Transcriptomic Deconvolution of Neuroendocrine Neoplasms Predicts Clinically Relevant Characteristics

Area of Science:

• Oncology
• Computational Biology
• Genomics

Background:

• Pancreatic neuroendocrine neoplasms (panNENs) present diagnostic challenges due to their rarity and heterogeneity.
• Accurate classification is crucial for effective treatment planning, but limited training data hinders machine learning (ML) development for rare panNENs.

Purpose of the Study:

• To develop and validate a novel ML framework for classifying panNENs using readily available healthy tissue data.
• To assess the framework's ability to predict key clinical-pathological characteristics and patient outcomes.

Main Methods:

• A multi-step ML framework was designed to deconvolve panNEN transcriptomes into cell type proportions by comparing gene expression profiles with healthy pancreatic cells.
• The model was trained exclusively on data from healthy pancreatic tissues.
• Performance was evaluated against established methods, including those using proliferation markers like MKI67.

Main Results:

• The ML framework successfully predicted overall patient survival, neoplastic grade, and tumor subclassification (carcinoma vs. tumor).
• The deconvolution approach demonstrated prognostic value, comparable to models trained on proliferation data.
• The method quantifies panNEN dedifferentiation in silico, enhancing classification of aggressive subtypes.

Conclusions:

• This ML framework offers a powerful, data-efficient tool for classifying panNENs, leveraging healthy tissue transcriptomic data.
• The approach provides prognostic insights and refines classification beyond traditional proliferation-based methods.
• It holds significant potential for improving diagnostic accuracy and clinical management of challenging panNEN cases.