Abstract 1071P
Background
Predicting immunotherapy (IO) efficacy still remains a crucial unmet need for NSCLC. Compared with PD-L1, no other biomarker has demonstrated superiority. Due to the complexity of the immune system, it is unlikely that a single biomarker will have a high performance. Due to the high dimensional data in oncology, new methodologies such as artificial intelligence (AI) are necessary for integration aiming at looking at new software biomarkers.
Methods
Among 480 advanced NSCLC patients (pts) treated with IO at Istituto Nazionale Tumori of Milan, 113 pts who performed cell-free DNA sequencing with Guardant360CDx/Archer LiquidPlex ctDNA assay panels were collected. Clinical and radiological real-world (RW) data, for a total of 29 features, were considered to develop AI models (RW-M). The matched gene features among the two panels (20 genes) were used in input to create a combined model (RWG-M). Different machine learning (ML) techniques such as logistic regression, random forest, saptor vector machine and Catbust were used. Disease control rate and (DCR), overall response rate (ORR) and 6-month overall survival (6OS) outcomes were considered for classification. SHAP Explainable AI (XAI) method was applied to explain the "black box".
Results
Catbust was the most robust ML model. The RW-M test accuracy (ACC) for DCR, ORR and 6OS were 0.57, 0.65 and 0.65, respectively. The RWG-M reported ACC values of 0.61, 0.74 and 0.74, for DCR, ORR and 6OS, respectively. Among RW features, XAI SHAP identified high LDH, NLR, age, ECOG and stage which negatively correlate with outcome and high pack/year, (body mass index) BMI, PD-L1 positively correlated with DCR, ORR and OS. Among the G features, SHAP detected TP53 as the most relevant gene negatively correlated with outcome. KIT gene was identified as the second most important gene negatively correlated with the outcomes. BRAF was found to correlate with worse response.
Conclusions
Despite the low amount of data, this work showed that G features enrich the RW model in terms of prediction accuracy. The explainable AI algorithms suggested a negative predictive role of TP53 and KIT mutational status. More data are needed to improve the goodness of the algorithms.
Clinical trial identification
Editorial acknowledgement
Legal entity responsible for the study
The authors.
Funding
Has not received any funding.
Disclosure
A. Prelaj: Financial Interests, Personal, Other: AstraZeneca, Travel Roche, Italpharma; Financial Interests, Personal, Advisory Board: BMS. G. Lo Russo: Financial Interests, Personal, Other: BMS, MSD, Italfarmaco, Novartis, Roche, AstraZeneca, Sanofi, Pfizer. R. Ferrara: Financial Interests, Personal, Advisory Board: Celgene, MSD. C. Proto: Financial Interests, Personal, Advisory Board: AstraZeneca, Roche, MSD; Financial Interests, Personal, Invited Speaker: AstraZeneca, Roche, MSD, Bristol Myers Squibb; Financial Interests, Personal, Principal Investigator: Janssen, Pfizer, Lilly, Spectrum Pharmaceuticals, Roche, MSD, BMS, AstraZeneca. M.C. Garassino: Financial Interests, Personal, Other: AstraZeneca, MSD International GmbH, BMS, Boehringer Ingelheim Italia S.p.A., Celgene, Eli Lilly, Ignyta, Incyte, Inivata, MedImmune, Novartis, Pfizer, Roche, Takeda. F.G.M. De Braud: Financial Interests, Personal, Advisory Board: Ignyta, BMS, Daiichi Sankyo, Pfizer, Octimet Oncology, Incyte, Teofarma, Pierre Fabre, Roche, EMD Serono, Sanofi, NMS Nerviano Medical Science, Pharm Research Associated (U.K) Ltd; Financial Interests, Personal, Invited Speaker: BMS, Roche, MSD, Ignyta, Bayer, ACCMED, Dephaforum S.r.l., Nadirex, Merck, Biotechspert Ltd, PriME Oncology, Pfizer, Servier, Celgene, Tesaro, Loxo Oncology Inc., Sanofi, Healthcare Research and Pharmacoepidemiology; Financial Interests, Personal, Principal Investigator: Novartis, Roche, BMS, Celgene, Incyte, NMS, Merck KGAA, Kymab, Pfizer, Tesaro, MSD. All other authors have declared no conflicts of interest.