Caveolin Gene Expression Predicts Clinical Outcomes for Early-Stage HER2-Negative Breast Cancer Treated with Paclitaxel-Based Chemotherapy in the GeparSepto Trial

PURPOSE

Caveolin-1 and -2 (CAV1/2) dysregulation are implicated in driving cancer progression and may predict response to nab-paclitaxel. We explored the prognostic and predictive potential of CAV1/2 expression for patients with early-stage HER2-negative breast cancer receiving neoadjuvant paclitaxel-based chemotherapy regimens, followed by epirubicin and cyclophosphamide.

EXPERIMENTAL DESIGN

We correlated tumor CAV1/2 RNA expression with pathologic complete response (pCR), disease-free survival (DFS), and overall survival (OS) in the GeparSepto trial, which randomized patients to neoadjuvant paclitaxel- versus nab-paclitaxel-based chemotherapy.

RESULTS

RNA sequencing data were available for 279 patients, of which 74 (26.5%) were hormone receptor (HR)-negative, thus triple-negative breast cancer (TNBC). Patients treated with nab-paclitaxel with high CAV1/2 had higher probability of obtaining a pCR [CAV1 OR, 4.92; 95% confidence interval (CI), 1.70-14.22; P = 0.003; CAV2 OR, 5.39; 95% CI, 1.76-16.47; P = 0.003] as compared with patients with high CAV1/2 treated with solvent-based paclitaxel (CAV1 OR, 0.33; 95% CI, 0.11-0.95; P = 0.040; CAV2 OR, 0.37; 95% CI, 0.12-1.13; P = 0.082). High CAV1 expression was significantly associated with worse DFS and OS in paclitaxel-treated patients (DFS HR, 2.29; 95% CI, 1.08-4.87; P = 0.030; OS HR, 4.97; 95% CI, 1.73-14.31; P = 0.003). High CAV2 was associated with worse DFS and OS in all patients (DFS HR, 2.12; 95% CI, 1.23-3.63; P = 0.006; OS HR, 2.51; 95% CI, 1.22-5.17; P = 0.013), in paclitaxel-treated patients (DFS HR, 2.47; 95% CI, 1.12-5.43; P = 0.025; OS HR, 4.24; 95% CI, 1.48-12.09; P = 0.007) and in patients with TNBC (DFS HR, 4.68; 95% CI, 1.48-14.85; P = 0.009; OS HR, 10.43; 95% CI, 1.22-89.28; P = 0.032).

CONCLUSIONS

Our findings indicate high CAV1/2 expression is associated with worse DFS and OS in paclitaxel-treated patients. Conversely, in nab-paclitaxel-treated patients, high CAV1/2 expression is associated with increased pCR and no significant detriment to DFS or OS compared with low CAV1/2 expression.

CD47-blocking Antibody ZL-1201 Promotes Tumor-associated Macrophage Phagocytic Activity and Enhances the Efficacy of the Therapeutic Antibodies and Chemotherapy

Tumor-associated macrophages (TAM) are the most abundant immune cells in the tumor microenvironment. They consist of various subsets but primarily resemble the M2 macrophage phenotype. TAMs are known to promote tumor progression and are associated with poor clinical outcomes. CD47 on tumor cells and SIRPα on TAMs facilitate a "don't-eat-me" signal which prevents cancer cells from immune clearance. Therefore, blockade of the CD47-SIRPα interaction represents a promising strategy for tumor immunotherapy. Here, we present the results on ZL-1201, a differentiated and potent anti-CD47 antibody with improved hematologic safety profile compared with 5F9 benchmark. ZL-1201 enhanced phagocytosis in combination with standards of care (SoC) therapeutic antibodies in in vitro coculture systems using a panel of tumor models and differentiated macrophages, and these combinational effects are Fc dependent while potently enhancing M2 phagocytosis. In vivo xenograft studies showed that enhanced antitumor activities were seen in a variety of tumor models treated with ZL-1201 in combination with other therapeutic mAbs, and maximal antitumor activities were achieved in the presence of chemotherapy in addition to the combination of ZL-1201 with other mAbs. Moreover, tumor-infiltrating immune cells and cytokine analysis showed that ZL-1201 and chemotherapies remodel the tumor microenvironment, which increases antitumor immunity, leading to augmented antitumor efficacy when combined with mAbs.

