• No results found

Machine Learning, Artificial Intelligence, and Computational Modeling

P726

Applying deep learning classification for tumor identification across immunohistochemical markers on serial sections to eliminate the need for image co-registration

Mark Anderson, BSc, Mark Anderson, BSc , Karen McClymont, Lorcan Sherry, PhD, Alison Bigley, CSci, FIBMS

OracleBio, North Lanarkshire, United Kingdom

Correspondence:Mark Anderson ([email protected])

Journal for ImmunoTherapy of Cancer2019,7(Suppl 1):P726

Background

Artificial intelligence deep learning networks are being increasingly applied to resolve complex pattern recognition challenges in quanti- tative digital pathology. Khosravi[1], utilized deep learning to discrim- inate and classify tumors and associated subtypes using a range of immunohistochemically (IHC) labelled tumor sections, even where tu- mors were significantly heterogeneous.

Image co-registration is commonly employed in brightfield quantita- tive workflows where serial sections have been IHC labelled with dif- ferent markers, one of which is a tumor marker such as pan- Cytokeratin, which is used to guide the tumor locations in the corre- sponding serial section. Frequently, alignment errors occur in particu- lar at the periphery of tumor foci, where either tumor cells are missed or adjacent stroma is incorrectly classified as tumor.

Deep learning classifiers can be trained across a range of sections and markers to define tumor regions of interest (ROI), utilizing a range of features from each stained sample to generate a classifier, capable of identifying tumor ROI independent of the tissue, stain or marker.

Here we use serial tissue microarray (TMA) sections of gastric adeno- carcinomas, labelled with pan-Cytokeratin and CD3, to exemplify the use of a deep learning approach to distinguish tumor foci without re- quiring serial section image co-registration.

Methods

Exemplar TMA serial sections of gastric adenocarcinomas were IHC labelled for pan-Cytokeratin and CD3, using DAB chromogen and

counterstained with hematoxylin. The stained slides were digitized using a Zeiss scanner.

The sections were analyzed using Deeplabv3 network in Visiopharm Oncotopix Software. Classes were established for tumor, stroma and background, at x10 input magnification, using identical example training labels across the IHC images.

Results

Tumor and stroma were accurately identified across the serial sec- tions using a single deep learning algorithm. Applying a Dice similar- ity coefficient confirmed a high-level correlation between manual ROI and deep learning classifications for the 2 markers.

Conclusions

The exemplar data demonstrated the application of a single classifier, across IHC images, enabling the identification of tumor and stroma, irrespective of marker. This approach has demonstrated that it is pos- sible to accurately utilize multifaceted features, from different IHC markers on TMA tissue sections, for the accurate classification of tumor and stroma ROI. This approach provides an alternative method to the co-registration of images for tumor assignment across serial sections.

References

1. Khosravi P, Kazemi E, Imielinskid M, Elemento O and Hajirasouliha I. Deep Convolutional Neural Networks Enable Discrimination of Heterogeneous Digital Pathology Images. EBioMedicine 2018; 27: 317–328

P727

Single-cell systems neuroimmunology reveals immunosuppressive checkpoint blockade receptor expression correlates with

ventricular stem cell niche contact in human glioblastoma

Todd Bartkowiak, PhD1, Sierra Barone, BS1, Allison Greenplate, PhD1, Justine Sinnaeve, BS1, Nalin Leelatian1, Akshitkumar Mistry, MD2, Caroline Roe, BS1, Bret Mobley, MD1, Lola Chambless, MD2, Reid Thompson, MD2, Kyle Weaver, MD2, Rebecca Ihrie, PhD1, Jonathan Irish, PhD1

1

Vanderbilt University, Nashville, TN, United States;2Vanderbilt University Medical Center, Nashville, TX, United States

Correspondence:Rebecca Ihrie ([email protected]), Jonathan Irish ([email protected])

Journal for ImmunoTherapy of Cancer2019,7(Suppl 1):P727

Background

Glioblastomas make up more than 60% of adult primary brain tu- mors and carry a median survival of less than 15 months despite ag- gressive standard therapy. Immunotherapy, now standard of care for many solid tumors, offers an appealing alternative platform that may improve survival outcomes for patients with glioblastoma; however, predictive features that could inform responsiveness to different im- munotherapeutic modalities remains to be elucidated. Recent studies have demonstrated that patients whose tumors show radiographic contact with the ventricular-subventricular zone (V-SVZ) have dimin- ished survival outcomes compared to patients whose tumors do not contact the V-SVZ. We therefore hypothesized that V-SVZ contact may provide a unique, immunosuppressive microenvironment within the brain that promotes tumor growth by suppressing anti-tumor im- munity that may be overcome via targeting the appropriate immune suppressive immune population or receptor target.

