GANs for Genomics

Motivation

Multiple papers have shown that neural networks are generally less effective for prediction of complex genetic disorders than polygenic risk scores (PRSs) even though PRSs are linear predictors [Pinto et al., 2019] [Mamani, 2020] [Badré et al., 2020]. I found this surprising at first but grew to understand why after discovering what challenges arrise when working with genomics datasets for predictive modeling.

1. High-dimensionality feature sets: Input vectors can have on the order of 100,000+ features.
2. Extremely low number of training samples: Deep neural nets need plenty of data, and in the case of genomics, we probably won't have plenty of relevant samples.
3. Datasets are often heavily imbalanced: This is especially true for case-control predictions. Usually, you'll have a small number of cases and lots of controls. The phenotypes tend not to have an even spread.

These challenges make training performant deep learning models an arduous task. I decided to try to alleviate some of the burden of having insufficient training samples.

TLDR: I research the effectiveness of using generative adversarial networks (GANs) to simulate synthetic samples. In doing so, I hope to create more plentiful training data, achieve higher performance on genomic predictions, and reach a greater understanding of the relevant loci in the genome. The central aim of this project is to investigate the viability of neural networks for identifying predictive variants in the genome.

Dataset

Samples: Sprague Dawley rats from GWAS

• sample size: 4106 rats
• Between two laboratories, there were high-quality genotypes at approximately 220,000 SNPs, of which ≈ 90,000 were common between the labs.

Features: Integer-encoded allele frequencies. For example, if A was the reference (most common) allele at some locus, it was represented as 0 and the next most common allele would be represented by 1.

Targets: Groupings from PavCA, or Pavlovian conditioned approach. The targets were also integer-encoded but were based on whether the rats were goal trackers (0), intermediate responders (1), or sign trackers(2).

PavCA: PavCA is a behavior that develops after response-independent presentation of a conditioned stimulus that predicts deliver of an unconditioned stimulus. In layman's terms, this would be something like giving a rat food or drugs (unconditioned stimulus) when it interacts with a lever or bell (conditioned stimulus) and observing what various responses. In general, 3 patterns of conditioned responses develop:

• GT (goal-tracking): Unconditioned stimulus-directed
  • Behavior directed by the goal
  • If the reward stops, the behavior stops
• IR (intermediate response): In between
• ST (sign-tracking): Persistent
  • Persistent motivation by reward-related cues, sometimes despite non-reinforcement
  • Susceptible to addiction-related behaviors

Data source:

• For more information on the data used in this analysis, refer to the following paper:

  Gileta, A. et al. (2018). Genetic characterization of outbred Sprague Dawley rats and utility for genome-wide association studies. bioRxiv, 412924. [paper]

• The data is available for download in variant call format (VCF) on Alex Gileta's Github.

References & Acknowledgements

This was a one-semester research project I worked on in the computer science department at Columbia University.

• PhD Student Mentor: Jie Yuan
• Advisor: Dr. Itsik Pe'er

Any opinions, findings, and conclusions or recommendations expressed in this material are those of the author(s) and do not necessarily reflect the views of Columbia University.

• Pinto, C., Gill, M., & Heron, E. A. (2019). Can artificial neural networks supplant the polygene risk score for risk prediction of complex disorders given very large sample sizes?. arXiv preprint arXiv:1911.08996.
• Mamani, N. M. (2020). Machine Learning techniques and Polygenic Risk Score application to prediction genetic diseases. ADCAIJ: Advances in Distributed Computing and Artificial Intelligence Journal, 9(1), 5-14.
• Badré, A., Zhang, L., Muchero, W., Reynolds, J. C., & Pan, C. (2020). Deep neural network improves the estimation of polygenic risk scores for breast cancer. Journal of Human Genetics, 1-11.
• Yelmen, B., Decelle, A., Ongaro, L., Marnetto, D., Montinaro, F., Furtlehner, C., ... & Jay, F. (2019). Creating Artificial Human Genomes Using Generative Models. [PDF]
• Gileta, A. et al. (2018). Genetic characterization of outbred Sprague Dawley rats and utility for genome-wide association studies. bioRxiv, 412924. [PDF]