SIGNIFICANCE

ZL-1201 is a novel anti-CD47 antibody that has improved hematologic safety profiles and combines with SoC, including mAbs and chemotherapies, to potently facilitate phagocytosis and antitumor efficacy.

Abstract 3171: Deep-learning image-based features stratify risk in HER2- breast cancer patients

Background: Tumor microenvironment (TME) characteristics are gaining acceptance as important biomarkers across all subtypes of breast cancer (BC). For example, stromal Tumor Infiltrating Lymphocytes (sTILs) have been demonstrated to be predictive in triple-negative and HER2+ BC on immune therapies (Hudeček, 2020), and TILs have been associated with recurrence-risk in HER2- BC (Kolberg-Liedtke, 2020). However, standardizing TME biomarkers remains a challenge to these prognostic & predictive studies (Kos, 2020).

Methods: Diagnostic H&E-stained pathology images from 506 HER2- breast cancer patients were acquired from TCGA sources. Pre-trained convolutional neural networks were used to classify each 100μm2 region as containing tumoral, stromal, and lymphocyte-infiltrated tissue as well as map their spatial co-distribution. Nine TME summary features were derived from these spatial maps, including total lymphocyte area, tumor-infiltrating lymphocytes (iTILs), stromal-infiltrating lymphocytes (sTILs), and tumor-adjacent lymphocytes (aTILs). Prognostic models relating these TME features to risk were fitted using Cox multiple-regression trained on 60% of patients and tested in the remaining 40%. Additional Cox models that incorporate seven standard clinicopathological features such as TNM staging, age, ethnicity, treatment type, and hormone-receptor status were also analyzed to establish independence of the TME features.

Results: A prognostic model developed using 9 TME-summarizing features accurately stratified unseen patients (HR=0.67 p=0.002) into high-risk (N=95) and low-risk (N=107) categories. Interestingly, tumor-adjacent stroma was significantly associated with higher risk (proportional HR=1.02, p=0.005) whereas tumor-infiltrating stroma was associated with lower risk (proportional HR=0.97, p=0.02). Incorporating standard clinicopathological features increased prognostic performance in test patients (HR=0.63, p=0.0005). The prognostic effect of tumor-adjacent and tumor-infiltrating stroma remained significant in multiple regression with clinicopathological features (p=0.003 & p=0.02 respectively).

Conclusions: Here we present using machine-vision to automate and standardize describing the TME, as well as demonstrate the prognostic potential of these descriptions by successfully stratifying risk in HER2- breast cancer. These TME descriptions add independent prognostic power to standard clinicopathological features. Machine-vision tools that produce interpretable features such as this can inform current pathology practices, as well as provide facile and scalable biomarkers for clinical studies going forward.

Citation Format: Mustafa I. Jaber, Liudmila Beziaeva, Stephen C. Benz, Shahrooz Rabizadeh, Patrick Soon-Shiong, Christopher W. Szeto. Deep-learning image-based features stratify risk in HER2- breast cancer patients [abstract]. In: Proceedings of the American Association for Cancer Research Annual Meeting 2021; 2021 Apr 10-15 and May 17-21. Philadelphia (PA): AACR; Cancer Res 2021;81(13_Suppl):Abstract nr 3171.

Abstract PO-030: Deep-learning image-based tumor, stroma, and lymphocytes spatial relationships and clinicopathological features that affect survival in pancreatic cancer patients

Background: Stromal and lymphocyte density have each been implicated in differential survival in pancreatic cancer (Torphy, 2018 & Orhan, 2020). In this study, we developed an automated deep-learning system to provide risk-assessment upon spatial relationships between tumor, stroma, and lymphocyte regions in pancreatic pathology images. Methods: Diagnostic H&E-stained pathology images from 82 pancreatic adenocarcinoma patients who underwent chemotherapy were acquired from TCGA sources. Thirty-two patients were held out for testing purposes. Tumor, stroma, and lymphocytes image masks were generated using pre-trained convolutional neural networks, and their co-distribution was summarized in nine numerical image-based features. Optimal thresholds in these image-based features were identified using 2-way Gaussian mixture models. This process found four spatial image features that significantly contributed to low-risk of early death: Low lymphocyte count, lymphocytes adjacent to tumor regions, and stromal adjacency to tumor regions. Ability to separate patients based on these features was evaluated using silhouette score, concordance index, and Cox proportional hazards ratios (HR). Results: Without using image-based features, exhaustive search of this cohort’s clinicopathological features found an optimal Cox proportional hazards model can yield a HR = 0.22 (p = 0.02) in 50 training examples and HR = 0.41 (p = 0.29) on 32 unseen test patients, ultimately utilizing just pathologically-determined T, and N information. The developed image-based risk predictor improved performance with HR = 0.51 (p = 0.06) on training data and HR = 0.52 (p = 0.09) on unseen test data. Combining the image-based risk models to selected clinicopathological features enhanced performance further to HR = 0.25 (p = 0.01) on the training set and HR = 0.37 (p = 0.07) on unseen test patients. Conclusions: Our interpretable image-based risk predictor shows high-risk pancreatic cancer patients have higher lymphocyte count overall but proportionally fewer tumor-infiltrating lymphocytes (TILs). In addition, this system shows high-risk patients have less stromal tissue within 100um from tumor compared to low risk patients. By aggregating both standard clinicopathological features with the proposed image-based risk assessment, superior separation in survival curves was achieved for both training and testing sets compared to either risk-model alone. Thus, our study demonstrates that image-based risk-associated features are independently prognostic of clinicopathological features. Despite the very limited sample-size of similarly-treated patients within the training dataset, these results trend towards significance and warrant further study within a larger cohort.

Citation Format: Mustafa I. Jaber, Liudmila Beziaeva, Robert J. Torphy, Stephen C. Benz, Shahrooz Rabizadeh, Patrick Soon-Shiong, Christopher W Szeto. Deep-learning image-based tumor, stroma, and lymphocytes spatial relationships and clinicopathological features that affect survival in pancreatic cancer patients [abstract]. In: Proceedings of the AACR Virtual Special Conference on Pancreatic Cancer; 2020 Sep 29-30. Philadelphia (PA): AACR; Cancer Res 2020;80(22 Suppl):Abstract nr PO-030.

Abstract 6712: Analysis of VEGF as a potential complimentary target with immunotherapy

Background: Angiogenesis is a hallmark of malignancy and greatly contributes to tumor growth and dissemination to metastatic sites. Vascular endothelial growth factors (VEGFs) are major members of angiogenesis signaling and have been shown to have immunosuppressive function. Bevacizumab (Bev), a VEGF inhibitor, in combination with immune checkpoint inhibitor (ICI) therapy and chemotherapy, has been shown in a previous clinical trial to improve outcomes in lung cancer versus chemotherapy plus Bev, even in PDL1-low and negative tumors. (IMPOWER150). Tumor mutation burden (TMB) has been established as a biomarker of ICI response, independent of PDL1. We set out to explore the relationship between VEGF signaling and TMB.

Methods: We performed a retrospective analysis of data from the NantHealth database of 2,737 commercial cases where matched WGS or WES and RNA-Seq were performed. Primary and metastatic breast tumors were included in the analysis. TMB was computed as the count of SNVs per tumor sample. Bowtie2, RSEM, and custom software are used for alignment, transcript quantification and variant analysis. Per-tumor type analysis was performed to find VEGF signaling correlates with TMB and a panel of immune markers (n = 9).

Results: Overall TMB does not correlate with PDL1, PDL2, CTLA4 and other immune markers' expression (spearman range [-0.01 - 0.15]). We also observed that overall TMB does not exhibit strong correlations with VEGF signaling markers (spearman range [-.17 - .17]), however when limited to TMB-low tumors VEGFA and VEGFD positively correlate with TMB (p = 1e-5 and 4.1e-2) and VEGFE negatively correlates with TMB (p = 1.6e-4). We then limited our analysis to PDL1-low tumors, of which the top three categories by sample counts are breast (n = 103), colon (n = 85), and sarcoma (n = 78). In whole-cohort analysis, we observed significantly higher expression of VEGFE and VEGFR2 (p = 1.4e-2 and 2.8e-2) in TMB-low tumors. When performing the same analysis in individual tissue types we observed significantly differentially higher expression of VEGFA and PGF (p = 2.1e-2 and 3.3e-2) in TMB-high colon samples, of VEGFR2 (p = 1.5e-2) in TMB-low NSCLC tumors, and of VEGFR2, FOS, and FOXO1(p = 1.9e-2, 3.4e-2, 6.2e-3) in TMB-low uterine endometrial tumors.

