Background
Women with high-risk early breast cancer are increasingly being offered chemotherapy before definitive surgery because neoadjuvant chemotherapy can enable breast-conserving surgery [
1]. Complete eradication of tumour cells in the surgically removed tumour bed or pathological complete response (pCR) is associated with improved survival [
2,
3]. The likelihood of pCR is profoundly affected by the oestrogen receptor (ER) and human epidermal growth factor receptor 2 (HER2) status of the primary tumour [
2,
4]. In spite of these differences, improved prediction of the probability of pCR is needed because alternative regimens of chemotherapy or enrolment in clinical trials might be offered to patients deemed unlikely to experience pCR at baseline.
Novel pathological and genomic predictors of pCR have been described. Hatzis et al
. used gene-expression microarrays to generate a gene set encompassing modules for response to endocrine therapy and cytotoxic chemotherapy, which identified patients likely to undergo pCR and to have longer survival [
5]. The proportion of tumour infiltrating lymphocytes has also been shown to predict pCR in several studies of neoadjuvant chemotherapy [
6‐
12]. Automated quantitative estimates of tumour morphology using digital images of tissue sections have been shown to be associated with prognosis [
13,
14]. Therefore, comparable computational analysis of histological sections might provide a similar method for prediction of pCR.
We hypothesized that systematic quantitative analysis of tumour morphology at diagnosis would objectively identify tissue characteristics associated with pCR. We undertook a digital pathology study using a newly developed image processing method for single cell detection and material from the Neo-tAnGo randomized controlled trial [
15], both from diagnosis and at surgery, in order to objectively identify tissue features associated with pCR and to investigate changes in quantitative morphological metrics between pre-treatment and post-treatment samples and their relationship to pCR.
Discussion
We used computational pathology to generate image metrics of both pre- and post-treatment biopsies in a randomized controlled trial of neoadjuvant chemotherapy in breast cancer in order to investigate associations with chemosensitivity. Median lymphocyte density in pre-treatment biopsies emerged as the best predictor of response to chemotherapy, improving prediction based on known clinical factors. In addition, change in lymphocyte density between pre and post-treatment samples revealed that an increase in lymphocyte density was, paradoxically, associated with relative chemoresistance.
Computational pathology was used here to generate objective quantitative estimates of the cellular composition of tissue samples and, importantly, of the spatial heterogeneity of different cell types across a tissue section. Median density of lymphocytes, a spatial estimate, outperformed simple cellular quantification. Similarly, we have previously reported that the spatial distribution of stromal cells is prognostic in breast cancer and that this feature is not easily measurable by genomic assays [
13]. Previous work in the context of breast and prostate cancer, has demonstrated the capacity of a computational approach to interrogate spatial, relational and geometric features of tissues for outcome prediction [
14,
19,
20]. Many of these parameters could not be practically estimated by other means, and in this respect machine learning can provide deeper insight into tissue morphology than is possible by manual evaluation.
We performed a data-driven selection of tissue features associated with pCR. Median lymphocyte density emerged as the best predictor of chemosensitivity. In a previous study we have reported an association between automated estimates of lymphocytic infiltration and breast cancer survival in ER-negative disease [
13,
21]. Similarly, studies of patients who received neo-adjuvant chemotherapy based on genomic assays and histopathology have also reported an association between the immune response and pCR [
22]. Ignatiadis et al
. conducted a meta-analysis of gene-expression data from pre-surgical specimens in 996 patients and investigated associations between 17 previously reported gene modules and pCR [
22]. They found that the gene modules most reliably associated with pCR across cancer subtypes were those relating to the immune response. Tumour infiltrating lymphocytes estimated by a pathologist from H&E sections have also been found to be associated with outcome and response to chemotherapy [
8,
23], including some in the neo-adjuvant setting [
6,
10], largely in accord with the findings of this study.
To our knowledge, this is the first report showing that an increase in lymphocyte density following the perturbation of chemotherapy is associated with a lower likelihood of pCR. Previous studies have also shown that the composition of the post-treatment immune repertoire is associated with survival [
24,
25]. In addition, we found that the sequence in which chemotherapy agents were administered affected the strength of this association. Where patients received a taxane second, increased lymphocyte density was more strongly associated with relative resistance to chemotherapy than in patients who received a taxane first.
