polytron' homogenizer. Standard 'Trizol' RNA extraction was carried out without use of a carrier according to the manufacturers' instructions. Samples not giving distinct 18S and 28S peaks on an Agilent Bioanalyzer (Agilent Technologies, Palo Alto, CA, USA) trace were excluded from the study. Multiple cores from the same patient were handled separately for RNA extraction. A single round of T7 linear RNA amplification was carried out using the RiboAmp kit (Arcturus, Sunnyvale, CA, USA) with a starting amount of 1 μg when available or 50% of the available RNA.

Reference RNA was generated from a pool of RNAs extracted from 20 independent breast cancer surgical samples.

Microarray analyses used in-house (Breakthrough Breast Cancer Research Centre, London, UK) human arrays spotted with DNA derived from 4,600 IMAGE cDNA clones in duplicate. This set of genes represents a subset of the 5,808 Cancer Research UK (London, UK) gene set that was designed to include a high proportion of genes documented as being involved in carcinogenesis or tumour biology. To improve 'coverage' of genes involved in breast cancer, a list of discriminatory genes cited in important microarray studies on clinical breast cancer samples at time of production was compiled 7202122, and the array was supplemented with these. One half to two micrograms of sample RNA and a matched amount of reference RNA was labeled using the Powerscript Labeling kit (Clontech, Mountain View, CA, USA) in combination with Amersham Cy dyes (GE Healthcare, Little Chalfont, Buckinghamshire, UK). A single dye swap experiment was performed for each clinical sample. Slides were scanned using a GenePix 4000B (Axon Instruments Inc., Union City, CA, USA) scanner and GenePix version 4.0 software.

Most of the data analyses were carried out with the S-plus statistical software package (Insightful, Seattle, USA) and purpose-written scripts (T Dexter, Breakthrough, UK). Raw expression values were transformed to Log₂ ratios (sample/reference). The loess function 23 was used to remove biases due to the spot position and spot intensity. Flagged spots were treated as missing values. Log ratio values from duplicate spots and hybridisations were averaged. Genes with consistently low intensity and those that exhibited little variation across samples were removed from the analysis. After the above pre-processing, 1,286 genes remained for the prediction analysis. Samples were clustered both by complete linkage and flexible beta (beta = -0.5) agglomerative algorithms with (1 - correlation) as a distance measure 2425. The correlations were estimated using Spearman's rank method.

The nearest neighbour class prediction algorithm (euclidean distance) was used for all classifications because of its simplicity and good performance on microarray data as reported by Dudoit and Fridlyand 26. We elected to use seven nearest neighbours throughout to give more stable error estimates and greater robustness than would result with smaller neighbourhoods. The weighted Kolmogorov-Smirnov statistic was used to rank genes for discriminatory information 27. Each predictor was built starting with the highest ranked two genes from the training set, and then genes were added to the predictor in decreasing rank order until the error rates no longer decreased.

Owing to the limited number of samples, the misclassification error was estimated by leave-one-out cross-validation (LOOCV). In this approach, the class (that is, response) of each sample was predicted in turn, using the other samples as the training set. To avoid selection bias, the genes that were used as predictors were re-selected for each of these leave-one-out classifications 28. To estimate the probability of the misclassification error arising by chance, a permutation p value was determined as suggested by Radmacher and colleagues 28. In this procedure, the LOOCV estimate was determined 1,000 times for permuted class labels; the fraction of these that gave an equal or lower error estimate than that with true labels is taken as the p value. We refer to the latter as a 'label permutation'.

To assess the significance of lower misclassification error estimates for selected subsets of the samples, we needed to control for the smaller sample sizes. To this end, we defined a 'subset permutation' p value as the fraction of random subsets (1,000 matched for size and class proportions) of the full set of samples that gave rise to LOOCV error estimates equal to or lower than the selected subset. The correct class labels were used for subset permutation. The minimum error for each permutation across the top 2 to 10 ranked genes was used to calculate the p values for both of the above types of permutation to avoid bias. Although this is an arbitrary range, the minimum error rates in all the permutations had increased by a cutoff of 10. Significance analysis of microarrays (SAM) analyses were performed using standard software 29. Pre-treatment samples only were used for prediction analyses.

