Multimodal AI Predicts Immune Checkpoint Inhibitor Response

Multimodal AI predicts immune checkpoint inhibitor response from clinically available inputs and whole-slide images with explainable tumor biology and combination therapy insights

Felicia Kuperwaser $ ^{1} $ , Sunil Kumar $ ^{1} $ , Momeneh Foroutan $ ^{1} $ , Dillon Tracy $ ^{1} $ , Kevin Freisen $ ^{1} $ , Taylor Wood $ ^{1} $ , Zong Miao $ ^{1} $ , Nathaniel Tann $ ^{1} $ , Fahad Khan $ ^{1} $ , Jiemin Liao $ ^{2} $ , Kevin Shah $ ^{2} $ , Aaron Hardin $ ^{2} $ , Jean-François Martini $ ^{2} $ , Jean Michel Rouly $ ^{1} $ , Allen Chao $ ^{1} $ , Anshu Jain $ ^{1} $ , Jeff Sherman $ ^{1} $ Maayan Baron $ ^{1} $ and Emily Vucic $ ^{1} $

Guardant Health, Palo Alto, CA

ICIs provide durable benefit for a subset of patients, but most do not respond, and current biomarkers (PD-L1, TMB, MSI) show limited predictive value across cancer types.
Improved patient stratification is needed to optimize treatment selection, reduce unnecessary toxicity, and identify patients who may benefit from rational combination strategies.
We developed AIMio, a multimodal, biologically interpretable AI model that predicts ICI response using routine clinical inputs, including DNA inputs derived from tissue and ctDNA, and as proof-of-concept, whole slide images (WSI) （Figure 1）
AIM-io reconstructs gene-expression and tumor-microenvironment programs, enabling biologically grounded response prediction without additional assays.
The model also predicts small-molecule sensitivities, supporting rational ICI combination strategies and scalable clinical deployment.

1 REAL-WORLD INPUTS

AIM-io Identifies ICI Responders From Routine Tissue DNA

AIM-io

AIM-io was trained on datasets from published studies (N = 1,663) [1-4] and Aster Insights (N = 1,113). In the validation cohort set (N = 470), AIM-io sensitivity predictions (ZephyrAl) outperformed conventional tumor mutational burden (TMB)-based stratification (Figures 2A-B). The ZephyrAl label identifies an inflammatory tumor phenotype, even among TMB low samples (Figure 2C). Within ZephyrAI+ tumors, TMB low samples are enriched for innate immune features, suggesting a distinct biological mechanism driving response (Figure 2D).

INTERPRETABLE OUTPUTS

Drug Combination Hypotheses

AIM-io Predicts Improved ICI Outcomes From Liquid Biopsy DNA

ZephyrAI predictions were generated in a cohort of liquid biopsy non-small cell lung cancer (NSCLC) samples from Guardant360 CDx $ ^{\circ} $ （Guardant Health, N =15,019). ZephyrAl+ patients showed significantly improved rwOS compared to ZephyrAl- patients (Figure 4A). Differential analyses of reconstructed TME signatures [6-7] and cell type fractions reveal enrichment of inflammatory features in ZephyrAl+ samples consistent with their predicted sensitivity (Figure 4B). Predicted drug sensitivities further suggest potential combination therapy strategies with ICI treatment (Figure 4C).

Figure 2 | AIM-io Identifies Patients with Improved Real-World (rw) Outcomes Over Conventional Stratification Approaches. (A) AIM-io-predicted ICI-sensitive patients had significantly improved OS compared to predicted less-sensitive patients (median 14 months; HR = 0.31 indexed from treatment start, left panel). AIM-io predictions (ZephyrAl) were significantly associated with survival in a univariate Cox PH analysis adjusting for available confounders (right panel). (B) ZephyrAl outperforms tumor mutational burden (TMB) for patient stratification (all pairwise p < 0.05, except ZephyrAI+/TMB+ vs ZephyrAI+/TMB-). (C) Among TMB low samples, ZephyrAl+ tumors exhibit increased immune cell fractions and tumor microenvironment (TME) features, consistent with an inflammatory phenotype (FDR < 0.1). (D) Within ZephyrAI+ samples, TMB low samples are enriched for innate immune features and immune signaling pathways (GSEA, FDR <

AIM-io Predicts Improved ICI Outcomes From Whole-Slide Images (H&E)

DE TME Features (FDR < 0.1) Zephyr+ vs Zephyr-

= 11473 7868 5843 4330 3287 2489 1862 1400 1029 = 3546 2483 1805 1374 1066 809 593 436 322

AI-Enabled Patient Stratification and Rational Combination Hypotheses for ICI Development Using Routine Clinical Inputs

AIM-io is a biologically interpretable, multimodal framework for predicting ICI response from clinically accessible inputs, including ctDNA liquid biopsy DNA and WSI embeddings. By integrating reconstructed expression, tumor microenvironment features, and predicted therapeutic vulnerabilities, AIM-io enables assay-agnostic evaluation of immunotherapy response and identification of rational combination strategies. This approach supports real-world validation, prospective stratification, and biological characterization of responders for current and next-generation ICI therapies. Prospective evaluation is warranted.