Conclusions: We provide evidence that targeting VEGF signaling in some tumor types, especially in tumors with low TMB, might provide a boost to therapeutic effects of checkpoint inhibitors. Our findings warrant further translational interrogation of the potential efficacy with a combination of VEGF inhibitors and ICI in various cancer types. For example, NCT03971474, a trial of ramucirumab (anti-VEGFR2 MAb) plus pembrolizumab in NSCLC, may produce differential activity in TMB-High v TMB-Low cohorts.

Citation Format: Yulia Newton, Christopher W. Szeto, Stephen C. Benz, Sandeep K. Reddy. Analysis of VEGF as a potential complimentary target with immunotherapy [abstract]. In: Proceedings of the Annual Meeting of the American Association for Cancer Research 2020; 2020 Apr 27-28 and Jun 22-24. Philadelphia (PA): AACR; Cancer Res 2020;80(16 Suppl):Abstract nr 6712.

Abstract 2224: Validated differential expression of immunoregulatory molecules that coincide with targetable mutations may provide novel insights into strategic trial design for therapeutics

Background: Immune checkpoint inhibitors (ICI) are rarely administered as first-line monotherapies for cancer. ICI are currently being trialed as combination ICI therapies, ICI in combination with standard chemo-(or other) therapies, or as second-line therapies following relapse, but many of these trials fail due to lack of optimal design. Here we sought to identify which checkpoint genes are differentially expressed (DE) in the presence of therapeutically targetable DNA mutations prior to treatment in order to facilitate the rational design of clinical trials for first-line ICI combination therapies.

Methods: Whole-exome (WES) variant calls and whole-transcriptome expression values (RNAseq) were acquired for 5767 samples across 28 solid tumor subtypes from TCGA sources including breast (N=981), thymic (N=404), melanoma (N=344), prostate (N=331), gliobastoma multiforme (GBM, N=289) among others. A curated list of 274 targetable single nucleotide variants (SNVs) was obtained from ImmunityBio sources based on FDA labels, literature review, and trial notes. Ten checkpoint genes significantly DE between targetable SNV mutant (mt) vs. wild-type (wt) were identified by t-tests corrected for multiple hypothesis testing. Checkpoint DE was then validated as significant in an external cohort of 2739 unselected later-stage clinical cases from the NantHealth database with similarly profiled paired WES & RNAseq, comprised of breast (N=576), colon (N=314), lung (N=283), pancreatic (N=221), ovarian (N=196), among others. Tissue subtype enrichment for targetable mutations was assessed by Fisher's exact test.

Results: Twenty-three significant associations between targetable mt and DE checkpoint genes were identified in TCGA cases; 10 were validated in the external cohort. The vemurafenib target BRAF V600E was found coinciding with increased PD1, PDL1, and CTLA4 (adj. p=2.4e-3, 1.1e-23, 6.6e-7 respectively) as well as decreased IDO1 (adj. p=3.4e-6), and the effect size was larger than that of tissue-type. TIM3 was found significantly elevated in lapatinib-sensitive EGFR G598V patients (adj. p=1.0e-5) most prevalent in GBM, and conversely suppressed in FGFR3 S249C patients (adj. p=0.04) that are significantly enriched in bladder cancers. PIK3CA E545K patients, mostly cervical cancers, showed higher IDO1 expression (adj. p=3.3e-4) suggesting sensitivity to combined alpelisib/epacadostat.

Conclusions: NGS data inform decisions for use of gene mutation-targeted therapies; use of similar analysis when paired with RNAseq may support efforts to replace chemotherapy with more efficacious/safer combined immuno- and mutation-targeted therapies. Our findings here suggest future studies may result in the optimization of ICI - gene targeted therapy trials.

Citation Format: Jacob J. Adashek, Christopher W. Szeto, Saihitha Veerapaneni, Andrea Preble, Sandeep K. Reddy, Philippe E. Spiess. Validated differential expression of immunoregulatory molecules that coincide with targetable mutations may provide novel insights into strategic trial design for therapeutics [abstract]. In: Proceedings of the Annual Meeting of the American Association for Cancer Research 2020; 2020 Apr 27-28 and Jun 22-24. Philadelphia (PA): AACR; Cancer Res 2020;80(16 Suppl):Abstract nr 2224.

