Beyond PD-L1: novel PD-1 biomarkers identified by driving T cell dysfunction in vitroSimarjot Pabla, Tenzing Khendu, Benjamin Duckless, Andrew Basinski, Cailin E. Joyce, Thomas Horn, Lukasz Swiech, David Savitsky, Jennifer S. Buell, Dhan Chand1Agenus Inc. or subsidiary thereof (current or former employee), Lexington, MA

Agenus Inc. or subsidiary thereof (current orformer employee), Lexington, MA 02421, USA.

Correspondence:Simarjot Pablasimarjot.pabla@agenusbio.com

Dhan Chanddhan.chand@agenusbio.comPresented at Society for Immunotherapy of Cancer (SITC), November 10 - 15, 2020, Virtual Meeting

References:• Riaz et. AL., Tumor and Microenvironment Evolution during

Immunotherapy with Nivolumab.• Hugo et. al., Genomic and Transcriptomic Features of Response to

Anti-PD-1 Therapy in Metastatic Melanoma. Cell. 2016 Mar 24;165(1):35-44.

• Morrison, C., Pabla, S., Conroy, J.M. et al. Predicting response to checkpoint inhibitors in melanoma beyond PD-L1 and mutational burden. j. immunotherapy cancer 6, 32 (2018).

Figure 1: Representation of VISION platform as an integrated tool to recapitulate tumor immune interactions allowing perturbations to interrogateconcepts and accelerate them to clinic.

Integrated discovery & development platform drives concepts to clinicAgenus VISION Platform

Conclusions

• Agenus’ VISION platform combines deep in vitro profiling and AI-based approaches to predict clinicaloutcomes, plus rational targets & combinations.

• We defined a predictive biomarker signature that outperforms standard PD-L1 IHC.• We identified a potential mechanism underlying the effective combination of anti-PD1 and anti-TIGITantibodies in the clinic.

VISION T cell dysfunction signature correlates with Objective Response Rate in anti-PD1/PDL1 treatment across indications

Figure 5: Correlation of Objective Response Rate of anti-PD1/PDL1 treatment in human tumors to A) Tumor Mutation Burden, B) Fraction of PD1 high patients (fPD1), C) CD8+ T-Cell Abundance, D) VISION T cell dysfunction signature, and E) Bivariate Tumor Mutation Burden and VISION T cell dysfunction signature.

F) Prediction of ORR for TCGA molecular subtypes using the bivariate model. TCGA indications were split into previously defined molecular subtypes (e.g. for BRCA - HER2, LumA, LumB, and basal) to validate correlations seen at the indication-level.

VISION deep-learning model identifies anti-PD1 responders in melanoma and cervical cancer

Figure 4: Development of a deep learning model to differentiate responders from non-responders in a population of anti-PD1 treated melanomapatients. Melanoma was selected for initial training and validation due to the availability of public, clinically-annotated tumor RNA-seq datasets fromlarge numbers of patients (Hugo et al. (n=25) and Riaz et al. (n=43) ). Feature selection comprised of Agenus VISION T cell state signaturescombined with differentially expressed genes between the two groups. Data partition for training and validation was followed by modelling andperformance evaluation.

Figure 3: (Left Panel) Average response rate to anti-PD1 therapies. (Right Panel) Accuracy of predicting response to anti-PD1 on a per patient basisusing either PD-L1 IHC or Agenus’ VISION machine learning models using melanoma (Riaz et. Al and Hugo et. al) and cervical (Agenus C-550 andC-700) patients treated with anti-PD1.

….

Input Layer Hidden Layer 1Hidden Layer 2

Hidden Layer 3

Output Layer

Responder

Non-Responder

68 Melanoma Patients

(Treated w/ anti-PD1)

Training Set

60/40 Partition

Test Set/VISION

Feature SelectionVISION

T cell Dysfunction Signature

Multi Layer Perceptron (MLP) Model

VISION PLATFORM DEFINES RATIONAL COMBINATIONSSingle cell transcriptomics of T cells during early dysfunction identifies T cell subsets that respond to anti-PD1

Figure 6: A) Heatmap showing mean expression of select activation and dysfunction markers based on single cell RNA-sequencing across 14 T cellclusters that are present during the early dysfunction state. (B) Violin plot depicting single cell expression levels for select dysfunction markers withinthe activated PD1high subset.