A limitation of this study is that we were not able to digitize and analyse samples from all patients enrolled in the Neo-tAnGo trial and that associated clinical data were not complete, leading to a loss of around one third of observations in multivariate analyses. This is inevitable in the context of large multicentre trials. A second limitation is that by using H&E sections we have not accounted for the immune phenotype or functional state of infiltrating lymphocytes.
The functional basis of the interaction we observed between the immune response and chemotherapy is uncertain. It should, however, be noted that all patients received an anthracycline (epirubicin) as part of their treatment. Anthracyclines have been extensively investigated in pre-clinical studies as a chemotherapeutic agent with a tumoricidal effect that can in part be attributed to stimulation of the immune response [
26]. For example, a recent study reported that upon exposure to an anthracycline tumour cells produce type I interferons evoking an immunological cascade reminiscent of that seen in cells infected by a virus and that this effect may be partly explained by the release of self-RNAs by dying cells [
27]. That greater clinical benefit of anthracyclines is significantly associated with the presence of a pre-existing immune response (tumour-infiltrating lymphocytes) has also been shown in several clinical studies [
8,
23,
28]. However, our results suggest that the effect of chemotherapy may be more complex than simply boosting pre-existing immune attack. First, we find that where the density of lymphocytes is increased following treatment, fewer tumours undergo pCR. That is, in this subset of around one quarter of patients, tumour cells apparently continue to resist the effects of immune attack in spite of its increased intensity. Second, we find that this effect is significantly greater where a taxane (paclitaxel) was administered after, as opposed to before, other agents. In the Neo-tAnGo trial, giving paclitaxel before the other agents significantly increased the proportion of cases with pCR [
15]. While the association between increased post-treatment lymphocyte density and treatment resistance holds whether paclitaxel is given first or second, the effect is significantly larger in cases where it is given second. Collectively, these findings raise the possibility that not only is a chemotherapy-stimulated immune response not universally effective, but that the efficacy of this response can be influenced by the sequence in which tumour cells are exposed to different chemotherapeutic agents, most notably taxanes and anthracyclines.
The existence of a substantial subgroup of relatively resistant tumours in which lymphocyte density is increased following chemotherapy further suggests a clinical opportunity. The variability in the immune response between primary breast tumours is well-known and recent genomic analyses suggest that some of this difference may be explained by the mutational burden of the primary tumour [
29]. Analyses of clinical trials of immune checkpoint inhibitors report that responses are best where there is a significant pre-existing immune response to the primary tumour [
30,
31]. Given that a large subset of breast tumours evoke only a mild immune response if any [
23,
32], methods for increasing immune attack against immunologically quiescent tumours are needed. Therefore the subgroup we have identified may benefit from receiving chemotherapy first, to increase the immune response, followed by immune checkpoint inhibitors to amplify its effect.
Digitization of pathology slides from clinical trials affords the important advantages of providing an enduring archive of tumour pathology and the opportunity for systematic image analysis. We anticipate analyses such as ours becoming more common as digital pathology is implemented more widely in clinical trials. We provide a valuable resource of digital pathology images to the research community, together with all our image-analysis codes. In addition, linked multiplatform genomic annotation (gene expression, copy number, targeted sequencing) will be made available for a subset of cases, following primary reporting of the data. While invaluable resources such as the Human Protein Atlas already provide access to an enormous array of tissue images [
33], some already utilized in translational studies [
34,
35], widely available digital images of tumour tissue from large high-quality clinical studies with molecular annotation such as ours are currently exceedingly rare. However, general access to such resources will be necessary to fulfil the potential of computational pathology as a novel modality that spans the research and clinical arenas.
Competing interests
The authors declare that they have no competing interests.
Authors’ contributions
HRA and CC conceived of and designed the study. AD and HRA conducted image analysis. HRA conducted statistical analyses. NW and MJI supervised image analysis methods. HB conducted slide scanning and associated data curation. EP was the lead trial pathologist and conducted pathology review of pre-treatment biopsies. LH was the trial statistician and provided matched, curated clinical data. HME, A-LV, LH, JD, SB and CC led the Neo-tAnGo trial. JA, MI, TH, KM and SH recruited patients and provided samples. HRA, AD, CC, PDP, JB, MI and NW led collaborative studies between Oncology and Astronomy. HRA and CC wrote the manuscript and all authors provided edits and agreed to its publication.