MoTERNN: Classifying the Mode of Cancer Evolution Using Recursive Neural Networks

Mohammadamin Edrisi, Huw A. Ogilvie, Meng Li, Luay NakhlehbioRxiv 2022 · DOI

TL;DR

This paper introduces MoTERNN, a novel deep learning model that uses recursive neural networks to classify the evolutionary mode of cancer based on phylogenetic trees derived from single-cell DNA sequencing data. MoTERNN achieves high accuracy in distinguishing between linear, branching, neutral, and punctuated cancer evolution, offering a powerful new tool for understanding tumor progression and informing clinical decisions.

What Problem Does This Paper Address?

Cancer is not a static disease; it evolves over time. Understanding this evolution, particularly the "mode" of evolution (how cancer cells diversify and adapt), is crucial for diagnosis, prognosis, and developing effective treatments. This evolutionary history is often represented as a phylogenetic tree, showing the relationships between different cancer cell populations. However, inferring and classifying these evolutionary modes from complex, high-dimensional single-cell sequencing data is a significant challenge. Existing methods often rely on simulations or mathematical models that can be computationally intensive or may not fully capture the nuances of tree structures. This paper aims to bridge this gap by developing a machine learning approach that can directly learn from and classify these evolutionary trees.

Methods

The core of this paper is the development of MoTERNN (Mode of Tumor Evolution using Recursive Neural Networks). This model leverages Recursive Neural Networks (RNNs), a type of neural network particularly well-suited for processing hierarchical or tree-like data structures.

Here's a breakdown of the key components:

  • Encoder Network: This network takes the genotype profile of each individual cell (represented as a binary vector of mutations) and maps it into a lower-dimensional embedding. This allows the model to represent complex genetic information in a more manageable form.
  • Compositionality Function: This is the heart of the recursive nature of the network. It takes the embeddings of child nodes in the phylogenetic tree and combines them to compute the embedding of their parent node. This process is applied recursively up the tree, effectively building a representation of the entire tree structure.
  • Classifier Network: Once the root of the phylogenetic tree has an embedding, this network takes it as input and outputs probability scores for each of the four evolutionary modes (linear, branching, neutral, punctuated). The mode with the highest probability is assigned as the predicted evolutionary mode for the tumor.

The model is trained in a supervised fashion using simulated data and then applied to real-world data.

Datasets

  • Simulated Data: The authors generated a large dataset of 8100 simulated phylogenetic trees and corresponding genotype matrices, with 2025 data points for each of the four evolutionary modes (linear, branching, neutral, punctuated). This dataset was created using a modified beta-splitting model (BSM) to generate trees with varying topologies and branch lengths, and an infinite-sites assumption (ISA) model for placing mutations.
  • Real Data: A single-cell DNA sequencing dataset from a triple-negative breast cancer (TNBC) patient was used for evaluation. This dataset included diploid, aneuploid, and hypodiploid cells, and a phylogeny was inferred from single-nucleotide variations (SNVs) using the Phylovar tool.

Key Insights and Ideas

  • RNNs are well-suited for phylogenetic tree analysis: The paper demonstrates that the inherent recursive structure of RNNs makes them a natural fit for processing and classifying phylogenetic trees, outperforming traditional methods that might struggle with variable tree sizes and complex relationships.
  • Capturing evolutionary modes is feasible with deep learning: MoTERNN successfully learns to distinguish between distinct evolutionary patterns (linear, branching, neutral, punctuated) based on tree topology and genetic data, suggesting that these modes leave discernible signatures in the evolutionary history.
  • Simulated data can effectively train models for real-world biological problems: The high accuracy achieved on both simulated and real data highlights the power of well-designed simulations in training robust machine learning models for complex biological questions.
  • SNV data can infer evolutionary modes similar to CNA data: The study found that MoTERNN, trained on SNV data, classified a TNBC patient's tumor evolution as punctuated, aligning with a previous study that used Copy Number Aberration (CNA) data, suggesting that different types of genetic data can converge on similar evolutionary conclusions.

Key Results

  • High Accuracy: MoTERNN achieved an average training accuracy of 99.98% and a test accuracy of 99.95% on the simulated datasets during cross-validation.
  • Real-world application: When applied to the TNBC patient data, MoTERNN predicted punctuated evolution as the most likely mode, with an association probability of approximately 0.7501. This result is in concordance with the original study that used CNA data.
  • Computational Efficiency: The training of MoTERNN on the simulated data was relatively fast, with an average training time of about five minutes and 14 seconds per cross-validation round on a powerful GPU.

Limitations and Future Work

  • Sampling Bias: The authors acknowledge that sampling bias in the collection of single-cell data can impact the inferred tree and, consequently, the model's performance. This is an ongoing challenge in phylogenetics.
  • Infinite-Sites Assumption (ISA): The simulation of mutations relies on the ISA, which assumes that each mutation occurs at a unique site and is never lost. This assumption may be violated in real cancer evolution, where back mutations or complex mutation patterns can occur.
  • Future Directions:
    • More Complex Evolutionary Models: The authors suggest exploring more sophisticated models for simulating phylogenetic trees, such as multi-type branching evolution processes, and incorporating other types of genetic data like copy number aberrations (CNAs) or even ternary genotypes.
  • Advanced Architectures: While MoTERNN is efficient, further complexity in the encoder, compositionality function, and classifier could potentially improve performance on more intricate datasets.
  • Broader Applications: The authors believe that variations of RNN models can be valuable for future studies in evolutionary biology beyond cancer.

Summary updated

Abstract

Abstract With the advent of single-cell DNA sequencing, it is now possible to infer the evolutionary history of thousands of tumor cells obtained from a single patient. This evolutionary history, which takes the shape of a tree, reveals the mode of evolution of the specific cancer under study and, in turn, helps with clinical diagnosis, prognosis, and therapeutic treatment. In this study we focus on the question of determining the mode of evolution of tumor cells from their inferred evolutionary history. In particular, we employ recursive neural networks that capture tree structures to classify the evolutionary history of tumor cells into one of four modes— linear , branching , neutral , and punctuated . We trained our model, MoTERNN, using simulated data in a supervised fashion and applied it to a real phylogenetic tree obtained from single-cell DNA sequencing data. MoTERNN is implemented in Python and is publicly available at https://github.com/NakhlehLab/MoTERNN .
BibTeX
@misc{edrisi2022moternn,
  author = {Mohammadamin Edrisi and Huw A. Ogilvie and Meng Li and Luay Nakhleh},
  title = {{MoTERNN: Classifying the Mode of Cancer Evolution Using Recursive Neural Networks}},
  year = {2022},
  doi = {10.1101/2022.08.21.504710},
  howpublished = {bioRxiv},
}

Loading related content...

Figure