1285P - Deep learning prediction of response to advanced-generation ALK inhibitors (aALKi) in ALK-rearranged advanced NSCLC (aNSCLC) patients using digital pathology

Background

About 4% of NSCLC patients harbor a rearrangement involving the ALK gene (ALK+), producing a novel chimeric kinase that drives cancer evolvement and progression. aALKi, i.e. 2nd and 3rd generation ALKi are currently the most active agents for the treatment of such patients, with a response rate around 80%. However, a subgroup of these patients do not respond to this treatment or have a short-term response followed by rapid progression. We aimed to use deep learning on the diagnostic H&E stained tumor sample images in order to predict response to aALKi.

Methods

Clinical data and digital pathology slides were collected from Sheba Medical Center and Mayo Clinic of NSCLC ALK+ patients who were treated with 1st line aALKi. Patients were categorized as rapid progressors (RP) if progression of disease (PD) occurred within a year of treatment initiation or as responders (R) if progression-free survival (PFS) was above one year. H&E slides from the diagnostic specimen were scanned at each center separately. The deep learning process was composed of extracting patches from the whole slide images (WSI), feature extraction using simple framework for contrastive learning of visual representations (simCLR) with ResNet18 as the Convolutional Neural Network backbone. We used a multiple instance learning approach for classification. Train and test ratio were 0.7:0.3, with random separation of the patients to the train and test cohorts.

Results

The study included 55 patients, all alectinib treated as 1^st line for aNSCLC. Median follow-up time was 38 months (IQR 22,55) with a median PFS on aALKi of 29 months (8, NR). Thirty-three (60%) were female. Median age at diagnosis was 57 (IQR 50-68). Forty-two patients were never smokers (76%), 12 (22%) were former smokers and one (2%) current smoker. TP53 mutation and ALK fusion variant 3 were seen in 12 (22%) and 9 (16%) of patients respectively. Fifteen (27%; 95% CI 16-41%) patients were RP. Our model achieved an AUC of 0.73 with an accuracy of 0.72 in assignment of patients to RP/R groups.

Conclusions

Our study shows the potential of using deep learning on H&E diagnosic slide images to predict RP/R status of ALK+ patients on aALKi.

Clinical trial identification

Editorial acknowledgement

Legal entity responsible for the study

The authors.

Funding

Has not received any funding.

Disclosure

K. Parikh: Financial Interests, Institutional, Advisory Board: AstraZeneca; Financial Interests, Personal, Invited Speaker: Clinical Care Options; Financial Interests, Institutional, Research Funding: Verastem; Financial Interests, Institutional, Speaker, Consultant, Advisor: OncLive, DAVA Oncology, Takeda; Other, Institutional, Writing Engagement, Research: Takeda. Y. Lo: Financial Interests, Personal, Advisory Board: Diaceutics, Sanofi, Takeda, ChromaCode; Other, Institutional, Research Funding, supplies: Biocartis; Financial Interests, Personal, Speaker, Consultant, Advisor: PRIME Education, CURE, Educated Patient. I. Vardi: Non-Financial Interests, Personal, Research Grant: Roche. D. Urban: Financial Interests, Personal, Speaker, Consultant, Advisor: Merck Sharp and Dohme Israel, J-C Healthcare, AstraZeneca, Roche Israel, BMS, Takeda Israel, Medison Pharma. H. Gantz Sorotsky: Financial Interests, Personal, Speaker, Consultant, Advisor: Roche-Genentech, MSD, AstraZeneca, Takeda, Astellas, BMS, Medison, Janssen, Pfizer, Merck, Novartis. A. Onn: Financial Interests, Personal, Speaker, Consultant, Advisor: MSD, AstraZeneca, Pfizer, Amgen, BMS; Financial Interests, Personal, Advisory Board: Medisin. G. Kalmanovitz: Financial Interests, Personal, Speaker, Consultant, Advisor: Merck. J. Bar: Financial Interests, Personal, Advisory Board: MSD, Takeda, Roche, Pfizer, AbbVie, Bayer, AstraZeneca, Merck-Serono, J-C healthcare, Medison Pharma, Amgen; Financial Interests, Personal, Invited Speaker: BMS; Financial Interests, Personal, Other, Steering committee: AbbVie, Merck Healthcare KGaA, Roche, AstraZeneca; Financial Interests, Institutional, Local PI: MSD, Roche, Takeda, AbbVie; Financial Interests, Institutional, Other, local sub-investigator: Novartis; Financial Interests, Institutional, Research Grant: OncoHost, Immunai, AstraZeneca; Non-Financial Interests, Other, Committee member: IASLC. All other authors have declared no conflicts of interest.