Methods

Primary glioblastoma tumors obtained in accordance with the Declar- ation of Helsinki and with institutional IRB approval (#131870, #030372, #181970) were disaggregated into single-cell suspensions. Radiographic contact with the V-SVZ was identified by MRI imaging and confirmed by a trained neurosurgeon. Multi-dimensional single- cell mass cytometry (CyTOF) then measured >30 immune identity markers in thirteen immune populations infiltrating human glioblast- omas, including CD4 and CD8 T cells,γδT cells, natural killer cells, B cells, microglia, peripheral macrophages, and myeloid-derived sup- pressors cells (MDSC). Advanced computational dimensionality- reduction tools including Citrus, t-SNE, FlowSOM, and MEM along with traditional biaxial gating strategies identified key differences in the abundance and phenotypes of immune infiltrates.

Results

On the basis of tumor contact with the V-SVZ, the Citrus and Flow- SOM clustering algorithms computationally identified consequential distinctions in the abundance of five T cell, macrophage, and micro- glia subsets among glioblastomas. In addition, differential expression of five functional immune markers (three activating receptors and two inhibitory receptors) was observed in seven distinct immune cell subsets infiltrating tumors. Critically, the abundance of identified im- mune subsets and relative immune receptor expression levels corre- lated significantly with tumor contact status and patient outcomes. Biaxial gating analysis and parallel computational pipelines confirmed that comparable cell subsets could be identified with traditional ap- proaches and unsupervised algorithmic analysis.

Conclusions

Single-cell mass cytometry in conjunction with the Citrus and FlowSOM clustering tools identified key differences in immune cell abundance between V-SVZ contacting and non-contacting glioblastomas. These results provide key insights into the immune microenvironment of glioblastomas and elucidate several clinically actionable immunotherapeutic targets that may be used to optimize treatment strategies for glioblastomas based on V-SVZ contact status.

Ethics Approval

This study was approved by the Institutional Review Board at Vander- bilt University, approval #131870, #030372, and #181970.

P728

Inference of immunotherapy response-predictive biomarkers in lung adenocarcinoma from hematoxylin and eosin (H&E) stained images

Kimary Kulig, PhD, MPH1, Cory Batenchuk, PhD1, Peter Cimermancic1, Huang-Wei Chang, PhD1, Eunhee Yi, MD2, Vanessa Velez1, Hardik Patel1, Ali Behrooz1, Kamilla Tekiela1, Robert Findlater1

1Verily Life Sciences, South San Francisco, CA, United States;2Mayo Clinic, Rochester, MN, United States

Correspondence:Kimary Kulig ([email protected])

Journal for ImmunoTherapy of Cancer2019,7(Suppl 1):P728

Background

The current work-up for diagnosis, gene mutation, and programmed death ligand-1 (PD-L1) testing in metastatic NSCLC can exhaust an entire tumor specimen. Testing for tumor mutation burden (TMB), a candidate predictive biomarker for checkpoint inhibitor therapy, cur- rently requires 10 tissue slides and ranges from 10 days to 3 weeks from sample acquisition to test result. As more CDx-restricted drugs are developed for lung adenocarcinoma, rapid, tissue-sparing tests are sorely needed. We investigated whether these biomarkers can be inferred from digital H&E images alone.

