# Identifying Novel Drivers of Drug Sensitivity using an AI-Enabled Multi-Modal Biomarker – Osimertinib sensitivity beyond EGFR

### Authors
Maayan Baron, Felicia Kuperwaser, Colin Tang, Sunil Kumar, Dillon Tracy, Zong Miao, Sepideh Foroutan, Amy Sheide, Nathaniel Tann, Samantha Pindak, Jacob Kaffey, Jordan Wolinsky, Sean Klei, Brandon Funkhouser, Nick Lee, Patrick Bohan, Taylor Wood, Fahad Khan, Ripple Khera, Jean Michel Rouly, Anshu Jain, Jeff Sherman and Emily Vucic  
Zephyr AI 1800 Tysons Blvd Suite 901, McLean VA 22102 | [www.zephyrai.bio](http://www.zephyrai.bio/)

## Bridging the Predictive Biomarker Gap: AI-Powered Drug Response Prediction Using Clinicogenomics Data Available in Real-World Settings

### Abstract # 4640

- Precision oncology remains limited by the lack of effective biomarkers. Few tumors harbor actionable alterations, and even when present, many patients fail to benefit. Conversely, some patients without recognized biomarkers respond to therapies—highlighting the need for more robust, biologically informed approaches.
- Traditional biomarkers, based on single mutations or genomic signatures, often miss critical drivers of drug response—especially in late-line, heterogeneous populations.
- AIM-Bx, Zephyr AI’s multi-modal, AI-enabled software-based platform, addresses this gap by integrating clinical, molecular, pharmacological, and phenotypic data to generate real-world validated drug response predictions.
- As a use case, AIM-Bx predicted osimertinib sensitivity in NSCLC using tissue and liquid biopsy data, identifying EGFR+ non-responders and EGFR– responders. These results highlight AIM-Bx’s ability to uncover hidden biological drivers, expand targeted therapy utility, and enable longitudinal prediction in real-world contexts.

### Fig 1 | AI-Enabled Multi-Modal Biomarkers (AIM-Bx) Architecture Overview.

Compatible with NGS data from most commercial and LDT panels. NGS profiles were filtered to biopsies collected within 2 years prior to osimertinib treatment start.

### AIM-Bx Outputs

Clinical & NGS data to run AIM-Bx predictions, along with real-world treatment outcomes for post-hoc evaluation, were sourced from public (AACR Project GENIE) and non-public data (CancerLinQ, Aster TM Insights Optum® Market Clarity, TM Guardant Health). Sensitive and non-sensitive groups were biologically characterized post hoc using Vulnerability Networks™ and gene

### Fig 2 | Evaluation Framework.

### AIM-Bx Stratifies NSCLC Patients by Osimertinib Sensitivity with Clinically Meaningful Outcomes Across Diverse Real-World Cohorts

### Fig 3 | AIM-Bx Identifies Osimertinib Responders Independent of EGFR Status.

NSCLC clinicogenomic and outcome data were sourced from AACR Project GENIE (N=97), CancerLinQ (N=32), and Aster Insights (N=21). 
- AIM-Bx–predicted osimertinib-sensitive NSCLC patients had significantly improved rw-OS compared to predicted insensitive patients (median 10 months; HR = 0.54 indexed from treatment start). 
- Most patients were on-label (EGFR+) and AIM-Bx predictions (ZephyrAI) were significantly associated with survival in a univariable Cox PH analysis adjusting for available confounders. 
- EGFR variants were not significantly enriched in AIM-Bx–predicted sensitive vs insensitive groups.

### Fig 5 | AIM-Bx Validates Osimertinib Predictions Using Liquid Biopsy–Derived NGS Data.

A NSCLC cohort (N=435) treated with osimertinib and profiled via ctDNA (Guardant Health) was used to evaluate AIM-Bx on liquid biopsy inputs. AIM-Bx–predicted sensitive patients had significantly improved outcomes: 
- rw-OS: +18 months; HR = 0.70, p < 0.05 
- rw-PFS: +20 months; HR = 0.72, p < 0.05, indexed to osimertinib treatment start dates. AIM-Bx predictions are significantly associated with improved survival in univariable Cox models, alongside patient sex (female) as an independent factor. Results support AIM-Bx performance across diverse NGS inputs, including ctDNA, positioning AIM-Bx for broader clinical integration and longitudinal modeling.

### Uncovering Tumor Dependencies and Transcriptomic Programs Driving Osimertinib Response via VNs and Reconstructed mRNA

### Fig 6 | Vulnerability Networks Reveal Distinct Tumor Dependencies Associated with Osimertinib Response Not Captured by EGFR Genomic Status.

An expanded NSCLC cohort (N=308) from Optum® Market Clarity was analyzed in a secure ODDW environment. AIM-Bx–predicted sensitive patients had significantly improved outcomes compared to non-sensitive patients:
- rw-PFS: +16 months median survival (HR = 0.51, p < 0.001) 
- rw-OS: +17 months median survival (HR = 0.50, p < 0.001).

Real-world outcomes were indexed to osimertinib treatment start. AIM-Bx predictions (ZephyrAI) were the only factor significantly associated with improved survival in univariable Cox PH models adjusting for available confounders. Results support AIM-Bx performance across diverse real-world settings and heterogeneous NGS inputs.

VNs were generated for NSCLC tumors, with two networks significantly enriched in each AIM-Bx–predicted sensitive and insensitive groups. VNs reveal predicted patient-specific mechanisms underlying AIM-Bx response predictions.

- Responder VNs were characterized by dependencies on EGFR, EGFR signaling and key survival and metabolic pathways, suggesting that network-level activity—beyond EGFR DNA level alterations—drive sensitivity to osimertinib.
- Non-responding NSCLC VNs were characterized by dependencies in cell cycle, transcriptional, and integrin signaling pathways, consistent with EGFR independence and known osimertinib resistance mechanisms.

### Fig 7 | GSEA of Reconstructed mRNA Profiles

Reconstructed mRNA profiles for tumors were generated using the AIM-Bx Patient Encoder from clinical and genomic inputs, and analyzed by GSEA. Less-sensitive tumors were enriched for oncogenic programs linked to proliferation and poor outcomes in NSCLC, including DREAM, E2F, MYC, KRAS, and NRF2—pathways previously associated with TKI resistance, including resistance to osimertinib. In contrast, predicted osimertinib-sensitive tumors were

## Platform Summary & Future Directions

- AIM-Bx predicts drug response using routinely available clinical and genomic data from tissue or liquid biopsy, revealing tumor vulnerabilities and biological programs beyond traditional DNA biomarkers.
- Our software seamlessly integrates into clinical workflows and supports rapid deployment in trials or retrospective studies using standard NGS and LDT panel data.