A platform-independent AI tumor lineage and site (ATLAS) classifier

Our plain-English summary of A platform-independent AI tumor lineage and site (ATLAS) classifier, by N.R. Rydzewski, Y. Shi, C. Li, et al (Communications Biology, 2024).

The study

• Authors: N.R. Rydzewski, Y. Shi, C. Li, M.R. Chrostek, H. Bakhtiar, K.T. Helzer, M.L. Bootsma, T.J. Berg, P.M. Harari, J.M. Floberg, G.C. Blitzer, D. Kosoff, A.K. Taylor, M.N. Sharifi, M. Yu, J.M. Lang, K.R. Patel, D.E. Citrin, K.E. Sundling, S.G. Zhao
• Journal: Communications Biology
• Published: 2024;7(1):314
• Cited by: 1 article, as counted by NIH iCite
• Read it: PubMed · Publisher · Free full text

Abstract

Histopathologic diagnosis and classification of cancer plays a critical role in guiding treatment. Advances in next-generation sequencing have ushered in new complementary molecular frameworks. However, existing approaches do not independently assess both site-of-origin (e.g. prostate) and lineage (e.g. adenocarcinoma) and have minimal validation in metastatic disease, where classification is more difficult. Utilizing gradient-boosted machine learning, we developed ATLAS, a pair of separate AI Tumor Lineage and Site-of-origin models from RNA expression data on 8249 tumor samples. We assessed performance independently in 10,376 total tumor samples, including 1490 metastatic samples, achieving an accuracy of 91.4% for cancer site-of-origin and 97.1% for cancer lineage. High confidence predictions (encompassing the majority of cases) were accurate 98-99% of the time in both localized and remarkably even in metastatic samples. We also identified emergent properties of our lineage scores for tumor types on which the model was never trained (zero-shot learning). Adenocarcinoma/sarcoma lineage scores differentiated epithelioid from biphasic/sarcomatoid mesothelioma. Also, predicted lineage de-differentiation identified neuroendocrine/small cell tumors and was associated with poor outcomes across tumor types. Our platform-independent single-sample approach can be easily translated to existing RNA-seq platforms. ATLAS can complement and guide traditional histopathologic assessment in challenging situations and tumors of unknown primary.

© 2024. This is a U.S. Government work and not under copyright protection in the US; foreign copyright protection may apply.

In plain English

This study built and tested a computer program, called ATLAS, that uses tumor gene-activity (RNA expression) data to predict where a cancer started (site of origin) and its lineage, such as adenocarcinoma or squamous cell carcinoma. It did not study asbestos exposure. The authors trained two machine-learning models on 8249 samples from public databases, made up of patient tumors and laboratory cell lines. They then tested the models on 10,376 other tumor samples, including 1490 from cancers that had spread (metastatic). Overall accuracy was 91.4% for site of origin and 97.1% for lineage. For site of origin, accuracy was 92.1% in primary tumors and 86.8% in metastatic tumors. The model had never been trained on mesothelioma. In 88 lung mesothelioma samples, its sarcoma score was higher in the biphasic/sarcomatoid subtypes (25 samples) than in the epithelioid subtype (63 samples). The score separated these subtypes with an AUC of 0.81. Median survival was 15.0 months in the high-score group and 23.9 months in the low-score group (log-rank P = 0.049). The authors concluded the tool could complement traditional pathology, especially in challenging cases.

We wrote this plain-English summary from the full article with the help of AI, and a second AI model checked each sentence against the article. It is not the authors' wording; the abstract and the article are their own account.

