SurvODE: Extrapolating Gene Expression Distribution for Early Cancer Identification

Tong Chen, Sheng WangarXiv 2021 · arXiv

Abstract

Understanding cancer progression requires modeling how gene expression evolves over time. However, clinical datasets often suffer from censored survival data and irregular temporal snapshots. We present SurvODE, a framework that combines Variational Autoencoders (VAEs), Neural Ordinary Differential Equations (Neural ODEs), and Cox Proportional Hazards (Cox-PH) models. By estimating pseudo-survival times for censored observations, SurvODE enables the continuous modeling of gene expression distributions, allowing for the extrapolation of biological states to early disease stages where data is typically unavailable.

Introduction

Cancer research increasingly relies on large-scale genomic datasets like The Cancer Genome Atlas (TCGA). Despite the wealth of data, two primary obstacles hinder the reconstruction of disease progression: censored data and irregular population snapshots. Censoring occurs when a patient's exact survival time is unknown (e.g., they are still alive at the study's end), providing only a lower bound on survival. Furthermore, genomic profiles are usually single-point observations across a population rather than continuous time-series for individuals. Existing methods often fail to capture the underlying distribution of gene expressions over time, limiting their utility for early cancer detection—the "holy grail" of oncology.

The SurvODE Framework

SurvODE addresses these challenges through a modular architecture designed to handle high-dimensional, noisy, and censored biological data. The framework operates in two distinct phases.

The SurvODE architecture consists of two phases: Phase 1 uses a Cox-PH model to estimate survival probabilities, while Phase 2 integrates a VAE with a Neural ODE module to model continuous latent dynamics.

The core components include:

  1. Variational Autoencoder (VAE): Compresses high-dimensional gene expression data into a low-dimensional latent space, mitigating noise and capturing essential biological features.
  2. Neural ODE Module: Models the evolution of latent representations as a continuous-time process. Unlike discrete neural networks, Neural ODEs can be queried at any arbitrary time point, making them ideal for irregular clinical data.
  3. Cox Proportional Hazards (Cox-PH) Module: Bridges the gap between survival analysis and generative modeling. It estimates "pseudo-survival times" for censored patients, allowing the model to utilize the entire dataset for training.

Experimental Results

The model was evaluated on eight cancer types from TCGA, focusing on the most time-relevant genes. A key test involved "extrapolation": predicting gene distributions for the 20% of patients with the longest survival times (representing early-stage cancer) which were withheld during training.

Comparison of Wasserstein distances across eight TCGA cancer types. SurvODE (red) consistently achieves lower distances compared to standard VAE and CVAE baselines.

As shown in the results, SurvODE significantly outperformed standard VAEs and Conditional VAEs (CVAEs) across nearly all cancer types, including Bladder (BLCA), Breast (BRCA), and Glioma (GBMLGG). Beyond quantitative metrics, the model demonstrated biological relevance. In Glioma, SurvODE identified 16 potential early-stage biomarkers, such as TAGLN2, which has been independently linked to tumorigenesis in existing literature.

Discussion and Conclusion

SurvODE represents a significant step forward in modeling the "arrow of time" in cancer progression. By predicting distributions rather than just mean values, it captures the increasing variance often associated with disease development. While the model currently relies on the proportional hazards assumption of the Cox model, its ability to extrapolate to unobserved early stages offers a powerful tool for biomarker discovery. Future work may extend this framework to model complex gene-gene interactions, further refining our understanding of early-stage oncogenesis.

Summary updated

Abstract

With the increasingly available large-scale cancer genomics datasets, machine learning approaches have played an important role in revealing novel insights into cancer development. Existing methods have shown encouraging performance in identifying genes that are predictive for cancer survival, but are still limited in modeling the distribution over genes. Here, we proposed a novel method that can simulate the gene expression distribution at any given time point, including those that are out of the range of the observed time points. In order to model the irregular time series where each patient is one observation, we integrated a neural ordinary differential equation (neural ODE) with cox regression into our framework. We evaluated our method on eight cancer types on TCGA and observed a substantial improvement over existing approaches. Our visualization results and further analysis indicate how our method can be used to simulate expression at the early cancer stage, offering the possibility for early cancer identification.
BibTeX
@misc{chen2021survode,
  author = {Tong Chen and Sheng Wang},
  title = {{SurvODE: Extrapolating Gene Expression Distribution for Early Cancer Identification}},
  year = {2021},
  eprint = {2111.15080},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2111.15080},
}
Figure