Deep Learning Models Expedite Biomarker Discovery, Detection in Lung Cancer

One of the studies presented evaluated analyzed lung tumors and their microenvironment across patients who did or did not respond to checkpoint inhibitor therapy, which detected distinct metabolic patterns of response for immunotherapy.

Two strategies highlighted at the IASLC 2024 World Conference on Lung Cancer demonstrated the feasibility of combining pathology findings with deep learning artificial intelligence (AI) to speed up biomarker detection and discovery for patients with lung cancer.

In the first,¹ investigators from China found a high degree of accuracy for detecting biomarkers using whole slide images (WSIs) obtained from lymph node biopsies. In the second,² researchers from Australia analyzed the tumor and its microenvironment (TME) across patients who did or did not respond to checkpoint inhibitor therapy, which detected distinct metabolic patterns of response for immunotherapy.

“I think any additional tools that we can develop to help us understand the disease better are going to be highly important, and we need to be open to new technology and figure out how to rigorously develop it,” said Matthew Smeltzer, PhD, of the University of Memphis, in Memphis, TN, who presented new findings from the IASLC Global Survey on Biomarker Testing.³ In these findings, the highest barriers to biomarker testing were cost (reported by 27.2%) and time (13.9%). The presenters of the sessions postulated that deep learning methods could help alleviate some of these challenges. In the survey, the median turnaround time for tissue testing was 14 days and only a quarter received full reimbursement for all genetic tests.

Deep Learning for Biomarker Detection

In the first study highlighted at IASLC,¹ AI was trained using WSIs paired with known gene mutation status. The images came from a mixture of excisional and aspiration biopsy samples across different cohorts. The first cohort, which was used for training and was internal, consisted of samples from 1 center in China, which included 1717 WSIs from 1716 patients. A separate external cohort was also formed that included 15 different centers with 1718 patients and 1719 WSIs. This was further paired with a third cohort of multiple centers from The Cancer Genome Atlas. This cohort included 473 patients and 535 WSIs. The platform, which was labeled DeepGEM, was then also tested on 331 annotation-free WSIs from lymph node metastasis biopsies that were collected from 203 patients with lung cancer.

“DeepGEM model trained on primary region biopsies can be generalized to biopsies from lymph node metastases and shows potential for prognostic prediction of targeted therapy,” senior author Wenhua Liang, MD, from the Guangzhou Medical University, China, said during a presentation of the findings.

Other deep learning models have also been created to examine pathological images from various biopsy sources, the presenter noted. Prior to moving beyond cohort 1, the model was compared to these other tools. In this initial comparison, the DeepGEM model was found to be superior to DeepPATH-a, DeepPATH-p, MIL-RNN, CLAM, and TOAD (area under the curve [AUC], 0.881; accuracy 0.825). From the WSIs, the DeepGEM modal created spatial gene mutation maps that were like those performed by a pathologist analyzing immunohistochemistry findings.

After passing this initial test, the AI tool was further examined on the cohort 2 external dataset. In this group, the AUC was 0.842 across all types of biopsies for gene mutation detection. For excisional biopsies, the AUC was 0.860 and for aspiration it was 0.853. These findings remained consistent across common lung cancer mutations, including EGFR (AUC, 0.862), TP53 (AUC, 0.879), and KRAS (AUC, 0.822). Similar findings were seen in cohort 3, with an AUC of 0.872. This dataset was more comprehensive, containing a more diverse racial background, the presenter noted.

“Compared to previous studies, DeepGEM achieved robust and superior predictive performance across various genes validating on the largest multicenter datasets to date,” Liang, said in a statement. “The rapid prediction capabilities of DeepGEM allow for quicker decision-making in treatment, enabling patients with severe symptoms to receive targeted therapies sooner. Furthermore, it presents opportunities for multigene mutation detection and precision treatment in economically underdeveloped areas where genomic testing is unaffordable.”

DeepGEM continued to show a high level of specificity and sensitivity in the fourth cohort, which contained unannotated lymph node biopsies from metastatic lesions. The AUC for sensitivity and specificity for EGFR detection from the WSIs was 0.911 and the AUC was 0.882 for KRAS. These findings were further corroborated with overall survival. Those marked as having EGFR had a significantly better survival, showing a prognostic value for the EGFR prediction (P = .024).

“This innovative approach has the potential to transform the clinical management of lung cancer patients, making advanced genomic insights more accessible and actionable,” said Liang.

Deep Learning Finds Immunotherapy Response Signature

In the second study highlighted by IASLC,² patients with lung cancer were analyzed using the 45-plex PhenoCycler. For the analysis, patient samples were used from those who responded to immune checkpoint inhibitors (ICIs; n = 16) and those who did not (n = 25) to effectively identify different characteristics between the 2 groups. The most common ICIs administered in the responding arm were nivolumab (Opdivo; n = 11), pembrolizumab (Keytruda; n = 4), and durvalumab (Imfinzi; n = 1). There were 22, 3, and 0 patients receiving these medications in the non-responding arm, respectively.

Samples for the experiment went through a series of modularized spatial analyses. In the first step, the samples underwent multiplex staining for 36 different protein markers. These were imaged in cycles by CODEX. After this process was complete, images from the staining were processed with QuPath and segmented with Cellpose and a cyto2 pretrained model. A quality control step was applied to remove incomplete images and to ensure staining effectiveness. After these processes, the cells underwent phenotyping and were imported into Anndata, which is a Python package for handling annotated data. They were also inspected with QuPath. From here, spatial feature analysis of neighborhoods, clustering of cell types, cell proportions, and 3-way interactions were analyzed using the SpatialScore method.

In the initial analysis, a lower proportion of T regulatory cells was found more commonly in the responder population compared with the non-responder group (P = .01). When looking at cellular neighborhoods, those with macrophages and mixed tumor phenotypes were also found most in those who responded. To further explore the potential of deep learning methods, cell type exploration was performed using unsupervised clustering. This process identified more than a thousand spatial features.

In the unsupervised deep learning analysis of these features, 15 cell types and 43 cell subsets were identified, which could be clustered into 3 main metabolic states labeled as OXPHOS+, OXPHOS-, and PPP+, the latter of which consisted of various regulators of the pentose phosphate pathway. The cells in each of these deep learning-identified metabolic states were associated with higher proliferation and upregulation of the tumor stemless marker CD44. Specifically, tumors that tested high (over 40%) for the PPP+ metabolic state were found to be resistant to PD-1 agents and had lower overall survival rates.

“This research reveals complex relationships between metabolic states, immune cell functionality, and responses to immunotherapy, and offers a promising pathway toward developing predictive biomarkers for ICIs,” primary investigator Arutha Kulasinghe, PhD, from The University of Queensland, said in a statement.

References

1. Xiong S, Zhao Y, Ren Q, et al. Deep multiple instance learning-enabled gene mutation prediction of lung cancer from histopathology images. Presented at: 2024 IASLC World Conference on Lung Cancer; September 7-11, 2024; San Diego, CA. Abstract 1453.
2. Kulasinghe A, Monkman J, Markovits E, et al. Deciphering single-cell resolved tumour microenvironment profiles using spatial metabolic mapping. Presented at: 2024 IASLC World Conference on Lung Cancer; September 7-11, 2024; San Diego, CA. Abstract 1664.
3. Smeltzer M, Connolly C, Lantuejoul S, et al. The 2024 International Association for the Study of Lung Cancer (IASLC) Global Survey on Biomarker Testing. Presented at: 2024 IASLC World Conference on Lung Cancer; September 7-11, 2024; San Diego, CA. Abstract 3000.