RNA sequencing can measure the activity of tens of thousands of genes simultaneously, giving a comprehensive picture of what is happening inside a tumour cell. However, this data overload creates a fundamental challenge: which genes are actually driving cancer growth, and which are irrelevant background noise?
Standard gene selection methods often require large datasets to be reliable, but collecting tissue from pancreatic cancer patients is inherently difficult because surgery is the only way to obtain samples. This study addresses the challenge of finding meaningful genes from a small patient cohort.
The researchers developed a novel deep learning model with a dedicated feature-selection layer — an additional component that learns to assign importance scores to individual genes during training. Genes with high weights in this layer are the ones the network relied on most when making its predictions.
Cancer tissue and adjacent normal pancreatic tissue were collected from 13 patients with pancreatic ductal adenocarcinoma during surgery and immediately frozen to preserve gene expression. RNA sequencing was then performed to generate gene expression profiles.
The model was applied to two distinct tasks. Task 1 used samples from 6 patients to teach the model to distinguish cancerous tissue from normal pancreatic tissue. Task 2 used all 13 patients to predict whether a patient survived more than one year after surgery — a critical prognostic question.
By training the same model architecture on two different classification tasks, the researchers were able to identify different sets of biologically meaningful genes for each question: genes involved in cancer development versus genes associated with survival outcomes.
For distinguishing cancer from normal tissue (Task 1), the most frequently selected genes were ACACB, ADAMTS6, NCAM1, and CADPS. These genes are involved in fatty acid metabolism, extracellular matrix remodelling, and cell adhesion — all processes known to be dysregulated in pancreatic cancer.
For predicting one-year survival after surgery (Task 2), the top genes were CD1D, PLA2G16, DACH1, and SOWAHA. These are linked to immune function, lipid signalling, developmental transcription factors, and tumour suppression — reflecting the diverse biological pathways that influence patient outcomes.
Crucially, all eight identified genes were validated as prognostic factors in The Cancer Genome Atlas (TCGA) dataset — a large independent repository of cancer genomic data. This external validation confirms that the model selected genuinely biologically meaningful genes, not statistical artefacts.
Standard deep neural networks are often criticised as 'black boxes' — they make accurate predictions but cannot explain which inputs drove those predictions. This limits their usefulness in biology, where understanding mechanism is as important as accurate classification.
The key innovation in this study is the dedicated feature-selection layer added before the standard neural network layers. After training, the weights of this layer directly indicate how much each gene contributed to the network's decisions — providing a transparent, interpretable ranking of gene importance.
This approach is particularly valuable for gene expression data, where unlike images, there is no natural local structure between features. All 20,000+ genes interact globally, so a dense feature-selection layer that considers all possible interactions is more appropriate than the convolutional layers used in image analysis.
The genes identified by this study are not just diagnostic markers — they represent potential therapeutic targets. For example, ADAMTS6's role in extracellular matrix remodelling suggests it may be involved in the dense stromal barrier that makes pancreatic cancer so resistant to drug delivery.
The survival-associated genes identified in Task 2 could form the basis of a molecular prognostic panel — a blood or tissue test that predicts which patients are likely to have poor outcomes after surgery and might therefore benefit from more aggressive adjuvant treatment.
The authors demonstrate that deep learning can perform meaningful gene discovery even from very small datasets of 13 patients — a significant finding for rare cancers and surgical studies where large cohort sizes are inherently limited. This validates the approach for future genomic discovery studies.