In the world of precision medicine, predicting how long a cancer patient might survive is a critical but challenging task. The data used to make these predictions—genomic profiles—is notoriously difficult to handle. Researchers often have thousands of gene expressions (features) but only a few hundred patients (samples), leading to the classic "curse of dimensionality."
A paper from ICML 2019, "A Multitask Multiple Kernel Learning Algorithm for Survival Analysis with Application to Cancer Biology," introduces a clever way to bypass these limitations by combining two powerful machine learning paradigms: Multitask Learning and Multiple Kernel Learning.
The Challenge of Genomic Survival Analysis
Survival analysis aims to predict the time until an event (like death) occurs, accounting for "censored" data (patients who are still alive at the end of a study or drop out). Doing this with genomic data presents two major hurdles:
- Small Sample Sizes: While data exists for many cancers, individual cohorts (e.g., pancreatic cancer) are often too small to train complex models without severe overfitting.
- Biological Complexity: Genes do not act in isolation; they work in interconnected pathways. Standard models often treat genes as independent features, ignoring the rich "prior knowledge" we have about biological systems.
Existing methods either select features before training or train separate models for every pathway, both of which struggle with the highly correlated nature of genomic data.
Methods: Introducing Path2MSurv
To address these challenges, the authors developed Path2MSurv, an algorithm that extends the Survival Support Vector Machine (SVM) in two significant ways:
- Multiple Kernel Learning (MKL): Instead of using one giant feature matrix, the authors create separate "kernels" (similarity matrices) for different biological pathways or gene sets. The model learns a weight ($\eta$) for each kernel, effectively "picking" which pathways are most informative for survival.
- Multitask Learning (MTL): Rather than training a separate model for Breast Cancer and another for Lung Cancer, Path2MSurv trains on 20 different cancer types simultaneously. It assumes these cancers share some underlying biological mechanisms, allowing the model to "borrow strength" from larger cohorts to improve predictions for smaller ones.
The optimization uses an alternating approach: it solves a standard survival SVM for each cancer type given a set of pathway weights, then updates those weights based on the performance across all cancer types. Because the weights lie on a simplex (they sum to 1), the model naturally produces a sparse selection of pathways, making the results highly interpretable for biologists.
Datasets
The study utilized a massive real-world dataset from The Cancer Genome Atlas (TCGA), comprising gene expression profiles and clinical data for 7,655 patients across 20 different cancer types. To build the kernels, they used the Hallmark gene set collection (50 well-defined biological states) and the Pathway Interaction Database (PID) (196 human signaling pathways).
Key Insights and Results
By forcing the model to find a shared set of important pathways across all 20 cancers, Path2MSurv avoids chasing noise in small, specific datasets. Furthermore, by using MKL, the model doesn't just give a prediction; it provides a ranked list of biological processes, offering researchers potential therapeutic targets.
The experimental results highlight several major advantages of this approach:
- Superior Performance: Path2MSurv (referred to as MTMKL in the results) outperformed standard Random Forests (RF) on 13 out of 20 cancer datasets.
- The Multitask Advantage: The multitask version outperformed its single-task counterpart (learning one cancer at a time) on roughly 70-75% of the datasets.
- Robustness: In several cancers (like GBM and LUSC), standard Random Forests failed to produce better-than-random predictions (C-index < 0.5), while Path2MSurv maintained its predictive power.
- Biological Validation: The model identified Glycolysis and Angiogenesis as the most informative pathways for survival across almost all replications. These are well-known hallmarks of cancer progression, providing strong validation for the algorithm's biological relevance.
Limitations and Future Work
The current model assumes all 20 cancers are related enough to share the same pathway weights. In reality, some cancers (like squamous cell carcinomas) are more similar to each other than to others.
The authors suggest that a logical next step is Task Clustering: automatically grouping similar cancers together and learning shared weights within those clusters rather than across the entire pan-cancer suite. Additionally, while this paper focused on gene expression, the MKL framework is perfectly suited for "multi-omics" (combining DNA mutations, proteomics, and methylation data), which could further boost predictive accuracy in precision oncology.
Summary updated