Transfer learning enables prediction ofCYP2D6haplotype function

Abstract

A major limitation of phenotype prediction in genetics is the ability to model the complexities of genetic variation when sample sizes are small. This is especially true in pharmacogenetics, ahighly translational yet data-limited subfield of genetics. Drug metabolism is a critical facet of pharmacogenetics and can have consequences for drug safety and efficacy. CYP2D6 is an 1 important enzyme, metabolizing more than 25% of clinically used drugs. It is highly polymorphic which leads to a heterogeneous response to drugs among the population.We presentHubble2D6, a set of deep learning models for predicting metabolic activity of CYP2D6 genotype and predicting functional classification of CYP2D6 haplotypes. We train our models on 249 samples, addressing data scarcity by pretraining on simulated data, weakly supervised learning, and using a functional representation of genetic variants. We validate our models using in vitro data for haplotypes previously unseen by the model and explain 38% of the variance with the genotype-based activity predictor and predict haplotype function with an AUC of 0.85. We demonstrate a procedure to build a computational model of a complex gene using primarily simulated and unlabeled data which can then be used to make functional predictions about novel genetic variation, and present a model that may be of clinical significance for an important application of genetics.

Pharmacogenetics at scale: An analysis of the UK Biobank

“…An alternative to defining individual haplotypes and/or driving mutations, is to take a top down approach, in which regions of the gene or genes themselves are deemed essential, and any deleterious variant within essential components can be assigned an inferred phenotype. Recent work on the development of data-driven PGx phenotyping methods indicates that given enough data, it might be possible to move away from variant level rule-based systems and towards data-driven machine-learning models capable of robustly handling unobserved genetic variation 13,24,25 . The challenges posed by rare variation is likely to be a consistent issue for the current PGx system and will likely grow over time as genotyping gives way to genome sequencing and more populations are studied in detail revealing more rare and/or private mutations harbored by individuals.…”

Variation in Actionable Pharmacogenetic Markers in Natives and Mestizos From Mexico

“…FATHMM was utilized to predict functional consequences of non-coding variants, it integrates functional annotations from ENCODE and calculates a score (0–1). Scores >0.5 are indicative of deleterious variation (McInnes et al, 2019; Zhou et al, 2018), and although non-coding variants may not have a direct impact on a protein the overall regulation may render faulty. These tools have been utilized in PGx research (Devarajan et al, 2019), and we have reported their use to filter variants with potential effects on atorvastatin pharmacokinetics with acceptable results.…”

A Review of the Important Role of CYP2D6 in Pharmacogenomics

“…This high-throughput functional screening approach has been successfully applied to NUDT15 and could be conceivably applied to CYP2D6 [111]. Analysis of large datasets is another approach to assign functional effects to these undetermined haplotypes, and is being increasingly carried out to interrogate CYP2D6 [112,113]. One example is the use of sequence data and phenotypic data from >450,000 UK biobank participants [112].…”

“…Analysis of large datasets is another approach to assign functional effects to these undetermined haplotypes, and is being increasingly carried out to interrogate CYP2D6 [112,113]. One example is the use of sequence data and phenotypic data from >450,000 UK biobank participants [112]. Analysis of these data has been used to elucidate haplotypes, but it was unable to capture structural variations.…”

“…This technique was able to detect UM individuals with sensitivity and specificity values of 85-87%, and PM individuals with sensitivity and specificity values of 71-79% [115]. Another recent study utilised a deep learning approach to predict the function of CYP2D6 genotypes, including those with rare variants, and then derive a subsequent classification model that predicts whether a CYP2D6 haplotype has a normal or reduced function, and was shown to have a high degree of accuracy [116]. Lastly, a neural network model was recently developed and trained on long-read CYP2D6 sequencing data in conjunction with information on the rate of CYP2D6-mediated tamoxifen metabolism from 561 patients with breast cancer.…”