What the study did and found

• Design: Development and validation of a machine-learning (gradient-boosted) classifier using previously published tumor RNA expression data
• Who was studied: Training set: 8249 samples, 7196 from The Cancer Genome Atlas (TCGA) and 1053 cell lines from the Cancer Cell Line Encyclopedia (CCLE). Validation set: 10,376 tumor samples (8886 primary, 1490 metastatic) from TCGA and 41 other datasets. Separate analyses covered 88 lung mesothelioma samples excluded from training and validation, and 198 neuroendocrine samples (neuroendocrine prostate cancer and small cell lung cancer).
• Exposure: No asbestos or other exposure was studied. The data analyzed were tumor gene-expression data. Asbestos is mentioned only in the background, as a common cause of mesothelioma.
• Main results:
  • In all 10,376 validation samples, overall accuracy was 91.4% for cancer site of origin and 97.1% for cancer lineage.
  • Site-of-origin accuracy was 92.1% in primary tumors (N = 8886) and 86.8% in metastatic tumors (N = 1490). Lineage accuracy was 97.4% in primary and 95.7% in metastatic samples.
  • When the classifier's prediction score was 0.99 or higher, accuracy was 98 to 99%, even in metastatic samples. This applied to 58.5% of validation samples for site and 75.1% for lineage.
  • Sarcoma lineage scores were higher in the 88 lung mesothelioma samples (median 0.0065) than in 637 non-small cell lung cancer samples (median 0.000007; P less than 0.001).
  • Within mesothelioma, the sarcomatoid and biphasic subtypes (N = 25, median 0.042) had higher sarcoma scores than the epithelioid subtype (N = 63, median 0.003; P less than 0.001). The score separated these groups with an AUC of 0.81.
  • Mesothelioma samples were split into high and low sarcoma-score groups. Median survival was 15.0 months in the high-score group and 23.9 months in the low-score group (log-rank P = 0.049).
  • Two mesothelioma samples scored high enough to be classified as sarcoma. They had originally been reported as one biphasic and one epithelioid. On blinded re-review, a pathologist described both as biphasic with approximately 90% sarcomatoid differentiation.
  • A lower 'differentiation score' separated small cell lung cancer (N = 137) from non-small cell lung cancer (N = 606) with an AUC of 0.963. It separated neuroendocrine prostate cancer (N = 61) from metastatic prostate adenocarcinoma (N = 721) with an AUC of 0.834.
  • In eight cancer subgroups, survival was significantly better with a higher differentiation score. The other subgroups trended in the same direction without reaching statistical significance.
• The authors' conclusion: The authors concluded that their platform-independent classifiers could complement traditional pathologic assessment, especially in challenging cases. They stated that histopathologic assessment remains the gold standard.
• Limitations the authors state:
  • Site-of-origin accuracy was lower in metastatic tumors, which the authors attributed partly to the models being trained on primary tumors.
  • Gastrointestinal and gynecologic tumors had worse classification accuracy.
  • Samples with low tumor purity (below 50%) had worse accuracy.
  • Rare tumor types are not currently well represented in the models.

Topics

The National Library of Medicine indexes this study under subjects our site covers: Mesothelioma.

Medical terms in this study, explained in our Lexicon: Next-Generation Sequencing (NGS), Adenocarcinoma, Sarcoma, Epithelioid, Biphasic Mesothelioma, Sarcomatoid, Mesothelioma.

Related studies

• Poorly differentiated neoplasms of unknown primary site: diagnostic usefulness of a molecular cancer classifier assay — F.A. Greco, W.J. Lennington, et al. (2015)
• Insulinoma-associated protein 1 is a sensitive and specific marker of neuroendocrine lung neoplasms in cytology specimens — E.E. Doxtader, S. Mukhopadhyay (2018)
• Pre-treatment high-sensitivity troponin T for the short-term prediction of cardiac outcomes in patients on immune checkpoint inhibitors — S. Petricciuolo, M.G. Delle Donne, et al. (2021)
• Whole exome sequencing of independent lung adenocarcinoma, lung squamous cell carcinoma, and malignant peritoneal mesothelioma: A case report — I. Vanni, S. Coco, et al. (2016)
• Peripheral PD-1 and Tim-3 percentages are associated with primary sites and pathological types of peritoneal neoplasms — H. Hu, J. Zhao, et al. (2023)
• Comparison of imaging based single-cell resolution spatial transcriptomics profiling platforms using formalin-fixed paraffin-embedded tumor samples — N. Ozirmak Lermi, M. Molina Ayala, et al. (2025)

Browse return to medical research.