CancerSiamese

CancerSiamese predicts primary and metastatic tumor types unseen during training from gene expression profiles using one-shot deep learning with Siamese convolutional neural networks.

Key Features:

• Type-Agnostic Expression Representations: Hypothesizes a set of expression representations that encode similarities or dissimilarities between cancer samples regardless of specific cancer type.
• Convolutional Neural Networks (CNNs): Two parallel CNNs process pairs of query and support gene expression profiles and are joined by a similarity function to learn representations of cancer-type similarity.
• Network Transfer Learning: Employs network transfer learning with training on datasets such as the Cancer Genome Atlas (TCGA) and MET500.
• Performance Metrics: In N-way prediction tests, reported average accuracy improvements of 8% for primary tumors and 4% for metastatic tumors compared to the 1-Nearest Neighbor (1-NN) classifier.
• Marker-Gene Analysis: Uses guided gradient saliency maps and feature selection to identify top marker-gene candidates (100 genes for primary and 200 genes for metastatic cancers) and performs functional analysis to reveal cancer-related functions.

Scientific Applications:

• Unseen tumor-type classification: Predicts rare or novel primary and metastatic tumor types from limited samples without prior examples in the training set.
• Diagnostic and prognostic support: Enables classification of tumor samples with few-shot input to support diagnostic and prognostic studies.
• Marker discovery and therapeutic insight: Identifies marker-gene candidates and associated functions that can inform targeted therapy development and personalized medicine research.

Methodology:

Applies one-shot learning with two parallel Siamese CNNs on pairs of query and support gene expression profiles combined via a similarity function, uses network transfer learning with TCGA and MET500 data, evaluates performance in N-way prediction scenarios against 1-NN, and derives marker genes using guided gradient saliency maps and feature selection.