Deep learning with limited data: a transfer learning approach for transcriptomic survival prediction.

Area of Science:

• Oncology
• Bioinformatics
• Machine Learning

Background:

• Deep neural networks (DNNs) show promise for analyzing complex cancer transcriptomic data.
• Data scarcity and heterogeneity limit DNNs in cancer research.
• Transfer learning (TL) is underutilized for cancer survival prediction despite its success in other areas.

Purpose of the Study:

• To evaluate transfer learning for disease-free survival prediction using The Cancer Genome Atlas (TCGA) RNA-seq data.
• To compare TL models against tumor-specific baseline models.
• To assess the biological relevance of TL models in cancer research.

Main Methods:

• Utilized RNA-seq data from 7509 patients across 27 tumor types.
• Trained a pan-cancer DNN and compared TL models (pre-trained and fine-tuned) with independent tumor-specific models.
• Assessed performance using concordance index and biological relevance via LIME-like interpretability and gene-set enrichment analysis.

Main Results:

• Transfer learning outperformed tumor-specific models in 24 out of 27 tumor types.
• Performance gains were most significant in smaller patient cohorts.
• Fine-tuning adapted general cancer patterns to specific tumor data, confirmed by enrichment analysis.

Conclusions:

• Transfer learning is a powerful strategy for improving cancer survival prediction from transcriptomic data, particularly for data-limited scenarios.
• TL effectively exploits shared molecular patterns across diverse cancer types.
• This work supports TL as a robust method for survival modeling and advances the development of foundation models in molecular oncology.