Download pdf - Algorithmic Prediction of Response to Checkpoint Inhibitors€¦ · Roswell Park Cancer Institute with complete treatment and clinical outcome data for 87. For algorithmic prediction,

Testing for Primary and Secondary Immune Biomarkers

Gandhi S2#, Pabla S1, Nesline M1, Pandey M2, Ernstoff M2, Dy GK2, Conroy JM1,2, Glenn ST1,2, Burgher B1, Qin MC1, Andreas J1, Giamo V1, Papanicolau-Sengos A1, Galluzzi L3,4, He J1,Gardner M1, Morrison CD1,2*

OmniSeq,Inc1, Buffalo, NY; Roswell Park Cancer Institute2, Buffalo, NY; Weill Cornell Medical College, New York, NY3; Université Paris Descartes/Paris V, Paris, France4

Algorithmic Prediction of Response to Checkpoint Inhibitors

PD-L1 immunohistochemistry (PD-L1 IHC), mutational burden (MuB) andtumor infiltrating lymphocytes (TILS) are the current standard or “primary”immune biomarkers used for selection of immunotherapy today. Thesemarkers however, are suboptimal for response prediction to currentlyapproved checkpoint inhibitors (CPIs) across a broad range of tumor types.Machine learning (ML) decision models have great potential for predictingresponse, however, they risk overfitting data, only performing well in theinitial training cohort and lacking generalizability when extended to othercohorts. Alternatively, biology-based decision models can initiallyunderperform due to limited data and a simplified rule set, but oftenperform equally well when extended to larger similar patient cohorts. Wetook a combinatorial approach to algorithmic prediction to address theseshortcomings.

Background

We evaluated response prediction to currently approved checkpointinhibitors using both standard primary immune biomarkers and algorithmicapproaches. We developed a reference population of 167 patients fromRoswell Park Cancer Institute with complete treatment and clinical outcomedata for 87. For algorithmic prediction, a ML polynomial regression modelbased on 54 immune-related genes combined with mutational burden wasoptimized for prediction of response. A biological 4-gene decision treemodel was constructed independently based on ML. A second biologicaldecision tree incorporated the weighted average relative rank of theexpression of multiple genes in 4 different immune functions includingimmune cell infiltration, regulation, activation, and cytokine signaling.Bayesian model average (BMA) incorporated all three models’ results into afinal prediction.

Methods

Immune Response NGS WorkflowFormalin-fixed paraffin embedded (FFPE) cancer samples were evaluated byRNA-Seq with the OncomineTM Immune Response Research Assay and DNA-seq with AmpliSeq capture of 409 cancer related genes with ComprehensiveCancer PanelTM using the Ion ChefTM and S5XLTM. RNA-Seq analysis wasperformed with the Torrent SuiteTM v5.2.0, followed by data normalization.

Figure 3: Primary Immune Biomarkers. PD-L1 IHC (Top left), CD8 IHC (Topright) are commonly used biological measurements to predict response toimmune checkpoint blockade in a variety of malignancies. An immuneresponse gene expression (GEX) panel was used to measure normalized readsper million (nRPM) for 54 validated genes. nRPM values were then rankedfrom 0 to 100 based on a reference population of 167 patients to derive GEXinterpretation of High and Low. These interpretations were then used tovisualize the immune GEX landscape of inflamed vs non-inflamed tumors.DNA Seq was used to estimate mutational burden (MuB), defined as numberof nonsynonymous somatic mutations per million exonic bases.

Prediction of response to checkpoint

inhibitors can be improved by combining

primary and secondary immune related

biomarkers. Our initial study of 87

patients is a framework for expanding to

a larger multi-institutional collaboration

to prove that combining primary and

secondary immune related biomarkers

for prediction of response to checkpoint

inhibitors has value. The same approach

can be used for assigning patients to

combination immunotherapies.

Figure 4. A subset of 87 samples in this reference population of 167 tumorshad RECISTv1.1 data. Response was defined as clinical benefit which wascomplete response, partial response, or stable disease with at least 1 yearsurvival.

• The majority of responders were MuB low and negative for PD-L1• Some tumors with a high MuB or PD-L1 positive were nonresponders

Prediction of Response Using Primary Immune Biomarkers

Secondary Immune Biomarkers for Nonresponders

52; 91%

3; 5%2; 4%

20; 67%

6; 20%

4; 13%

MuB Low PD-L1 Low

MuB High

PD-L1 High

Inflamed tumors with MuB and PD-L1 in a reference population of 167 tumors• Tumors with high MuB are as likely to be non-inflamed as inflamed• Tumors with high PD-L1 are often non-inflamed

RespondersNonresponders

Primary Immune Biomarkers

Figure 1: Treatment Regimen by Tumor Type for Patients (n=87) Treated byCheckpoint Inhibitors at Roswell Park Cancer Institute with completeresponse data by RECIST v1.1

Inflamed Tumors Non-Inflamed Tumors

10

MuB Not High

64

MuB High MuB High

MuB Not High

78

15

Inflamed Tumors Non-Inflamed Tumors

14

PD-L1 Not High

60

PD-L1 High PD-L1 High

PD-L1 Not High

85

8

RNA-seq

DNA-seqIHC

FISH MSI

Gene expression of >350 immune related genes for tumor infiltrating lymphocytes and t-cell

receptor signaling

Full exon sequencing >400 genes for mutational burden

Microsatellite instability (MSI) by PCR for DNA mismatch repair deficiency

Presence of PD-L1/2 gene cluster copy number gains

CD3, CD8, FDA-approved assay

for PD-L1 (22C3)

Primary Immune Markers expand eligible patients for checkpoint inhibitorsSecondary Immune Markers inform clinical trial selection for non-responders

2017 ASCO Annual MeetingCategory: Tumor BiologySub-category: Molecular Diagnostics and ImagingAbstract No: 11565Poster Board Number: Poster Session (Board #265)Citation: J Clin Oncol 35, 2017 (suppl; abstr 11565)Contact: [email protected], [email protected],[email protected]

PD-L1 IHC TPS=0 PD-L1 IHC TPS=50 CD8 IHC Non-inflamed CD8 IHC Inflamed

Figure 2: Immune Response workflow

Algorithmic Prediction of Response

Figure 6: Algorithmic Response Prediction. Using a Bayesian Model Averaging algorithm, gene expression andmutational burden data from three independent prediction models result in a high negative predictive value(NPV) for non-responders and a high positive predictive values (PPV) for responders in a 87 sample cohort ofpatients treated with checkpoint inhibitors

Figure 5: Secondary Immune Biomarkers. Understanding T-cell receptor signaling in individual genes implicatedin the adaptive immune response cycle can identify why patients are not responsive to monotherapy.

Conclusions: