Deep representation learning for temporal inference in cancer omics: a systematic review

This systematic review examines the application of deep representation learning, particularly Variational Autoencoders, in cancer omics, highlighting their current dominance in subtyping and prognosis while identifying the scarcity of longitudinal data as a major barrier to modeling cancer's temporal dynamics and proposing the use of VAEs as generative models to advance time-based cancer staging.

The Gist

The Big Picture: Cancer is a Movie, but We Only Have Snapshots

Imagine cancer isn't just a static lump; it's a movie. It starts small, changes, grows, and evolves over time. To truly understand how to stop it, we need to watch the whole film.

However, most medical data we have today is like a photo album. We have a photo of a patient when they are diagnosed (the beginning), maybe one when they get sick again (the middle), and one at the end. We don't have the video recording of the movie in between.

This paper is a systematic review (a big, organized search) of how scientists are using a specific type of Artificial Intelligence called Deep Representation Learning (DRL), and specifically a tool called the Variational Autoencoder (VAE), to try to turn those scattered photos into a continuous movie.

The Star of the Show: The Variational Autoencoder (VAE)

Think of a VAE as a super-smart translator and a creative artist rolled into one.

1. The Translator (The Encoder): Cancer data (like gene sequences) is messy, huge, and confusing. It's like a library with millions of books written in a language no one understands. The VAE's "Encoder" reads all those books and summarizes the story into a single, simple sentence. This sentence is called the Latent Space. It captures the essence of the cancer without all the noise.
2. The Artist (The Decoder): The VAE's "Decoder" can take that simple sentence and try to rewrite the whole story (reconstruct the original data). But here's the magic: because it understands the rules of the story, it can also imagine new scenes. It can write a chapter that doesn't exist yet.

What the Review Found: The Good, The Bad, and The Missing

The authors looked at 440 papers and narrowed them down to 21 that were truly relevant. Here is what they discovered:

1. The Current Focus: Taking Photos, Not Filming Movies
Most scientists are using these AI tools to take better "photos." They use the VAE to:

• Sort patients: "You have Type A cancer, you have Type B." (Subtyping)
• Diagnose: "Is this a tumor or a mole?" (Diagnosis)
• Predict the future: "Based on this photo, how long will the patient live?" (Prognosis)

The Problem: They are mostly looking at a single moment in time. They aren't really using the AI to watch the cancer move or evolve.

2. The Missing Ingredient: Longitudinal Data
To make a movie, you need a camera that keeps rolling. In medicine, this is called longitudinal data (taking samples from the same patient over and over again).

• Why is this hard? Cancer cells are tiny. To test them, doctors often have to destroy the sample (like eating a cookie to see what's inside). You can't eat the same cookie twice.
• The Result: We have thousands of photos of different people at different times, but very few "time-lapse" videos of the same person.

3. The Workaround: Using "Stages" as a Proxy
Since we can't film the movie, scientists are trying to stitch together photos from different people to guess the plot. They use Cancer Stages (Stage 1, Stage 2, Stage 3) as a stand-in for time.

• The Analogy: Imagine trying to figure out how a human grows from a baby to an adult. You don't have a video of one person. Instead, you take a photo of a baby, a photo of a toddler, and a photo of a teenager from three different families. You line them up and say, "Okay, this must be the order of growth."
• The Catch: Not every baby grows at the same speed. Not every cancer grows at the same speed. Lining them up perfectly is very difficult.

4. The New Frontier: Single-Cell "Pseudo-Time"
The most exciting part of the review is about Single-Cell Data. Instead of looking at a whole tissue sample (a smoothie), scientists are looking at individual cells (the fruit pieces).

• By looking at thousands of cells at once, the AI can guess which cells are "younger" and which are "older" based on their genetic makeup.
• This creates a Pseudo-Time trajectory. It's like looking at a crowd of people and guessing who is walking toward the exit and who just entered, even though you didn't see them move.
• This is currently the most popular way to study cancer "progression," but it's still an estimate, not a real-time video.

The Big Idea: Why This Matters

The authors argue that we need to stop just using AI to classify (sort) cancer and start using it to simulate (predict) cancer.

The Proposal:
Imagine the VAE as a Time Machine.
If we train the AI on the "photos" we have (Stage 1, Stage 2, Stage 3), we can ask it to generate the missing chapters.

• "If a patient is at Stage 2, what will their cancer look like in 6 months?"
• "If we give them this drug, how does the movie change?"

This would allow doctors to test treatments in a virtual simulation before trying them on a real person.

The Hurdles Ahead

The paper ends with a reality check. To make this work, we need:

1. Better Data: We need more "time-lapse" videos (longitudinal data) to teach the AI the rules of the movie.
2. Better Validation: We need to make sure the AI isn't just "hallucinating" (making up fake biology). If the AI invents a cancer progression that doesn't exist, it could be dangerous.
3. Ethics: We need to ensure these AI models don't learn biases from the data (e.g., only learning from one type of patient).

Summary in One Sentence

This paper says that while AI is great at taking "snapshots" of cancer to diagnose it, we need to teach it how to "film the movie" of cancer progression, but we are currently held back by a lack of continuous data and the difficulty of tracking how cancer changes over time in real patients.

Technical Summary

1. Problem Statement

Cancer is a highly complex, dynamic, and heterogeneous disease. While Deep Learning (DL) and Deep Representation Learning (DRL) have shown remarkable success in analyzing high-dimensional omics data (genomics, transcriptomics, proteomics, etc.) for static tasks like subtyping, diagnosis, and prognosis, they often fail to capture the temporal dimension of cancer progression.

Key challenges identified include:

• Lack of Longitudinal Data: Clinical omics data is typically cross-sectional (single time point). True longitudinal data (repeated measurements of the same patient over time) is scarce due to ethical constraints, the destructive nature of sequencing assays, and logistical difficulties in patient retention.
• Misalignment: Even when temporal data exists, samples are often unaligned across patients because cancer progression rates vary individually.
• Limited Temporal Modeling: Most DRL applications treat cancer as a static state rather than a dynamic process, missing evolutionary dynamics and trajectory inference.

2. Methodology

The authors conducted a Systematic Literature Review (SLR) following PRISMA guidelines to investigate the application of DRL, with a specific focus on Variational Autoencoders (VAEs), in cancer omics studies involving time-related processes.

• Search Strategy: Queries were executed across Google Scholar, Scopus, and Web of Science (2014–2024) using keywords related to deep learning, representation learning, VAEs, omics, cancer, and temporal concepts (progression, evolution, trajectory, longitudinal).
• Selection Criteria:
  • Inclusion: Peer-reviewed papers using DRL/VAEs on cancer omics data with explicit or implicit temporal modeling (e.g., longitudinal data, cancer stages, pseudo-time).
  • Exclusion: Reviews, non-English papers, studies focusing solely on survival analysis without biological insights, and papers lacking temporal characterization.
• Data Extraction: From an initial pool of 440 papers, 44 were selected for full-text review, resulting in 21 papers included in the final analysis. These were categorized based on their "Temporal Proxy" (Actual Time, Stages, or Pseudo-time).

3. Key Contributions

The paper provides a comprehensive mapping of the current landscape and identifies critical gaps:

1. Categorization of Temporal Approaches: The review classifies existing studies into three main strategies for handling time:
   • Actual Time Series: Rare in cancer; mostly found in non-cancer diseases (e.g., Alzheimer's, Cardiovascular) or specific liquid biopsy studies.
   • Clinical Stages as Proxies: Using TNM staging (0–IV) as a surrogate for time to align cross-sectional data.
   • Pseudo-time Inference: Using single-cell data to order cells along a differentiation trajectory based on transcriptional similarity (e.g., Monocle, PAGA).
2. Analysis of Model Architectures: The authors compare various DRL models:
   • VAEs (and variants like CVAE, ConvVAE, GraphVAE): Highlighted for their ability to learn probabilistic latent spaces, generate synthetic data, and interpolate between states. They are the most versatile for temporal omics.
   • RNNs (LSTM/GRU) & Transformers: Effective for explicit sequential data but less common in cancer omics due to data scarcity.
   • Standard Neural Networks: Primarily used for classification, lacking generative or temporal interpolation capabilities.
3. Identification of Data Limitations: The review underscores that single-cell omics (especially Spatial Transcriptomics) is the dominant data source for temporal inference, while bulk omics and true longitudinal patient data remain underutilized.

4. Key Results

• Dominance of Static Tasks: The vast majority of DRL applications in cancer focus on subtyping, diagnosis, and prognosis. Explicit modeling of temporal progression is rare.
• The "Stage" Proxy: When longitudinal data is unavailable, researchers often use cancer stages (e.g., Stage I vs. Stage IV) as a time unit. However, these studies often struggle with class separability, leading to simplified groupings (e.g., "Early" vs. "Late" stages).
• Generative Potential: VAEs are uniquely positioned to address data scarcity. They can generate synthetic samples to balance unbalanced stage distributions (e.g., generating synthetic Stage II data from Stage I and III) and infer trajectories between stages.
• Single-Cell & Spatial Data: The most sophisticated temporal inferences are currently found in single-cell studies using pseudo-time algorithms. Recent trends show a surge in combining Spatial Transcriptomics (ST) with DRL to map cell states in both spatial and pseudo-temporal dimensions.
• Validation Gap: There is a significant lack of rigorous validation frameworks for synthetic temporal data. Most studies rely on internal metrics rather than external biological validation of the generated trajectories.

5. Significance and Future Directions

This review highlights a critical bottleneck in cancer research: the inability to effectively model disease evolution due to data and methodological limitations.

• Proposed Solution: The authors advocate for leveraging VAEs as generative models to synthesize temporally aligned instances across cancer stages. This could allow researchers to "fill in the gaps" of missing longitudinal data, enabling the study of progression trajectories without requiring destructive, repeated sampling of patients.
• Clinical Impact: Moving from static snapshots to dynamic models could revolutionize personalized medicine by predicting individual patient progression paths, optimizing treatment timing, and understanding resistance mechanisms.
• Future Requirements:
  • Development of standardized validation frameworks for synthetic omics data (focusing on fidelity, utility, and privacy).
  • Creation of open, curated longitudinal datasets (e.g., leveraging resources like TRACERx or PDX models).
  • Integration of explainable AI (XAI) to ensure generative models do not produce biologically implausible "hallucinations" and to build clinical trust.

In conclusion, while DRL has transformed cancer omics analysis, its potential to model the temporal dynamics of cancer remains largely untapped. The paper calls for a paradigm shift toward generative, time-aware models to bridge the gap between static genomic data and the dynamic reality of cancer progression.