Transfer learning enables prediction of CYP2D6 haplotype function

Gregory McInnes, Rachel Dalton, Katrin Sangkuhl, Michelle Whirl‐Carrillo, Seung‐been Lee, Philip S. Tsao, Andrea Gaedigk, Russ B. Altman, Erica L. Woodahl

TL;DR

Researchers have developed Hubble.2D6, a deep learning model that predicts the functional status of CYP2D6 gene variants with 88% accuracy. This tool can help assign function to previously uncharacterized gene variants, improving personalized medicine by predicting drug metabolism and potential adverse reactions.

What Problem Does This Paper Address?

The CYP2D6 gene plays a crucial role in metabolizing over 20% of commonly prescribed drugs. Variations in this gene lead to significant differences in how individuals respond to medications, potentially causing toxicity or rendering treatments ineffective. Predicting a person's CYP2D6 "phenotype" (how well they metabolize drugs) relies on understanding the function of different gene variants, known as haplotypes. However, as new haplotypes are discovered through large-scale sequencing, manually assigning their function based on existing literature and experimental data is becoming increasingly challenging and time-consuming. This gap means many individuals may have their CYP2D6 status labeled as "indeterminate," hindering personalized pharmacogenetic guidance.

Methods

The authors developed a deep learning model called Hubble.2D6. They employed a transfer learning approach, where a neural network was first pre-trained on a large dataset (simulated and real data) to learn general patterns. This pre-trained network was then adapted to the specific task of predicting CYP2D6 haplotype function.

The model uses a convolutional neural network (CNN) architecture, specifically a three-layer CNN inspired by the Basset architecture. The input to the model is the DNA sequence of a CYP2D6 haplotype, represented as a one-hot encoded matrix. Crucially, this sequence data is augmented with eight functional variant annotations. These annotations provide context about each genetic position, including whether it's in a coding region, its population frequency, predicted deleteriousness, and other functional characteristics.

The model outputs two scores: the probability of being a "no function" allele and the probability of being a "normal function" allele. These scores are then combined using cutoffs to classify the haplotype into one of three functional statuses: normal, decreased, or no function. This approach frames the problem as an ordinal regression, allowing the model to learn the relationships between these functional categories.

Datasets

  • Simulated Diplotypes: Used for initial pre-training. 50,000 simulated diplotypes (pairs of haplotypes) were generated based on known CYP2D6 star alleles with curated functions, with additional diversity introduced by sampling alternate alleles from gnomAD.
  • Liver Microsome Data: Used for a second pre-training step. Activity data from 314 liver microsome samples, measured using dextromethorphan as a substrate, were utilized.
  • Star Allele Data: Used for training and validation. Sequences and functions for 31 star alleles with curated function were used for training, and 25 different star alleles with curated function were used for validation. An additional 71 star alleles with uncurated function were used for prediction.
  • In vitro data from a prior study: Used for further validation. This dataset included functional characterization of 46 CYP2D6 star alleles, with 30 having curated function and 16 lacking it.

Key Insights and Ideas

  • Transfer learning is key for small datasets: The study demonstrates that transfer learning is highly effective for training accurate models in genetics, where curated datasets for specific genes like CYP2D6 can be relatively small. Pre-training on larger, related datasets allows the model to learn generalizable features, significantly improving performance.
  • Functional annotations provide crucial context: Simply using raw DNA sequence wasn't enough. Augmenting the sequence with functional annotations (like deleteriousness or coding region status) dramatically improved the model's accuracy, highlighting the importance of biological context in genetic variant interpretation.
  • Deep learning can bridge the "indeterminate" gap: The model successfully predicts function for previously uncharacterized CYP2D6 haplotypes, offering a way to assign a functional status where manual curation is lacking. This has direct implications for clinical decision-making.
  • Model prioritizes "loss-of-function" variants conservatively: The model tends to predict "decreased function" rather than "no function" for variants that are strongly predicted to cause loss-of-function. While this might seem counterintuitive, it suggests a conservative approach that still flags abnormal drug metabolism, which can be beneficial.

Key Results

  • Hubble.2D6 achieved 88.0% accuracy on a held-out validation set of CYP2D6 star alleles with curated function.
  • The model's predictions for uncurated CYP2D6 haplotypes explained a significant amount of the variability in in vitro functional measurements, indicating good correlation with actual function.
  • For curated alleles used in evaluation, Hubble.2D6's predicted labels explained 71% of the variance in metabolic activity, comparable to labels assigned by CPIC.
  • For uncurated alleles, the model's predictions explained 47.5% of the variance in measured in vitro activity, with significantly different mean activities across predicted phenotype groups.
  • The study found that the annotation indicating whether a variant was rare in the population most improved classification accuracy (68% test accuracy), followed by predicting deleteriousness (52% test accuracy).

Limitations and Future Work

  • Exclusion of Structural Variation: The current model does not account for structural variations like gene duplications, deletions, or hybrid genes (e.g., CYP2D7-CYP2D6). This means it cannot predict increased function alleles, which are primarily due to gene duplications.
  • Limited Sequence Window: The model considers a narrow range of upstream and downstream sequences, potentially missing distal regulatory effects that could influence gene expression.
  • Single Expression System for Validation: The in vitro validation data was generated using a single expression system in one laboratory, which may not fully capture inter-laboratory variability in functional measurements.
  • Focus on Genetic Variation: The model only considers genetic variation within the CYP2D6 locus and does not account for external factors like drug interactions (e.g., with SSRIs) or variations in other genes that can affect CYP2D6 activity.

Future work could involve incorporating structural variation, expanding the sequence window, and integrating external factors to create a more comprehensive predictive model. Further validation with diverse in vitro data and more training data, especially for loss-of-function variants, could also improve future versions of Hubble.2D6.

Summary updated

Abstract

Abstract Cytochrome P450 2D6 ( CYP2D6 ) is a highly polymorphic gene whose protein product metabolizes more than 20% of clinically used drugs. Genetic variations in CYP2D6 are responsible for interindividual heterogeneity in drug response that can lead to drug toxicity and ineffective treatment, making CYP2D6 one of the most important pharmacogenes. Prediction of CYP2D6 phenotype relies on curation of literature-derived functional studies to assign a functional status to CYP2D6 haplotypes. As the number of large-scale sequencing efforts grows, new haplotypes continue to be discovered, and assignment of function is challenging to maintain. To address this challenge, we have trained a deep learning model to predict functional status of CYP2D6 haplotypes, called Hubble.2D6. We find that Hubble.2D6 predicts CYP2D6 haplotype functional status with 88% accuracy in a held out test set and explains a significant amount of the variability in in vitro functional data. Hubble.2D6 may be a useful tool for assigning function to haplotypes with uncurated function, which may be used for screening individuals who are at risk of being poor metabolizers.
BibTeX
@misc{mcinnes2019transfer,
  author = {Gregory McInnes and Rachel Dalton and Katrin Sangkuhl and Michelle Whirl‐Carrillo and Seung‐been Lee and Philip S. Tsao and Andrea Gaedigk and Russ B. Altman and Erica L. Woodahl},
  title = {{Transfer learning enables prediction of <i>CYP2D6</i> haplotype function}},
  year = {2019},
  doi = {10.1101/684357},
  howpublished = {bioRxiv},
}
Figure