Autoencoder/RandomForest–TabPFN for Cross-Cancer Metabolomics: Prostate and Breast Cancer Diagnosis Using Paper Spray and Ion Mobility-Mass Spectrometry Techniques

Accurate and rapid disease diagnosis, particularly in prostate cancer (PC) and breast cancer (BC), is critical for early intervention and improved patient outcomes. Metabolomic signatures represent a robust molecular framework for elucidating cancer-associated biochemical reprogramming. The use of Artificial Intelligence (AI) in biology in recent years has become widespread and promising. This study introduces a novel predictive method that integrates an Autoencoder, random forest-based feature selection and Tabular Prior-data Fitted Network (TabPFN) to achieve high diagnostic accuracy from metabolomics data of prostate and breast cancer patients. The datasets were acquired using Paper Spray Ionization Mass Spectrometry (PSI-MS) and Flow Injection;Traveling-Wave Ion Mobility–Mass Spectrometry (FI-TWIM-MS) of individuals diagnosed with PC and BC. When leveraging metabolomic profiling data from two distinct sources, prostate cancer urine and serum samples, the proposed model achieved an accuracy up to 98.75% in distinguishing diseased from healthy conditions. Additionally, we employed a breast cancer dataset containing metabolic and lipidomic signatures acquired from core needle biopsies using a miniature MS platform coupled with PSI to assess the fidelity of our implementation across distinct cancer types. Our results on a well-characterized targeted dataset show that we can effectively reduce high-dimensional data into latent feature representations. At the same time, TabPFN captures tumor progression-related changes and feature interaction, thereby enhancing the possibility that the model will be a highly potent and effective tool for stage-specific diagnostic precision. Most existing machine learning approaches for disease diagnosis primarily rely on imaging, genomics, or clinical parameters, often overlooking the critical role of metabolites in identifying disease-specific biochemical signatures. By integrating metabolite-specific data with a robust deep-learning approach, this study demonstrates the transformative potential of AI in metabolomics-based diagnostics. The proposed model offers scalability and versatility, with applications extending beyond oncology to a much broader disease profiling aspect. These findings emphasise the value of combining multi-source metabolomic data with deep learning to advance personalised medicine and enhance diagnostic efficiency in clinical practice.