Aggregation Tool for Genomic Concepts (ATGC): A deep learning framework for somatic mutations and other sparse genomic measures

Jordan Anaya, John-William Sidhom, Faisal Mahmood, Alexander S. BarasbioRxiv 2020 · DOI

TL;DR

This paper introduces ATGC, a novel deep learning framework that uses attention-based multiple instance learning (MIL) to analyze sparse genomic data like somatic mutations. ATGC can learn to encode, aggregate, and predict from this data in a single end-to-end model, outperforming traditional methods in tasks like cancer type and microsatellite instability classification while also offering improved model explainability.

What Problem Does This Paper Address?

Analyzing large-scale genomic data, such as somatic mutations, presents a significant challenge. While deep learning excels at feature extraction, genomic data often has labels at the sample level (e.g., cancer type) rather than at the individual genomic measure level (e.g., a specific mutation). Traditional methods typically involve separate, static steps for feature engineering and aggregation before feeding the data into a predictive model. This approach can lead to suboptimal results because it doesn't dynamically learn which genomic features are most important for a given task. Furthermore, the high dimensionality of genomic data (the "curse of dimensionality") makes it sparse and prone to overfitting, and existing dimension reduction techniques might discard valuable information. The paper aims to bridge this gap by developing a unified, end-to-end deep learning framework that can learn to represent, aggregate, and predict from sparse genomic data, while also providing insights into why it makes certain predictions.

Methods

The core of ATGC is an attention-based Multiple Instance Learning (MIL) framework. In MIL, a "bag" (in this case, a sample's set of genomic measures) is labeled, but individual "instances" (specific mutations) within the bag are not. ATGC addresses this by:

  • Trainable Encoders: Instead of relying on pre-defined features, ATGC uses trainable encoders (e.g., a sequence encoder for mutation context, a position encoder) to learn meaningful representations of individual genomic instances.
  • Attention Mechanism: A key innovation is the use of an attention mechanism that dynamically weights the importance of each instance within a sample. This allows the model to focus on the most salient mutations for a given task.
  • End-to-End Learning: The entire process—encoding, attention-based aggregation, and prediction—is performed within a single, trainable model. This contrasts with traditional methods where these steps are separate and static.
  • Aggregation Functions: ATGC offers flexible aggregation functions, including sum, mean, and a novel dynamic aggregation that allows for sample-level information to influence instance weighting.
  • Custom Activation Functions: The paper introduces adaptive square root (ASR), adaptive rectifying unit (ARU), and adaptive sigmoid unit (ASU) functions designed to handle the numerical characteristics of genomic data and avoid overflow issues.

Datasets

The study utilizes several publicly available datasets:

  • TCGA MC3 public MAF: This dataset of exome-based somatic mutations was used for tumor classification and MSI status prediction.
  • TCGA controlled MAF: Used for MSI classification.
  • PCAWG (Pan-Cancer Analysis of Whole Genomes) data: Used for tumor classification.
  • Synthetic tasks: Used to demonstrate the framework's superiority over conventional approaches.

The paper also mentions the use of MSI PCR labels and UCSC simpleRepeat.txt for MSI analysis.

Key Insights and Ideas

  • Attention is Crucial for Sparse Genomic Data: The attention mechanism is not just a nice-to-have; it's essential for identifying the truly informative mutations within a large, noisy set of genomic alterations.
  • End-to-End Learning Unlocks Performance: By integrating feature learning, aggregation, and prediction, ATGC can discover task-specific representations that static, pre-engineered features miss.
  • Model Explainability is a Direct Output: The attention weights provide a direct mechanism to understand which genomic features (mutations, their context, or location) are driving the model's predictions, which is critical for clinical adoption.
  • Genomic Location Matters, But Genes are More Informative: While genomic bins showed some predictive power, using gene-level information as input further improved performance, suggesting that understanding the biological context of mutations is key.
  • Learning Signatures of Cancer Biology: The framework can uncover known cancer biology, such as UV radiation signatures in SKCM or mismatch repair deficiency signatures in COAD, by analyzing attention patterns and learned representations.

Key Results

  • Tumor Classification: ATGC achieved best-in-class performance on tumor classification tasks using TCGA exome data, outperforming logistic regression, random forests, and neural nets. Using 6 bp windows as input provided a significant ~14% increase over the best standard model.
  • Microsatellite Instability (MSI) Prediction: ATGC slightly outperformed state-of-the-art methods (MANTIS and MSIpred) for MSI status prediction, despite not being explicitly trained on simple repeat information, which is a key feature of MSI.
  • Model Explainability: The attention mechanism allowed for the visualization of which genomic regions or mutation types were most important for classification, revealing biologically relevant patterns. For example, in SKCM, attention clusters corresponded to known UV signatures.
  • Synthetic Tasks: ATGC successfully solved synthetic tasks where conventional approaches failed, demonstrating its robustness.

Limitations and Future Work

  • Limitations:
    • The paper acknowledges that their MIL approach is "fairly standard" and may not represent the "full potential" of applying MIL to genomics.
  • While the model can identify important instances, it doesn't inherently distinguish between instances that are important due to their biological significance versus those that are simply rare.
  • The current instance featurization is limited to basic mutation properties (chromosome, position, alleles), and incorporating outside knowledge (e.g., variant consequence, pathways) could further enhance performance.
  • Future Work:
    • The authors suggest incorporating clustering steps at the instance level, similar to recent work in computational pathology.
  • They propose integrating external biological knowledge into the variant encoding process.
  • The potential for combining predictions from multiple models (e.g., ATGC with MANTIS for MSI prediction) is highlighted.
  • The framework's modularity allows for easy extension to other genomic data types and tasks.

Summary updated

Abstract

ABSTRACT Deep learning can extract meaningful features from data given enough training examples. Large-scale genomic data are well suited for this class of machine learning algorithms; however, for many of these data the labels are at the level of the sample instead of at the level of the individual genomic measures. Conventional approaches to this data statically featurise and aggregate the measures separately from prediction. We propose to featurise, aggregate, and predict with a single trainable end-to-end model by turning to attention-based multiple instance learning. This allows for direct modelling of instance importance to sample-level classification in addition to trainable encoding strategies of genomic descriptions, such as mutations. We first demonstrate this approach by successfully solving synthetic tasks conventional approaches fail. Subsequently we applied the approach to somatic variants and achieved best-in-class performance when classifying tumour type or microsatellite status, while simultaneously providing an improved level of model explainability. Our results suggest that this framework could lead to biological insights and improve performance on tasks that aggregate information from sets of genomic data.
BibTeX
@misc{anaya2020aggregation,
  author = {Jordan Anaya and John-William Sidhom and Faisal Mahmood and Alexander S. Baras},
  title = {{Aggregation Tool for Genomic Concepts (ATGC): A deep learning framework for somatic mutations and other sparse genomic measures}},
  year = {2020},
  doi = {10.1101/2020.08.05.237206},
  howpublished = {bioRxiv},
}
Figure