Methods

Whole slide (40x-magnified) H&E images of lung adenocarcinoma from TCGA were used to train a computational model to infer TMB and PD-L1 status. Modeling occurred in two layers: first, a neural network was trained to segment known histologic and other tumor microenvironment features from H&E images. Each feature was manually annotated and confirmed by two rounds of pathologist review. Second, these feature maps, along with ground truth biomarker labels and patient demographics, were analyzed by second-level models to infer biomarker status. Ground truth biomarker status of H&E-associated tumor samples included WES for TMB and reverse-phase protein array for PD-L1, per TCGA. TCGA samples (n=40) generating 40,000 feature anno- tations were used to train and evaluate the first layer model. Held-out TCGA samples were used to train (n=200) and test (n= 194) the TMB and PD-L1 classification models.

Results

In a "hold-one-TCGA-site-out" approach using 40 images, the overall accuracy of the first model layer in classifying pathologist-confirmed histologic features was 98-99%. The two-layer model achieved an ac- curacy of 81% in predicting both TMB (AUC = 0.81) and PD-L1 status (AUC = 0.86) in the testing dataset.

Conclusions

Biomarker inference from digital H&E images has the potential to be highly accurate and to supplement or replace certain tissue-based tests. Use of known biological features makes the model interpret- able and verifiable by pathologists. This method may enable testing in patients whose tumor tissue has been exhausted or who cannot undergo re-biopsy to enable biomarker testing. This technology can return test results within hours instead of days or weeks, enabling rapid treatment decision-making and start. Ongoing work to improve accuracy of the model, assess model correlation to PD-L1 by IHC and gene panel-derived TMB, and to evaluate the models correlation to clinical outcomes is in progress.

P730

An empirical framework for validating artificial intelligence-derived PD-L1 positivity predictions using samples from patients with urothelial carcinoma

Andrew Beck, MD, PhD1, Benjamin Glass1, Hunter Elliott1, Jennifer Kerner1, Aditya Khosla1, Abhik Lahiri1, Harsha Pokkalla1, Dayong Wang1, Ilan Wapinski1, George Lee, PhD2, Vipul Baxi, MS2, Cyrus Hedvat, MD, PhD2, Dimple Pandya2, Michael Montalto2

1PathAI, Inc, Boston, MA, United States;2Bristol-Myers Squibb, Princeton, NJ, United States

Correspondence:Ilan Wapinski([email protected])

Journal for ImmunoTherapy of Cancer2019,7(Suppl 1):P730

Background

Assessing PD-L1 immunohistochemistry (IHC) expression may play a role in identifying patients likely to benefit from anti–PD-1/PD-L1 therapies in some advanced cancers, including urothelial carcinoma (UC). Studies have shown variable inter-observer agreement for path- ologist assessment of PD-L1 expression, with generally higher levels of concordance for tumor cell scoring compared with immune cell scoring [1,2]. Thus, immune cell PD-L1 scoring may be more challen- ging to implement than scoring on tumor cells. We evaluated the performance of an artificial intelligence (AI)-based predictor of PD-L1 expression on tumor and/or immune cells in the tumor microenvir- onment and compared this with manual evaluations by a network of pathologists to identify whether the AI platform performed more consistently.

Methods

PD-L1 was assessed using the PD-L1 IHC 28-8 pharmDx assay (Dako, Agilent Technologies Co). The training set for this exploratory ana- lysis consisted of 293 pretreatment samples from patients with ad- vanced UC, commercially procured or from clinical trials of nivolumab [3,4]. From these, we obtained 105,514 annotations of tumor and immune cells from 43 pathologists. To establish a refer- ence dataset for manual vs digital concordance using the platform (PathAI, not intended for diagnostic use), we generated 80 150x150- micron–sized“frames”sampled from 100 held-out clinical images, re- moving frames of inadequate tissue quality or with presence of arti- facts. For each frame, we collected exhaustive annotations from 5 pathologists to produce quantitative estimates of PD-L1 positivity on all tumor and immune cells. Altogether, 66,049 annotations were col- lected and used to compute pathologist consensus scores for each frame. These scores were then correlated with each individual path- ologist (inter-reader agreement) and with the PathAI-derived auto- mated scores for evaluation of manual vs digital agreement.

