Enhancing Multiomics With AI in Localized NSCLC

The thoughtful integration of artificial intelligence (AI) with multiomics is reshaping the diagnostic and therapeutic landscape in localized non–small cell lung cancer (NSCLC). Although the hype around AI often overshadows its practical limitations, the integration of digital pathology, radiomics, and multiparametric modeling is beginning to provide real clinical utility, particularly in the neoadjuvant setting.

In a presentation delivered at the 26th Annual International Lung Cancer Congress, Sandip Patel, MD, discussed the rise of AI-enabled imaging and digital pathology in predicting treatment response, the use of multiomics and machine learning for toxicity prediction and risk stratification, and a cautious but promising vision of AI as an assistive, not autonomous, tool in clinical oncology.¹

Patel is a professor at the University of California, San Diego (UCSD), medical director of Clinical Research Informatics, leader of Experimental Therapeutics, coleader of the Solid Tumor Therapeutics Program, and deputy director of the Sanford Stem Cell Clinical Center at UCSD Moores Cancer Center.

AI in Imaging and Digital Pathology

A standout example of AI’s evolving role in lung cancer care is Sybil, an AI model developed at Massachusetts General Hospital (MGH) by Lecia Sequist, MD, MPH, and colleagues. Designed for low-dose CT lung cancer screening, Sybil leverages deep learning to detect patterns that elude even expert radiologists. Importantly, Sybil has demonstrated the phenomenon of constructive hallucinations, or instances where the AI flags nodules missed by human readers that later manifest as radiologically visible lesions.² This ability to anticipate disease presence before human-detectable manifestation opens new possibilities in early diagnosis and surveillance.

“Everything AI produces is a hallucination; it’s just most things are useful hallucinations,” Patel said. “What’s interesting from the MGH data is that in the lung cancer screening population, there were several nodules that 2 radiologists felt were low risk or not even there. Sybil said, ‘Oh, there’s something there.’ Then 2 or 3 or 6 months later, a nodule showed up.”

Despite its promise, Sybil, like all AI, suffers from a lack of epistemological grounding. Without a true gold standard, clinicians must remain cautious of false positives and negatives. Equally transformative is AI’s role in evaluating responses to neoadjuvant therapy. Traditional assessment of pathologic response is highly subjective, especially in intermediate cases in which complete tumor eradication isn’t observed, Patel explained.

To address this, investigators are integrating computer vision algorithms into digital pathology workflows. One case study from the phase 2 LCMC3 trial (NCT02927301) used AI to standardize the measurement of pathologic response to neoadjuvant atezolizumab (Tecentriq).³ These models help quantify tumor regression more objectively.

Within the same study, led by Sanja Dacic, MD, PhD, of Yale School of Medicine, investigators found that digital major pathologic response, an AI-determined response metric, outperformed subjective human interpretation in predicting intermediate patient outcomes when there was discordance between pathologists.

Multiomics and Radiomics

AI is also proving useful in treatment planning by integrating diverse data streams to predict adverse effects. One area of focus is immune-related pneumonitis, a well-known toxicity in the context of immunotherapy. Here, radiomics, or the extraction of high-dimensional data from standard medical images, is combined with multiparametric modeling to identify patients at risk before symptoms emerge.

Research by Jarushka Naidoo, MBBCh, a consultant medical oncologist at Beaumont Hospital at the Royal College of Surgeons in Ireland Cancer Centre in Dublin, exemplifies this. Her multivariate analysis used imaging features, electronic medical record variables, and baseline laboratory data to predict pneumonitis risk.⁴ The model’s high negative predictive value (0.93; 95% CI, 0.90-0.96) makes it a valuable tool for ruling out toxicity and avoiding unnecessary treatment interruptions.

This is particularly important in the neoadjuvant setting, where radiologic findings such as nodal “flares” can mislead clinicians. What might appear to be progressive disease may represent immune cell infiltration rather than tumor growth.⁵ Radiomics could help differentiate these phenomena, guiding oncologists on whether to continue or halt immunotherapy prior to surgery.

“This is a red herring. These are immune cells ready to attack the tumor, and this patient needs to continue therapy. Radiomics can really help us in discerning which patients are going to benefit vs not,” Patel explained.

Moreover, radiomic features could eventually help discern which patients derive the most benefit from novel agents such as datopotamab deruxtecan (Datroway). For instance, AI-assisted image analysis has been used to evaluate whether TROP2 expression is predictive of therapeutic response via whole slide imaging and quantitative immunohistochemistry scoring. Work led by Marina Garassino, MD, of the University of Chicago Medicine, in this space suggests that combining AI with digital pathology may help match antibody-drug conjugates with biologically appropriate candidates, improving outcomes and avoiding overtreatment.⁶

Large Language Models

Although imaging and pathology applications are gaining traction, there’s growing skepticism about AI models purporting to replace oncologists. Much of this concern stems from flawed comparisons between AI and human performance in artificial test scenarios.