Transcriptomic silencing as a potential mechanism of treatment resistance

Next-generation sequencing (NGS) has not revealed all the mechanisms underlying resistance to genomically matched drugs. Here, we performed in 1417 tumors whole-exome tumor (somatic)/normal (germline) NGS and whole-transcriptome sequencing, the latter focusing on a clinically oriented 50-gene panel in order to examine transcriptomic silencing of putative driver alterations. In this large-scale study, approximately 13% of the somatic single nucleotide variants (SNVs) were unexpectedly not expressed as RNA; 23% of patients had ≥1 nonexpressed SNV. SNV-bearing genes consistently transcribed were TP53, PIK3CA, and KRAS; those with lower transcription rates were ALK, CSF1R, ERBB4, FLT3, GNAS, HNF1A, KDR, PDGFRA, RET, and SMO. We also determined the frequency of tumor mutations being germline, rather than somatic, in these and an additional 462 tumors with tumor/normal exomes; 33.8% of germline SNVs within the gene panel were rare (not found after filtering through variant information domains) and at risk of being falsely reported as somatic. Both the frequency of silenced variant transcription and the risk of falsely identifying germline mutations as somatic/tumor related are important phenomena. Therefore, transcriptomics is a critical adjunct to genomics when interrogating patient tumors for actionable alterations, because, without expression of the target aberrations, there will likely be therapeutic resistance.

Real-world data validation for differential expression of immunoregulatory molecules and targetable cancer genes may provide therapeutic insights into agnostic-driven trial designs

10

Background: Targeting actionable genes and using immunotherapy have increased treatment options. We previously reported that some immunoregulatory molecules are found differentially regulated in the presence of certain gene mutations regardless of cancer subtype. Here we validated a subset of these associations in an independent, real-world dataset with distinct clinicopathological characteristics. Methods: Previously, 2740 TCGA patients were identified to have at least one potentially oncogenic mutation (mt) within an established 50-gene hotspot panel. Differential expression of 10 immunoregulatory molecules (IRM) was analyzed between mutant (mt) vs. wild-type (wt). To ensure observed significant associations were not confounded by tumor-type, differential IRM expression within mt-enriched tumor-types was compared to that of mt vs. wt. Now, using the NantHealth external database of 2739 unselected clinical cases these associations were validated. Results: Within the TCGA cohort 19/50 gene mutations were found to be significantly associated with ≥1 IRM expression. In many, the mt effect-size was larger than that of tumor-type; e.g. head & neck carcinomas (HNSCC) are highly enriched for CDKN2A mt (OR = 4.9, p = 4.3e-9), yet CDKN2A mt are more associated with CTLA4 expression than HNSCC histology (t = 7.0 vs. 5.4). Of these 15 associations, 6 were validated within the independent later-stage NantHealth cohort. Most notably, CDKN2A mt was validated as associated with increased PD1 and CTLA4 expression while KRAS and APC mt were validated as associated with decreased PDL1/2 expression. Conclusions: The presented differential checkpoint expression patterns are strongly associated with mutation status and are not primarily driven by tissue-type, which have been further validated by an external database. Strategies combining genomic targets have been shown to yield success as well as using immunotherapies. Our data suggests there may be a role for combining NGS targets along with IRM expression patterns to better guide future design of clinical trials in combatting various cancers in a tissue-agnostic fashion.

A deep learning image-based intrinsic molecular subtype classifier of breast tumors reveals tumor heterogeneity that may affect survival

BACKGROUND

Breast cancer intrinsic molecular subtype (IMS) as classified by the expression-based PAM50 assay is considered a strong prognostic feature, even when controlled for by standard clinicopathological features such as age, grade, and nodal status, yet the molecular testing required to elucidate these subtypes is not routinely performed. Furthermore, when such bulk assays as RNA sequencing are performed, intratumoral heterogeneity that may affect prognosis and therapeutic decision-making can be missed.

METHODS

As a more facile and readily available method for determining IMS in breast cancer, we developed a deep learning approach for approximating PAM50 intrinsic subtyping using only whole-slide images of H&E-stained breast biopsy tissue sections. This algorithm was trained on images from 443 tumors that had previously undergone PAM50 subtyping to classify small patches of the images into four major molecular subtypes-Basal-like, HER2-enriched, Luminal A, and Luminal B-as well as Basal vs. non-Basal. The algorithm was subsequently used for subtype classification of a held-out set of 222 tumors.

