Abstract Background: Immune checkpoint inhibitors (ICIs) have significantly improved outcomes for patients with metastatic non-small cell lung cancer (NSCLC). However, clinical management is challenging. The PD-L1 biomarker does not optimally guide the choice between ICI monotherapy and combination ICI-chemotherapy, and there are currently no biomarkers to predict immune-related adverse events (irAEs). Here, we describe a versatile platform based on plasma proteomics and machine learning for generating predictive tools to support treatment decisions. Methods: Pre-treatment plasma samples were collected from 616 NSCLC patients receiving ICI-based therapy (NCT04056247). Clinical benefit (CB) data and irAEs were recorded. Proteomic profiling of plasma samples was performed using the SomaScan® assay, covering ~7000 proteins per sample. Proteins associated with Clinical Benefit (progression-free survival at 12 months) were classified as resistance-associated proteins (RAPs). Proteins associated with grade ≥3 irAEs occurring within the first 100 days of treatment were classified as toxicity-associated proteins (TAPs). RAPs and TAPs were integrated into two separate machine learning-based models, yielding CB and irAE probabilities, respectively. The CB model used a CB probability threshold to assign patients a PROphet-positive or PROphet-negative result. Bioinformatic analysis was performed on RAPs and TAPs. Results: CB and irAE models displayed strong predictive performance with a high correlation between predicted probability and observed CB or irAE rate (CB model: R2=0.98; p-value<0.0001; irAE model: R2= 0.92; p-value <0.0001). In the CB model, patients classified as PROphet-positive achieved significantly longer overall survival (OS) than PROphet-negative patients, with a median OS of 25.9 vs. 10.8 months (hazard ratio, HR = 0.51; p-value<0.001). Furthermore, patients with PD-L1 ≥50% tumors and a PROphet-negative result had significantly longer OS with ICI-chemotherapy in comparison to ICI monotherapy (HR=0.23, p-value = 0.0003), suggesting that combination therapy is preferable for such patients despite high PD-L1 levels. Bioinformatic analysis showed that RAPs related to immune regulation, angiogenesis, chemo-resistance, and other potential resistance mechanisms displayed higher expression levels in patients who did not benefit from treatment. Neutrophil- and inflammation-related proteins were significantly enriched in patients who experienced irAEs. Conclusions: We describe two machine-learning based predictive tools that may be used separately or in combination to inform treatment decisions based on the patient’s likelihood to benefit from ICI therapy and to develop serious irAEs. Our findings also provide biological insights related to treatment resistance and immune-related toxicity. Citation Format: Michal Harel, Petros Christopoulos, Igor Puzanov, Jair Bar, Iris Kamer, Niels Reinmuth, Ina Koch, Mor Moskovitz, Adva Levy-Barda, Alona Zer, Michal Lotem, David Farrugia, Rivka Katzenelson, Gillian Price, Abed Agbarya, Helen Cheley, Adam Hassani, Mahmoud Abu-Amna, Tom Geldart, Anirban Chatterjee, Andreas Polychronis, Maya Gottfried, Yanyan Lou, Tatiana Harkovsky, Alison Brewster, Ido Wolf, Ella Tepper, Ben Yellin, Itamar Sela, Coren Lahav, Yehonatan Elon, Raya Leibowitz, Adam P. Dicker, Young Kwang Chae, Ryan J. Sullivan, David Carbone, David Gandara, Jarushka Naidoo. Plasma proteomics-based models for predicting therapeutic benefit and immune-related adverse events in non-small cell lung cancer patients treated with immunotherapy [abstract]. In: Proceedings of the American Association for Cancer Research Annual Meeting 2024; Part 1 (Regular Abstracts); 2024 Apr 5-10; San Diego, CA. Philadelphia (PA): AACR; Cancer Res 2024;84(6_Suppl):Abstract nr 1208.