Quantifying unobserved protein-coding variants in human populations provides a roadmap for large-scale sequencing projects

James Zou, Gregory Valiant, Paul Valiant, Konrad J. Karczewski, Siu On Chan, Kaitlin E. Samocha, Monkol Lek, Shamil Sunyaev, Mark J. Daly, Daniel G. MacArthur

Abstract

As new proposals aim to sequence ever larger collection of humans, it is critical to have a quantitative framework to evaluate the statistical power of these projects. We developed a new algorithm, UnseenEst, and applied it to the exomes of 60,706 individuals to estimate the frequency distribution of all protein-coding variants, including rare variants that have not been observed yet in the current cohorts. Our results quantified the number of new variants that we expect to identify as sequencing cohorts reach hundreds of thousands of individuals. With 500K individuals, we find that we expect to capture 7.5% of all possible loss-of-function variants and 12% of all possible missense variants. We also estimate that 2,900 genes have loss-of-function frequency of <0.00001 in healthy humans, consistent with very strong intolerance to gene inactivation.

Introduction

As genomic sequencing scales to hundreds of thousands of individuals in initiatives like the UK Biobank or the U.S. All of Us Research Program, researchers face a "dark matter" problem: how many rare genetic variants remain undiscovered? While common genetic variants are well-mapped, the most medically relevant variants—those that break genes (Loss-of-Function, or LoF) or change protein shapes (missense)—are often vanishingly rare. This study introduces UnseenEst, a machine learning algorithm designed to predict the frequency of these unobserved variants, providing a quantitative roadmap for future large-scale sequencing projects.

Methodology

The core of this research is UnseenEst, an algorithm based on linear programming that addresses the classic statistical challenge of "estimating the unseen."

The UnseenEst Algorithm

Unlike previous Bayesian or "jackknife" estimators that rely on rigid evolutionary assumptions, UnseenEst takes the observed Site-Frequency Spectrum (SFS)—a histogram of variant occurrences—and solves an optimization problem. It identifies a frequency distribution that matches the observed data while remaining mathematically plausible, allowing for extrapolation to sample sizes more than 10-fold larger than the input.

Data and Validation

The model was primarily trained on the Exome Aggregation Consortium (ExAC) dataset, comprising exome sequences from 60,706 individuals. To ensure relevance to U.S. health initiatives, a sub-cohort of 33,778 individuals was weighted to match the 2010 U.S. Census ancestral breakdown. Validation involved training the model on a 10% data slice and successfully predicting the variants in the remaining 90%, where it significantly outperformed standard estimators.

Key Results and Insights

The application of UnseenEst reveals that we are still far from "saturation" in our understanding of human genetic variation.

Projections of variant discovery and gene-level analysis using UnseenEst.

Projections for Large Cohorts

In a cohort of 500,000 individuals, the model predicts the identification of approximately 400,000 distinct LoF variants, 7.5 million missense variants, and 3.4 million synonymous variants. Despite these large numbers, this represents only 7.5% of all possible LoF variants and 12% of missense variants, highlighting the immense "long tail" of human variation.

Mutation Rates and Discovery

Discovery rates are heavily influenced by mutation types. Variants at CpG sites, which are chemical hotspots for mutations, are discovered rapidly; nearly 50% have already been identified. Conversely, "transversions" (rarer chemical swaps) are discovered much more slowly.

Identifying "Human Knockouts"

The model serves as a guide for finding "human knockouts"—healthy individuals missing specific genes. These individuals are invaluable for drug discovery, as they demonstrate the safety of inhibiting specific gene products. At a sample size of 250,000, the study expects to find ~14,000 genes where at least 10 different individuals carry an LoF mutation.

Evolutionary Selection

The frequency distribution also provides a footprint of natural selection. The authors identified approximately 2,900 genes where LoF variants are extremely rare (<1 in 100,000), suggesting these genes are essential and the human body cannot tolerate their inactivation.

Limitations and Future Outlook

While UnseenEst provides a powerful predictive framework, it has limitations. Confidence intervals widen significantly for projections beyond 500,000 individuals. Additionally, the current study focuses exclusively on Single-Nucleotide Variants (SNVs), excluding insertions, deletions, and structural variations. Future work will need to account for these more complex variants and potential sampling biases in disease-specific cohorts to refine the roadmap for genomic discovery.

Summary updated

Abstract

As new proposals aim to sequence ever larger collection of humans, it is critical to have a quantitative framework to evaluate the statistical power of these projects. We developed a new algorithm, UnseenEst, and applied it to the exomes of 60,706 individuals to estimate the frequency distribution of all protein-coding variants, including rare variants that have not been observed yet in the current cohorts. Our results quantified the number of new variants that we expect to identify as sequencing cohorts reach hundreds of thousands of individuals. With 500K individuals, we find that we expect to capture 7.5% of all possible loss-of-function variants and 12% of all possible missense variants. We also estimate that 2,900 genes have loss-of-function frequency of <0.00001 in healthy humans, consistent with very strong intolerance to gene inactivation.
BibTeX
@inproceedings{zou2016quantifying,
  author = {James Zou and Gregory Valiant and Paul Valiant and Konrad J. Karczewski and Siu On Chan and Kaitlin E. Samocha and Monkol Lek and Shamil Sunyaev and Mark J. Daly and Daniel G. MacArthur},
  title = {{Quantifying unobserved protein-coding variants in human populations provides a roadmap for large-scale sequencing projects}},
  booktitle = {Nature Communications},
  year = {2016},
  doi = {10.1038/ncomms13293},
}
Figure