RESULTS

This deep learning image-based classifier correctly subtyped the majority of samples in the held-out set of tumors. However, in many cases, significant heterogeneity was observed in assigned subtypes across patches from within a single whole-slide image. We performed further analysis of heterogeneity, focusing on contrasting Luminal A and Basal-like subtypes because classifications from our deep learning algorithm-similar to PAM50-are associated with significant differences in survival between these two subtypes. Patients with tumors classified as heterogeneous were found to have survival intermediate between Luminal A and Basal patients, as well as more varied levels of hormone receptor expression patterns.

CONCLUSIONS

Here, we present a method for minimizing manual work required to identify cancer-rich patches among all multiscale patches in H&E-stained WSIs that can be generalized to any indication. These results suggest that advanced deep machine learning methods that use only routinely collected whole-slide images can approximate RNA-seq-based molecular tests such as PAM50 and, importantly, may increase detection of heterogeneous tumors that may require more detailed subtype analysis.

Abstract B093: 50 gene breast cancer (BC) RNA subtype classifier and colorectal cancer (CRC) CMS subtype classifier applied to 90 prostate cancer (PC) patients reveals distinct subtype differences

Background: BC and CRC are heterogeneous diseases with several distinct disease subtypes. A 50-gene qPCR assay (PAM50) identifies 5 intrinsic biological subtypes: luminal A, luminal B, HER2-enriched, basal-like, and normal-like in breast cancer. Consensus molecular subtype in CRC (CMS) identifies 4 subtypes: CMS1-4. PC can be classified by specific genomic rearrangements, such as the TMPRSS2-ERG fusion occur or whether specific mutations such as SPOP and FOXA1 or germline BRCA is present, yet 25% of PC remains difficult to classify based on molecular drivers. We used comprehensive genomic and transcriptomic data to determine the association of somatic and germline mutations in molecular PC subtypes using the CMS and 50-gene breast cancer classifiers. Methods: Retrospective analysis on Whole exome (WES) DNA tumor and paired germline and matched deep whole transcriptomic sequencing (RNA-Seq) (∼200x106reads per tumor) data from NantHealth was performed. BC Intrinsic Subtypes and CMS sorting based on RNAseq assay was used to classify PC into 5 BC subtypes and CMS 1-4. Results: 90 PC patients were classifiable using RNAseq and WES DNA. BC subtype distribution was 64.4% Luminal A, 31.1% Luminal B, 2.2% HER2-enriched, 2.2% normal-like, and 0 classified as Basal-like. The CMS subtype distribution was 83.3% CMS4, 15.6% CM1, 1.1% CMS3, and 0 CMS2. Overlap comparison revealed that almost all LumA are CMS4 (56/58, 96.6%) and CMS1 are LumB (11/14, 78.6%). Analysis of germline DDR mutations in LumB compared to LumA showed LumB has significantly more DDR mutants than LumA (OR 2.68, p = 0.035 one-sided Fishers exact test). Conclusions: PC molecularly profiled using 50-gene and CMS classification we found LumA subtype as the predominant subgroup characterized by low TMB and high CMS4. LumB was characterized as higher TMB and greater CMS1 suggesting potential immunotherapy eligibility in this subtype. The presence of potentially pathogenic germline variants in DDR genes is associated with subsequent somatic increase in TMB, however, somatic expression subtyping (by BC or CRC types) is a stronger indicator of TMB than germline pPV in DDR genes.

Citation Format: Jacob J Adashek, Sumanta K Pal, Saihitha Veerapaneni, Elias Obeid, Christopher W Szeto, Sandeep K Reddy, Roberto Nussenzveig, Neeraj Agarwal. 50 gene breast cancer (BC) RNA subtype classifier and colorectal cancer (CRC) CMS subtype classifier applied to 90 prostate cancer (PC) patients reveals distinct subtype differences [abstract]. In: Proceedings of the AACR-NCI-EORTC International Conference on Molecular Targets and Cancer Therapeutics; 2019 Oct 26-30; Boston, MA. Philadelphia (PA): AACR; Mol Cancer Ther 2019;18(12 Suppl):Abstract nr B093. doi:10.1158/1535-7163.TARG-19-B093

Abstract 1393: Automated adeno/squamous-cell NSCLC classification from diagnostic slide images: A deep-learning framework utilizing cell-density maps

Background: The most common form of lung cancer, non-small cell lung cancer (NSCLC), is further categorized into two major histopathological subtypes: ~40% Adenocarcinoma (LUAD), and ~30% squamous cell carcinoma (LUSC). Classifying patients accurately is important for prognosis and therapy decisions, but requires costly pathologist review. Here we present an automated algorithm to differentiate LUAD and LUSC diagnostic whole slide images (WSIs).