Results

The PathAI platform showed significantly stronger correlation with pathologist consensus scores compared with scores generated by in- dividual pathologists for quantifying PD-L1 positivity of lymphocytes (r-squared=0.744 vs 0.598) and macrophages (r-squared=0.68 vs 0.287). There was no difference in correlation with consensus be- tween PathAI-derived and individual pathologist-derived assessment of positivity on tumor cells (r-squared=0.837 vs 0.857).

Conclusions

We validated performance of the PathAI platform for automated as- sessment of PD-L1 expression on tumor and immune cells and dem- onstrated that the AI-predictors perform similar to or better than

pathologist-based scoring in all cell types tested, and especially for immune cells where manual correlation is low. These results suggest that AI-powered assessment represents a reproducible and poten- tially generalizable approach to interpretation of IHC assays.

Acknowledgements

Bristol-Myers Squibb and PathAI, Inc. Trial Registration

NCT02387996, NCT01928394

References

1. Rimm DL, Han G, Taube JM, et al. A prospective, multi-institutional, pathologist-based assessment of 4 immunohistochemistry assays for PD- L1 expression in non-small cell lung cancer. JAMA Oncol. 2017;3:1051- 1058.

2. Eckstein M, Wirtz RM, Pfannstil C, et al. A multicenter round robin test of PD-L1 expression assessment in urothelial bladder cancer by immunohis- tochemistry and RT-qPCR with emphasis on prognosis prediction after radical cystectomy. Oncotarget. 2018;9:15001-15014.

3. Sharma P, Retz M, Siefker-Radtke A, et al. Nivolumab in metastatic urothe- lial carcinoma after platinum therapy (CheckMate 275): a multicentre, single-arm, phase 2 trial. Lancet Oncol. 2017;18:312-322.

4. Sharma P, Callahan MK, Bono P, et al. Nivolumab monotherapy in recurrent metastatic urothelial carcinoma (CheckMate 032): a multicentre, open-label, two-stage, multi-arm, phase 1/2 trial. Lancet Oncol. 2016;17:1590-1598.

Ethics Approval

The protocol was approved by site institutional review boards or independent ethics committees and conducted according to Good Clinical Practice guidelines, per the International Conference on Harmonisation. Patients provided written informed consent based on Declaration of Helsinki principles.

P731

Identification and validation of shared neoantigens for cancer immunotherapy

Jennifer Busby, PhD, Melissa Rotunno, Tyler Murphy, William Brinton, Michael Zhong, Lina Kim, Abu Jalloh, Amanda Costa, Michael Fray, Aaron Yang, Meghan Hart, Matthew Davis, Rita Zhou, Elizabeth Maloney, Alexis Mantila, Karin Jooss, PhD, Aleksandra Nowicka, Christine Palmer, PhD, James Sun, Roman Yelensky, PhD , Jennifer Busby, PhD

Gritstone Onclogy, Inc, Cambridge, MA, United States Correspondence:Roman Yelensky ([email protected])

Journal for ImmunoTherapy of Cancer2019,7(Suppl 1):P731

Background

Neoantigens can be targeted with cancer immunotherapies, includ- ing cell therapy and immunization, but most neoantigens are private to each patient, necessitating fully personalized treatments. Identifi- cation of solid tumor-presented neoantigens that may be shared across patients holds great value for the generation of off-the-shelf therapies.

Methods

We pursued an immunopeptidomics strategy to identify which recur- rent mutations in cancer lead to neoantigen peptides presented on the surface of tumor cells. 498 frozen resected tumor samples and 6 cell lines across 14 tumor types (including 139 NSCLC and 139 CRC) were subjected to transcriptome RNASeq. Transcriptomes were ana- lyzed for expression of recurrent driver mutations and used to pre- dict HLA peptide epitopes with our deep learning model EDGE [1]. Additionally, single HLA class I allele expressing K562 cells were transduced with vectors expressing shared mutations of interest. Pre- dicted peptides were assayed in the tumors and cell lines by Class I HLA immunoprecipitation and targeted mass spec. For a subset of peptides, we evaluated if CD8 T-cell precursors are found in the naïve repertoire of HLA-matched donors.