Combination of PD-1 and TIGIT blockade enhanced T cell cytotoxicity of tumor cells relative to monotherapies

Figure 7: A) Tumor killing capacity of anti-PD1 and co-blockade B) Tumor killing capacity of PD1 and anti-TIM3 co-blockade.

Hours

% k

illin

g

Isotypeanti-PD1anti-TIGITCo-blockade

Enhanced cytotoxicity with a PD1-TIGIT co-blockade

Isotypeanti-PD1anti-TIM3Co-blockade

Hours

Donor 1 Donor 2 Donor 3 Donor 4 Donor 8

No benefit from PD1-TIM3 co-blockade over PD1 mono-blockade

% k

illin

g

T cell dysfunction system drives CD8+ T cells to dysfunction through chronic cancer antigen exposure

Figure 2: Method for driving T cell dysfunction in vitro. T cell preparation: PBMCs were isolated from three healthy donors on a Ficoll gradient andrested overnight. CD8+ T cells were isolated on magnetic beads, stimulated with CD3/CD28, transduced with NY-ESO-1 lentivirus, and expanded.TCR expression was confirmed by flow cytometry. Cancer cell preparation: U251 MG cell lines were transduced with lentivirus encoding a fusionprotein of b2-microgobulin, HLA2-A2 and NY-ESO-1 or MART-1 peptide and selected on blasticidin. Antigen expression was confirmed with an NFATreporter system in Jurkat T cells. For co-culture: A fixed number of irradiated U251 MG cells were cultured for 24h. CD8+ T cells were added toachieve the desired cancer:T cell ratio. The co-culture was monitored every 24h. If cancer cells were depleted, CD8+ T cells were transferred to afresh well of irradiated U251 MG cells and the remainder were collected for downstream analyses. If cancer cells were not depleted, media waschanged and reassessed the following day. Top Panel) Cytotoxicity kinetics as monitored by live cell fluorescence microscopy. Middle Panel)Secreted effector molecules measured by Luminex Bead Array, normalized to the maximum signal per molecule. Bottom Panel) Transcriptomiclandscape of T cell dysfunction across antigen exposures.

D E F

A B C

VISION T cell dysfunction signature correlate with Objective Response Rate in anti-PD1/PDL1 treatment

60% 87%PDL1 IHCAccuracySolid Tumors

Agenus ModelMelanomaAccuracy

Anti-PD1 Therapy Landscape (Cervical Cancer)

13 %Responders

Non-Responders 87 %

Identifying Responders and non-responders

86%Agenus ModelCervicalAccuracy

VISION PLATFORM PROVIDES CLINICAL INSIGHTS

A B

Cytotoxicity:

Cytokine secretion:

Naïve

Memory TCR signaling Co-activators

Early co-inhibitors

Late co-inhibitors TME conditioners

Molecular markers:

IFN-beta

IL2,3,4,10TNF-alpha

IL5,13

Perforin

IFN-gamma, GM-CSF, sCD40L

FGF2, sCD137, sFasL, TGF-alpha

Activation Early Dysfunction (PD-1 response

Interferon stimulated genes

Naïve window)

Effector

Late DysfunctionCombinations for durable PD-1 responses

ANTIGEN EXPOSURES

PRIME

ACTIVATE

KILL

EXHAUST

SUPPRESS

REGULATE

DIFFERENTIATE

REGULATE

• Gene editing libraries• Small molecule libraries• Cytokines• Tumor metabolites• Hypoxia inducers• Other perturbations

APC T cell

MDSC

Myeloid cell

CARs TCRs Abs

Tumor

iNKT cell

KILL

EXHAUST

KILL

EXHAUST

NK cell

A B

70Poster #

Expression of dysfunction markers on the activated PD1high subset

Low High

Normalized Expression

4-1BB

TNFRSF18

OX40

GZMB

TIGIT

CD244

CD38

TIM3

PD1

ENTPD1

CD101

CTLA4

LAG3

Dysf

unct

ion

mar

kers

Activ

atio

nm

arke

rs

Activated PD1high subset

Inhibited PD1high subset

1.5-fold increasedwith anti-PD1 Tx

6.5-fold decreased with anti-PD1 Tx