Methods: 488 subtyped NSCLC high-resolution diagnostic WSIs were obtained from TCGA sources. Adjacent normal regions were identified and excluded from analysis. Cancer cell-density maps were created based on cell counts within discrete patches. These maps were then binned into ten ranges of cell counts (20-30 cells per patch, 30-40, etc. up to >110 cells per patch). 2D color patches were transformed into 1D descriptive vectors using the inception v3 deep learning framework. Samples were randomly split into 70% training and 30% testing sets. Ten LUAD/LUSC linear SVM classifiers (one for each cell-density bin) were trained on such transformed data. Subtype prediction in unseen testing samples was achieved by averaging subtype predictions from the 10 subsequent models.

Results: 338 TCGA diagnostic WSIs (164 LUAD and 174 LUSC) were used to train, and 150 (71 LUAD and 79 LUSC) were used to test. The proposed system achieved an AUC of 0.9068 in test samples, corresponding to a classification accuracy of 83.33%. The (heretofore excluded) adjacent normal regions were classified correctly almost as accurately as tumor regions (74.7%).

Conclusions: This fully-automated histopathology subtyping method generates maps of regions-of-interest within WSIs, providing novel spatial information on tumoral organization. For example, our results on test data show tumor patches of size 100 square microns with 60 to 100 cells distinguish LUAD from LUSC better than other cell-density ranges. Moreover, adjacent normal tissue may provide additional insights into tumorigenesis mechanisms.

Citation Format: Mustafa I. Jaber, Liudmila Beziaeva, Christopher W. Szeto, John Elshimali, Shahrooz Rabizadeh, Bing Song. Automated adeno/squamous-cell NSCLC classification from diagnostic slide images: A deep-learning framework utilizing cell-density maps [abstract]. In: Proceedings of the American Association for Cancer Research Annual Meeting 2019; 2019 Mar 29-Apr 3; Atlanta, GA. Philadelphia (PA): AACR; Cancer Res 2019;79(13 Suppl):Abstract nr 1393.

Abstract 1188: Detecting intratumor heterogeneity of PAM50 subtypes in H&E-stained slides using deep learning

We present a methodology for identifying PAM50 intrinsic molecular breast cancer subtypes from only H&E-stained tissue section whole-slide images (WSI) of breast biopsies without using RNA expression data. We then use this system to identify patients presenting multiple subtypes simultaneously (i.e. intra-tumor heterogeneity), and validate the clinical utility of identifying patients with heterogeneous tumors.

Several methods have been proposed to stratify breast cancer subtypes including histological, immunohistochemical, and molecular. Intrinsic molecular subtypes such as PAM50 subgroups demonstrably outperform clinical factors and IHC in prognostic power. Yet molecular subtyping is fundamentally limited in two ways: 1) molecular characterization is relatively expensive and so not ubiquitously performed; and 2) non-single-cell molecular characterization assays the bulk tumor population, making studying intra-tumor heterogeneity difficult. The presented subtyping system uses routinely-gathered H&E stained WSIs to mimic molecular subtyping. Three modules form the proposed WSI-based subtyping system: First, WSIs are broken into multi-scale 400px x 400px patches and converted to descriptive tensors using the Inception-v3 neural net architecture. Next, a subset of cancer-enriched patches is automatically selected to summarize WSI tumor content and used in further analysis. Finally, each patch is assigned a subtype in a 4-way classifier (Basal, HER2-enriched, Luminal A, and Luminal B). Optionally, patient-based subtype classifications can be made by employing a voting mechanism upon the patch-based results.

We demonstrate this subtyping system using publicly available diagnostic WSIs from the TCGA BRCA cohort. We trained on 582 randomly selected patients, then tested subtyping accuracy on a held-out set of 223 patients. The subtype accuracy in held-out samples was 66% (compared to 34% from a random classifier, and 52% based on the majority-class classifier). We focused on 76 patients containing WSI patches with predictions for both Basal and Luminal A, and contrast them to 204 patients with majority Luminal A patches and 82 patients with majority Basal patches. We validated that this mixed-subtype population have outcomes and expression patterns that support a heterogeneous cellular population: The mixed subtype population have intermediate survival times between Luminal A and Basal in Kaplan-Meier analysis, varied hormone-receptor levels, and form a cluster equidistant between Luminal A and Basal in batch analysis.

