Predictive biomarkers for immune therapy must address a complex interface between the immune system and triple-negative breast cancer and still be technically reliable for diagnostic use. Two recent articles describe the assessment of spatial heterogeneity using digital methods that promise to improve the quantification of immune infiltrate or molecular targets.

See related articles by Bai et al., p. 5557 and Carter et al., p. 5628

In this issue of Clinical Cancer Research, Bai and colleagues (1) and Carter and colleagues (2) lead us in the direction of spatial resolution of the interface between immune environment and the invasive cancer as a potential biomarker strategy for immune–oncology treatment of triple-negative breast cancer (TNBC).

Our immune system might arguably be more complex and dynamically adaptive than a TNBC. So, how should we measure the immune–tumor interface that is neither uniform within the tumor nor challenged yet by the immune-oncology treatment we might then select to administer to augment an effective immune response to the cancer? The complexity of this challenge is profoundly important as a minority subset of those treated will realize significant benefits of tumor response and improved survival outcome, but many probably won't (3, 4). Adding to this issue is the currently imperfect efficacy of IHC evaluation of PD-L1 as integral biomarker, that was recently illustrated in phase III clinical trials, and the confusion, imprecision, and relative inaccuracy associated with interpreting and reporting this staining in primary or metastatic samples when implemented into broad diagnostic use (5, 6). Clearly, there is urgent need to translate new approaches to interpreting the immune activity in TNBC to predict the likelihood of activating immune response to achieve cure. Usual approaches of staining the therapeutic target or molecular signatures from a tumor sample in toto have not solved this challenge, so perhaps biomarkers that assess the more complex intersection of immune system and cancer might succeed.

The two articles employ different approaches and different technologies, but they both attempt to interpret activity of immune microenvironment across the histopathologic landscape of a cancer sample (1, 2). The article by Bai and colleagues describes their use of digital pathology and artificial intelligence methods to find the most prognostic model to quantify immune infiltrate from routine hematoxylin and eosin stains (1). Their neural network machine learning provided outputs of the proportion of tumor-infiltrating lymphocytes (TIL) relative to tumor cells, or stromal cells, or to all cells, either calculated from cell numbers or from area involvement. The five algorithms developed were all validated as prognostic after standard chemotherapy, with similar prognostic efficacy. Their method offers the potential to standardize interpretation method and reduce interobserver and intraobserver variance of quantitation of TILs from routine diagnostic slides (7).

The article by Carter and colleagues describes their use of spatial proteomic analysis to deconvolute the components of immune infiltrate in the tumor and their relationship to TNBC cells, supplementing this with bulk RNA sequencing to measure transcript abundance from the same area (2). The Nanostring technology that they employed uses area analysis of spatial relationships, albeit with less than the cellular resolution of histopathology. Nevertheless, they reported that TNBCs with PD-L1 expression were more enriched for dendritic cells, macrophages, B and T lymphocytes, fewer myeloid suppressor cells, and had higher expression of proteins such as HLA-DR, IDO-1, CD163, and CD40. Proteins associated with antigen presentation and stimulation of IFN genes (STING) were among those with slight enrichment in the intraepithelial component. Expression of the checkpoint protein CTLA-4 was lower in both compartments.

The immune infiltrate in TNBC is usually heterogeneously distributed (Fig. 1), except for medullary cancer that represents an uncommon and more indolent subtype, and this was addressed as a subset in the article by Carter and colleagues So, intratumoral heterogeneity of immune infiltrate is measurable and might then be tested for a predictive relevance that has been elusive so far. Similarly, the measurement of TILs has struggled to distinguish whether the intraepithelial or stromal component is the more informative functional interface between cancer and immune system. Stromal measurements of TILs from histopathology offer more reliable prognostication, but that might reflect more reliable method of interpretation. Results from Bai and colleagues demonstrated that intraepithelial and stromal TILs were similarly prognostic when neural network digital methods of interpretation were used. That finding could influence the interpretation of molecular activity using proteomic or genomic methods, as the components might be interpreted differently when machine learning is applied. Going forward, investigators would still need to be aware of the propensity for machine learning methods to overfit while training and to suffer regression to the mean in subsequent analyses. This same caution would also apply to the combination of multiparameter molecular measurements and digital pathology variables into an algorithm. So, stringent study design and repeated independent validation will be essential. Nevertheless, accurate and standardized interpretation is possible and would greatly help with reliable diagnostic implementation in the future.

Figure 1.

TNBC at a glance. Tumor-immune interface appears complex. Immune infiltrate is variable in intensity and distribution, and the intraepithelial and stromal components are present but difficult to interpret separately.

Figure 1.

TNBC at a glance. Tumor-immune interface appears complex. Immune infiltrate is variable in intensity and distribution, and the intraepithelial and stromal components are present but difficult to interpret separately.

Close modal

Both articles report early results that are somewhat descriptive. However, they both demonstrate the initial feasibility of a more complex analytic approach toward development of a robust immune biomarker. At some point, it is likely that these approaches of quantitative digital pathology algorithms and functional molecular characterization beyond cell type will converge and will use technologies that are amenable to routine diagnostic application. Of course, there is still more to understand about where to take measurements along the heterogeneous interface between cancer and immune infiltrate (intratumoral heterogeneity), and what characteristics of the pretreated tumor landscape (steady-state) will best represent the propensity for activation by immune-oncology agents (pharmacodynamic response). However, those approaches should probably be developed using samples from clinical trials of immune therapy, and perhaps also applied to on-treatment biopsies to evaluate pharmacodynamic response compared with baseline tumor sample. These reports describe approaches that are likely to be developed for application to more specific predictive hypotheses.

By analogy, during a brief summer respite while gazing down a valley in the Hill Country, I was reminded that the beauty and interest in a real landscape from this overall vantage might not provide specific details about the interface between humankind and nature, details of what's happening at each site where they interface or how different parts of the valley would respond to change or evolve with time. One needs to know where and when to look more closely to address those questions. Thus, I am also reminded of the limitations of analyzing small samples or relying on an average when interpreting a complex and varied relationship such as a functional tumor-immune interface. Perhaps spatial diversity of immune activation can predict whether there will be pockets of resistance or complete response when immune-oncology agents are added to the treatment of TNBC. Also, perhaps biomarker signals from relapsed metastatic disease (when immune depletion, tumoral molecular progression, and heterogeneity increase) represent different aspects or activity of tumor-stromal interface than residual tumor after neoadjuvant treatment (potentially emulating the clones of a future recurrence) or de novo disease and host immunity prior to neoadjuvant treatment (theoretically encouraging immune memory) or adjuvant treatment (directed at nascent and possibly dormant disease). For example, testing the archival untreated primary cancer sample might be less informative than a contemporaneous sample of metastasis. If so, then our diagnostic interpretation of biomarkers representing the tumor-immune interface should depend on clinical and treatment context, because predictive performance could be different in metastatic, neoadjuvant, and adjuvant treatment settings. I think these investigators have taken a promising methodologic path to characterize and measure the landscape of immune activity in a tumor sample. We should eagerly anticipate the refinement of their respective approaches and their eventual application to samples from clinical trials of immune therapy for adjuvant or neoadjuvant treatment of TNBC. Then we can infer function more precisely from the landscape that we see.

W.F. Symmans reports other support from Merck, is an unpaid member of a translational research committee for the Kaitlin trial (Genentech/Roche) and the PALLAS trial (Pfizer), has received research funding from Pfizer for pathology review of the NeoTALA trial, has intellectual property related to residual cancer burden (patent expired) and a molecular predictor of sensitivity to endocrine therapy (patent pending) licensed to Delphi Diagnostics, owns founder shares in Delphi Diagnostics and is an unpaid scientific advisor, and owns publicly traded shares in IONIS Pharmaceuticals and Eiger Biopharmaceuticals outside the submitted work. No other disclosures were reported by the author.

1.
Bai
Y
,
Cole
K
,
Martinez-Morilla
S
,
Ahmed
FS
,
Zugazagoitia
J
,
Staaf
J
, et al
An open source, automated tumor infiltrating lymphocyte algorithm for prognosis in triple-negative breast cancer
.
Clin Cancer Res
2021
;
27
:
5557
65
.
2.
Carter
JM
,
Polley
M-YC
,
Leon-Ferre
RA
,
Sinnwell
J
,
Thompson
KJ
,
Wang
X
, et al
Characteristics and spatially-defined immune (micro)landscapes of early-stage PD-L1-positive triple-negative breast cancer
.
Clin Cancer Res
2021
;
27
:
5628
37
.
3.
Schmid
P
,
Salgado
R
,
Park
YH
,
Muñoz-Couselo
E
,
Kim
SB
,
Sohn
J
, et al
Pembrolizumab plus chemotherapy as neoadjuvant treatment of high-risk, early-stage triple-negative breast cancer: results from the phase 1b open-label, multicohort KEYNOTE-173 study
.
Ann Oncol
2020
;
31
:
569
81
.
4.
Winer
EP
,
Lipatov
O
,
Im
S-A
,
Goncalves
A
,
Muñoz-Couselo
E
,
Lee
KS
, et al
Pembrolizumab versus investigator-choice chemotherapy for metastatic triple-negative breast cancer (KEYNOTE-119): a randomised, open-label, phase 3 trial
.
Lancet Oncol
2021
;
22
:
499
511
.
5.
Mittendorf
EA
,
Zhang
H
,
Barrios
CH
,
Saji
S
,
Jung
KH
,
Hegg
R
, et al
Neoadjuvant atezolizumab in combination with sequential nab-paclitaxel and anthracycline-based chemotherapy versus placebo and chemotherapy in patients with early-stage triple-negative breast cancer (IMpassion031): a randomised, double-blind, phase 3 trial
.
Lancet
2020
;
396
:
1090
100
.
6.
Reisenbichler
ES
,
Han
G
,
Bellizzi
A
,
Bossuyt
V
,
Brock
J
,
Cole
K
, et al
Prospective multi-institutional evaluation of pathologist assessment of PD-L1 assays for patient selection in triple negative breast cancer
.
Mod Pathol
2020
;
33
:
1746
52
.
7.
Salgado
R
,
Denkert
C
,
Demaria
S
,
Sirtaine
N
,
Klauschen
F
,
Pruneri
G
, et al
The evaluation of tumor-infiltrating lymphocytes (TILs) in breast cancer: recommendations by an International TILs Working Group 2014
.
Ann Oncol
2015
;
26
:
259
71
.