A case in point: Microsoft’s claim of achieving “medical superintelligence” by having an AI outperform physicians on the New England Journal of Medicine case vignettes.⁷ These studies typically favor the AI by removing the contextual ambiguity that real-life clinicians face, such as missing data, unclear symptom onset, and differential diagnoses.

In contrast, tools such as the American Society of Clinical Oncology’s retrieval- augmented generation–based guidelines assistant represent a more grounded application.

These systems combine large language model architecture with curated clinical guidelines, enhancing the reliability of AI-generated recommendations.⁸ Similarly, models such as the ScoutAI tool ingest structured clinical data and professional meeting content rather than open internet sources, creating a more trustworthy information substrate for oncologists.

Ambient AI for documentation is another area with immediate clinical utility. These systems transcribe and summarize clinical conversations in real-time, reducing the burden of electronic charting, improving documentation accuracy, and freeing physicians to focus on patient care.⁹

However, AI’s reliability is still influenced by how it’s prompted and how many iterative queries are made. Large language models can “degenerate” over time in chained prompts, leading to increasingly inaccurate responses—a phenomenon evident in older models that hallucinated PubMed IDs or recommended contraindicated drug combinations when pushed with edge-case scenarios, Patel noted.

Integration, Not Replacement

AI is unlikely to replace oncologists anytime soon, but it is poised to augment them in profound ways. From standardizing pathologic response measurements to refining imaging interpretations and predicting toxicities, AI offers practical tools that address some of oncology’s most nuanced challenges.

In localized NSCLC, particularly in the neoadjuvant space, these tools are already beginning to influence trial design, regulatory end points, and even patient selection. As multiomics and machine learning models become more integrated into routine workflows, the field must ensure that their implementation is guided by rigorous validation, clinician oversight, and, most importantly, patient benefit.

The promise of AI in NSCLC lies not in sensational headlines, but in practical, iterative progress. By embracing the technology’s strengths, while remaining vigilant to its limitations, oncologists can unlock new avenues of precision medicine that truly move the needle for patients.

Editor’s Note: This article was written with the help of ChatGPT.

REFERENCES:

1. Patel S. Multiomic and AI-assisted approaches in localized NSCLC. Presented at: 26th Annual International Lung Cancer Congress; July 25-26, 2025; Huntington Beach, CA.

2. Mikhael PG, Wohlwend J, Yala A, et al. Sybil: a validated deep learning model to predict future lung cancer risk from a single low-dose chest computed tomography. J Clin Oncol. 2023;41(suppl 12):2191-2200. doi:10.1200/JCO.22.01345

3. Dacic S, Travis WD, Giltnane JM, et al. Artificial intelligence–powered assessment of pathologic response to neoadjuvant atezolizumab in patients with NSCLC: results from the LCMC3 study. J Thorac Oncol. 2024;19(5):719-731. doi:10.1016/j.jtho.2023.12.010

4. Naidoo J, Haakensen VD, Bar J, et al. Quantitative radiomics for the detection of symptomatic pneumonitis following chemoradiotherapy in patients with stage III unresectable NSCLC. Ann Oncol. 2024;35(suppl 2):S794. doi:10.1016/j.annonc.2024.08.1298

5. Cascone T, Weissferdt A, Godoy MCB, et al. Nodal immune flare mimics nodal disease progression following neoadjuvant immune checkpoint inhibitors in non-small cell lung cancer. Nat Commun. 2021;12(1):5045. doi:10.1038/s41467-021-25188-0

6. Garassino MC, Sands J, Paz-Ares L, et al. Normalized membrane ratio of TROP2 by quantitative continuous scoring is predictive of clinical outcomes in TROPION-Lung 01. J Thorac Oncol. 2024;19(10):S2-S3. doi:10.1016/j.jtho.2024.09.015

7. Nori H, Daswani M, Kelly C, et al. Sequential diagnosis with language models. arXiv. Updated July 2, 2025. Accessed July 28, 2025. doi:10.48550/arXiv.2506.22405

8. Introducing ASCO guidelines assistant. ASCO. Accessed July 28, 2025. https://www.asco.org/practice-patients/guidelines/assistant

9. Tierney AA, Gayre G, Hoberman B, et al. Ambient artificial intelligence scribes to alleviate the burden of clinical documentation. NEJM Catal Innov Care Deliv. 2024;5(3). doi:10.1056/CAT.23.0404