RNA of adequate amount and quality was available from 43 tumours before treatment. Of these, 23 (53%) demonstrated a 'good' and 20 (47%) a 'poor' clinical response. For 16 of these tumours, a paired 21-day sample was also available. These 'on treatment' biopsies are included in the analysis of reproducibility shown below to increase the number of paired samples but are excluded from the analysis of response prediction. The good responses comprised 16 (37%) that underwent a cCR, three (7%) that exhibited ill-defined thickening (MRD) at the end of treatment, and four (9%) that underwent a pCR (all cCR or MRD). The patient and tumour characteristics are shown in Table 1. The only feature differing between good and poor responders was tumour size, with pre-treatment size greater in poor compared with good responders (Mann-Whitney, p = 0.03). Pre-treatment size did not relate to expression profile (data not shown).

A total of 147 cores were sectioned in the course of this study, including 104 that were sectioned at three levels (levels approximately 50 μm apart). Of these 104 cores, only 16 cores showed more than 10% absolute variation in invasive tumour content across all three sections and only one core showed more than 20% variation. Only four cores were found to have tumour at some levels and none at others; all four contained not more than 15% invasive tumour at the lowest level and were therefore excluded from the study. This suggests that the histological composition did not vary widely over the width of the core. The histological result from the lowest section for a given biopsy was taken as that most representative of the remaining biopsy used for profiling and hence the level on which percentage invasive tumour was assessed. The distribution of percentage invasive content of core (in cases of multiple cores, that core used for prediction analysis) did not relate to response as assessed by Wilcoxon rank sum test (p = 0.3) or t test (p = 0.27).

Percentage invasive content in core biopsies included in the prediction study varied from 20% to 95% (median 50%). For most core biopsies, the majority of the non-malignant tissue consisted of connective tissue. Inflammatory infiltrate was also scored on the lowest section as nil, mild, moderate, or severe. Of all 147 core biopsies, only six (4%) were scored as having a 'severe' inflammatory infiltrate at one or more levels.

To assess consistency of the expression profile between repeated biopsies, the 43 pre-treatment and 16 paired post-treatment samples were clustered (Additional file 1). The post-treatment samples were included in this validation study only to enhance numbers and were not used in subsequent predictive analyses. All but one pair (b223A and B) of duplicate biopsies taken at the same time point relative to treatment (in total, five pre-treatment and five post-treatment pairs) clustered as nearest neighbours despite some variation in percentage tumour between the pairs (Additional file 3). Of 16 pre/post-treatment pairs, 14 clustered with samples from the same tumour (Additional file 1). To validate our class prediction methodology, a supervised analysis was undertaken to obtain a gene list for prediction of oestrogen receptor (ER) status in pre-treatment samples (ER-α was excluded from the analysis). ER status was correctly assigned in 40 of 42 cases on LOOCV. Discriminatory genes are listed in the supplementary information (Additional file 4).

Before performing prediction analysis, an exploratory analysis was undertaken to assess the impact of variation in the histological content of biopsies on the expression profile. A SAM regression analysis was performed to establish whether it was possible to identify genes that correlated with the proportion of tumour cells in the 43 pre-treatment core biopsies. One hundred forty-four genes were significant at a false discovery rate of not more than 5%. Table 2 shows the most correlated genes from this analysis. Positive values for 'score' (bold) indicate positive correlation with high percentage tumours, and negative values (underlined) indicate negative correlation with proportion of tumour cells in the samples. More genes correlated positively with the stromal content (n = 128) than with the tumour content (n = 16), possibly reflecting the greater molecular heterogeneity of tumour types across the samples than that of their associated stromas. This is also reflected in the higher false discovery rates for 'tumour-associated genes'.