Results

16 neoantigens (distinct HLA/mutation pairs) derived from 10 driver mutations were detected by MS in 27 tumors and 3 single allele cell lines. Only 1 of these neoantigens (KRAS G12V/HLA-A*11:01) was

previously reported in patients. The driver mutation with most neoantigens was KRAS G12V, with peptides detected on 4 HLA al- leles, accounting for ~40% of patients with the mutation. Import- antly, HLA-A*02:01 was not found to present KRAS G12V. Other neoantigen peptides were detected in genes including KRAS, CTNNB1 and TP53. When combined with the few published neoanti- gens, the set was estimated to cover 9% of NSCLC, 11% of MSS-CRC, and 19% of pancreatic cancer patients, accounting for HLA restric- tion. Precursor CD8 T cells were detected for 6 out of 7 tested HLA/ peptide pairs arising from mutations in KRAS and CTNNB1, recog- nized by a median of 46 TCR clonotypes, highlighting their immuno- genic potential.

Conclusions

We analyzed immunopeptidomes of human tumors and cell lines and validated 15 novel shared neoantigens, the largest set to-date. The study enables off-the-shelf immunotherapies for significant frac- tions of patients and highlights the value of HLA screening, as muta- tions are presented by specific HLA alleles. Shared neoantigen targets identified in this study, among others, are currently being tested in patients in the SLATE clinical immunotherapy program. Pa- tient coverage will likely expand as additional classes of neoantigens are identified and added to off-the-shelf immunotherapy products.

References

1. Bulik-Sullivan B, Busby J, et al. Deep learning using tumor HLA peptide mass spectrometry datasets improves neoantigen identification. Nature Biotech. 2018; 37: 55-63. (DOI: 10.1038/nbt.4313)

P732

Reverse causal inferencing on lung adenocarcinoma patients reveals a stem cell-like molecular subtype associated with pack year history

Renee Deehan, Sergey Korkhov, Alexis Foroozan, Scott Marshall, PhD, Renee Deehan, Nimisha Schneider

QuartzBio, a Precision for Medicine company, Plan-les-Ouates, Switzerland

Correspondence:Renee Deehan (renee.deehan- [email protected])

Journal for ImmunoTherapy of Cancer2019,7(Suppl 1):P732

Background

Advances in high throughput measurement technologies (-omics data) have made it possible, and increasingly affordable, to generate high complexity, high volume data for oncology research. Coordinate efforts in computational modeling and machine learning applications to biological data have yielded increasingly sophisticated methods to deeply characterize the mechanisms of disease pathogenesis and heterogeneity among patients; which are predicates to the rapid de- velopment and timely and safe administration of efficacious treat- ments. Open source repositories that catalog, harmonize and host -omics data collected from clinical or preclinical studies, generously donated by patients and researchers, provide revolutionary access to otherwise siloed data. The combination of these advancements en- abled us to characterize the molecular phenotypic heterogeneity that exists within a lung adenocarcinoma cohort from The Cancer Gen- ome Atlas [1].

Methods

We applied a reverse inferencing approach that systematically inter- rogates RNAseq measurements from tumor and control biopsies against our library of cause and effect gene networks curated from published experiments. If patterns observed in the data are signifi- cantly similar to those in a network, an inference about the direc- tional activity of that network can be made. Our library was nucleated through an open sourced knowledge graph [2] and en- hanced with updated and relevant knowledge using the Biological Expression Language framework [3].

Results

In LUAD tumor cells, we detected a pattern of gene signatures which indicated a tumor stem cell-like phenotype characterized by pre- dicted decreases in the activity of pro-differentiation factors FOXP2

and PHOX2B and an increased response to hypoxia. Analysis of pa- tients with heavy (>40) versus light (<10) pack-year burden sug- gested an augmented dedifferentiation profile in heavy smokers, including decreased SOX6, HNF1A and increased FAT1 signaling, which has also been recently implicated in resistance to immune checkpoint inhibitors in NSCLC patients [4]. Expression of PD-L1 has been associated with cumulative inhaled smoke exposure in lung cancer patients, and with mechanisms inferred from our analysis (e.g., increased MIR140 signaling) [5,6].

Conclusions

Our in silico analysis of lung cancer patient biopsies generated hy- potheses implicating stem cell signaling in tumors, and a further stratification of this signal based on patient pack year burden. Given