These results demonstrate using readily-available data to characterize tumor subtypes and sub-populations. Correctly identifying these sub-populations may provide crucial additional information that is lost to bulk-tumor assays.

Citation Format: Mustafa Jaber, Bing Song, Clive R. Taylor, Charles J. Vaske, Christopher W. Szeto. Detecting intratumor heterogeneity of PAM50 subtypes in H&E-stained slides using deep learning [abstract]. In: Proceedings of the American Association for Cancer Research Annual Meeting 2018; 2018 Apr 14-18; Chicago, IL. Philadelphia (PA): AACR; Cancer Res 2018;78(13 Suppl):Abstract nr 1188.

Abstract 44: Building patient-specific predictors of drug responses from cell line genomics

Hundreds of model cancer cell lines from diverse tissues have now been comprehensively assayed using several different genome-scale technologies. Additionally, many of these model lines have been challenged with dozens of clinically available oncotherapies. Here we present a set of high-throughput technologies suitable for turning these diverse datasources into actionable, patient-specific diagnostic predictors of response.

First, we computationally integrate diverse -omics data from the DNA, mRNA and protein levels into a pathway model of the cellular state. Next, parallel compute clusters are used to develop and evaluate accurate predictive models upon these pathway activity levels. Finally, these predictive models are used to suggest therapies in a patient-specific manner.

Here we show results from applying these technologies to learning a simple 50-gene signature for response to the the tyrosine kinase inhibitor Dasatinib. This signature utilizes features from the TP53/FOXM1, FOS/JUN & MYC pathways. We note that the exceptional responders to Dasatinib are enriched for nervous-system cancer cell lines. We applied our predictive signature to glioblastoma multiforme samples to correctly indicate which specific GBM patients may have respond to Dasatinib.

Citation Format: Christopher W. Szeto, Stephen Benz, Charlies Vaske. Building patient-specific predictors of drug responses from cell line genomics. [abstract]. In: Proceedings of the AACR Precision Medicine Series: Integrating Clinical Genomics and Cancer Therapy; Jun 13-16, 2015; Salt Lake City, UT. Philadelphia (PA): AACR; Clin Cancer Res 2016;22(1_Suppl):Abstract nr 44.

Abstract 5085: TopModel: An online resource for predictive models in cancer

One goal of characterizing the genome-wide landscape of cancer cells is to identify predictive signatures of onset, progression, and treatment outcomes. Many computational approaches have been developed to discover gene signatures with a range of success. The challenge still remains to identify the best approach that, when trained on one cohort, remains accurate in predicting outcomes on an unseen cohort. Thus far, no clear themes have emerged that might provide clues about which method works for a particular task. We have built a system called TopModel that facilitates the identification of top-performing machine-learning algorithms for a series of cancer-genomics challenges. The four components of the system include: 1) a benchmark that includes several cancer genomics datasets with outcome variables as targets to predict; 2) a database of results derived from the application of thousands of machine-learning and feature selection combinations; 3) a web interface that allows bioinformaticians to evaluate their own prediction results; and 4) a web interface that allows a biomedical researcher to upload data on a sample or set of samples in order to receive a report on the signatures predicted to exist in the sample(s). The cancer benchmark component provides a common ground for the development and evaluation of prediction methods for variables such as cancer subtype, drug response, survival, and others. Several datasets have been loaded including predicting survival in the TCGA cohorts, and the hundreds of drug sensitivities in several cancer cell line cohorts. We demonstrate the utility of the resource by comparing state-of-the art feature selection methods to a new approach that uses locality on a genetic interaction network. We evaluate the performance in terms of how well the features generalize across datasets as a trade-off to the accuracy of prediction. In addition to identifying high-value genome features, we explore the robustness of the cancer state in the absence of these features. We simulate gene knock outs by disconnecting these features in our pathway models, inferring the pathway interaction network in the absence of these features, and then reassessing using the top-performing predictive models of cancer.

Citation Format: {Authors}. {Abstract title} [abstract]. In: Proceedings of the 103rd Annual Meeting of the American Association for Cancer Research; 2012 Mar 31-Apr 4; Chicago, IL. Philadelphia (PA): AACR; Cancer Res 2012;72(8 Suppl):Abstract nr 5085. doi:1538-7445.AM2012-5085