Prediction analysis was initially undertaken on the full set of 43 biopsies. The optimum misclassification error estimate (LOOCV) for the whole sample set was 28% using a three-gene predictor (Figure 1). To explore the effect of biopsy tumour content (and its associated influence on the expression profile) on error of response prediction, we selected three overlapping subsets of tumour samples which were more homogeneous in terms of percentage tumour content: (a) ≥ 50% (50%–95%) invasive tumour (25 samples), (b) ≤ 50% (20%–50%) invasive tumour (24 samples), and (c) 35%–60% invasive tumour (24 samples).

The subset consisting of 35%–60% represented a group centered around the median sample in terms of percentage tumour content and of a size similar to or the same as the other two groups.

The minimum error of classification for each of the subsets (8%–13%) was lower than that for the superset of all samples (28%) (Figure 1), suggesting that homogeneity of tumour content rather than tumour content per se might be an important factor for response prediction. However, the comparison of the subset error estimates with that for all samples is not controlled for the different sample sizes involved. To address this, we determined a 'subset permutation' p value for each of the subsets (see Materials and methods). We also estimated the probability that the subset errors arose by chance (a 'label permutation' as discussed in Materials and methods). Error estimates and corresponding p values for the 'subset permutation' and the 'label permutation' support the hypothesis that homogeneity of biopsy tumour content improves response prediction with the nearest neighbour algorithm (Table 3). The identity of the genes used in the prediction for each of these subgroups is presented in Table 4. The mean differential expression between good and poor responders for these genes is given in the supplementary information (Additional file 5).

We found, in common with other studies 2022303132, that ER-positive and ER-negative tumours were very distinct in molecular terms (Additional file 4), which may confound response prediction. Therefore, response prediction was performed on ER-positive samples alone (16 good, 13 poor responders). Error rates, however, were high (31% with 8 genes, 31% with 25 genes). Within this subset, the effect of heterogeneity of tumour content on prediction appeared to be more pronounced (error rates: 31% (all samples, n = 29) versus 4.5% (≤ 60% tumour, n = 22)).

To explore further the effect of biopsy composition on prediction error, we attempted to create a single predictor for all samples by adding information about tumour content to the predictor as an extra dimension or 'histology gene'. The impact of variation in biopsy tumour content on prediction error was thought to be due partly to the fact that genes that discriminate between good and poor responders in high-tumour-content biopsies are often poor discriminators in low-tumour-content biopsies (and vice versa) as is evident from Table 4. To overcome this complication, 10 genes were selected that showed 100% support in the subset permutations (Table 4). By adding the proportion of tumour content to these 10 genes as an 11th dimension, we aimed to move the samples apart in 'prediction space' such that differences in tumour content also contributed to the distances between samples. Thus, samples with similar expression profiles become nearest neighbours of similar proportion of tumour content as well. In the nearest neighbour predictor, like is matched with like in terms of proportion of tumour content as well as expression profile.

It was necessary to rescale the histology gene's contribution to the distance between samples because the standard deviation of most genes lay between 0.5 and 1.5 (log₂ ratio scale). Therefore, the percentage tumour figures were standardised by subtracting the average and dividing by the standard deviation. The 'histology gene' or 11th dimension was then rescaled from 0 to 4 standard deviations, and the effect on cross-validation error determined (Figure 2). After an initial drop in error, as the scale approached 1.0, the error rose beyond the initial error as the influence of the histology gene became too dominant. To test whether the effect of the histology gene was specific to the particular set of 10 genes used above, we tested the effect of adding the histology gene to 1,000 predictors, each containing a random permutation of between 8 and 16 genes drawn from the top 20 genes ranked by a combined score (Additional file 6 a-c). In 91% of these, the addition of the histology gene resulted in a lower error; in 49%, the error drop was greater than 10%.