If you compare identical twins to fraternal twins, you can calculate exactly how much of a trait—such as height, bone density, or susceptibility to depression—is dictated by genetics. This classical method gives us an upper bound of genetic influence, known as heritability.
Yet, when scientists began scanning the actual DNA of hundreds of thousands of individuals using Genome-Wide Association Studies (GWAS), they hit a paradox. Summing up the effects of every discovered genetic variant accounted for only a tiny fraction of the heritability that twin studies promised was there. For instance, while family studies show that conditions like preeclampsia or osteoporosis are highly heritable, early polygenic risk models failed to translate this into accurate individual risk predictions.
This discrepancy is known as the missing heritability problem. The variance isn't actually missing from our biology; rather, it is hidden from our standard statistical tools. To uncover it, researchers have shifted from looking at isolated, common genetic variants toward strategies that aggregate weak signals, integrate intermediate biological layers, and map rare, population-specific mutations.
The Statistical Wall of Single-Variant Analysis
To understand why heritability goes missing, we have to look at how traditional GWAS works. A standard study tests millions of common single nucleotide polymorphisms (SNPs) one by one to see if any single letter change correlates with a disease. To avoid false positives across millions of tests, researchers must use an incredibly stringent statistical significance threshold (typically $p < 5 \times 10^{-8}$).
This creates a massive blind spot. If a complex trait is driven by thousands of different genetic variants, each exerting a tiny, subtle effect on the phenotype, most of these variants will fail to clear the high statistical hurdle. They remain buried in the background noise.
As Karen Kapur demonstrated in a large-scale meta-analysis of serum calcium levels, even a highly robust, strongly associated genetic locus might explain just over 1% of the total phenotypic variance. The remaining genetic architecture is composed of a vast sea of sub-threshold variants. When researchers build conventional Polygenic Risk Scores (PRS) by simply adding up these isolated top hits, the scores lack the predictive power needed to accurately model complex structural or behavioral phenotypes.
Strategy 1: Aggregating Whispers into Shouts via Pathways
If individual genetic variants are too weak to pass statistical thresholds on their own, how do we find them? The solution is to change the unit of analysis from the individual SNP to the biological pathway.
Think of a biological pathway as an assembly line in a factory. If ten different workers along the line slow down slightly, the factory's output drops significantly. Testing each worker individually might not reveal a statistically significant anomaly, but evaluating the assembly line as a whole makes the problem obvious.
Traditional GWAS: [SNP 1] [SNP 2] [SNP 3] --> Evaluated in isolation (Signals fail threshold)
Pathway Analysis: [ SNP 1 + SNP 2 + SNP 3 ] --> Aggregated into functional unit (Signal amplifies)
By grouping variants into known functional networks, researchers can amplify weak genetic signals. For example, Roxana Daneshjou et al. applied this pathway-based aggregation to improve clinical predictions for complex drug dosages, such as warfarin. Instead of treating variants as independent variables, they demonstrated that accounting for the collective interplay within biological pathways captures variance that standard single-variant models miss.
Similarly, Pingzhao Hu et al. showed that pathway-based joint analysis significantly boosts statistical power when analyzing rare genetic variants. By transforming a high-dimensional search for single mutations into a targeted analysis of disrupted biological systems, pathway modeling helps recover the missing heritability overlooked by standard genome-wide scans.
Strategy 2: Integrating the Transcriptome as an Intermediate Layer
Another reason heritability goes missing is that DNA does not act directly on a phenotype. A genetic variant on a chromosome must first be transcribed into RNA, which is then translated into proteins, which ultimately alter cells and tissues. Standard GWAS skips these intermediate steps, attempting to draw a straight line from a single DNA base pair to a complex trait.
Transcriptome-Wide Association Studies (TWAS) bridge this gap by integrating gene expression models directly into genetic risk assessments. Instead of asking whether a specific SNP correlates with a disease, TWAS uses reference datasets to predict how a person's collective genetic profile alters gene expression in specific tissues, and then correlates those expression levels with the trait.
Bingxin Zhao et al. leveraged this approach to study human brain structures and their overlap with neuropsychiatric traits. The authors demonstrated that transcriptome-based risk scores significantly outperform conventional variant-based models in predicting complex structural phenotypes. By capturing the cumulative effect of multiple variants on actual gene expression, intermediate biological layers preserve the genetic signals that get diluted in single-variant DNA models.
Strategy 3: Hunting for Rare and Population-Specific Variants
Standard genomic microarrays are designed to capture common variants—those present in more than 1% to 5% of the population. However, evolutionary selection tends to keep highly deleterious mutations rare. A variant with a massive impact on disease risk is unlikely to become common in the global population.
Consequently, focusing exclusively on common variations systematically misses low-frequency mutations that carry high individual risk. Yunzhi Zhang et al. argued that fully mapping the genetic architecture of complex conditions like osteoporosis requires deep sequencing rather than standard genotyping arrays, specifically to capture these low-frequency and rare mutations.
One of the most effective ways to catch these elusive variants is to study genetically isolated or founder populations. In a massive, diverse global population, a rare mutation is incredibly difficult to detect statistically. However, due to historical bottlenecks, that same mutation might be enriched and highly prevalent within a specific founder group.
Global Population: [Mutation X] is 1 in 100,000 (Statistically invisible)
Founder Population: [Mutation X] is 1 in 100 (Statistically detectable)
Eimear E. Kenny et al. demonstrated the power of this strategy by conducting a genome-wide scan for Crohn's disease within the Ashkenazi Jewish population. By leveraging the unique genetic architecture of this cohort, the authors identified novel susceptibility loci that had been completely missed by much larger studies conducted in broader, heterogeneous European populations.
Strategy 4: Correcting Mathematical Biases in Polygenic Estimators
Sometimes, the heritability isn't missing from the genome at all; it is simply missing from our mathematical estimators. When researchers calculate the genetic correlation between two different traits using summary statistics, standard polygenic risk score methods frequently introduce systematic downward biases.
To address this mathematical limitation, Bingxin Zhao developed a bias-corrected estimation framework specifically designed for GWAS summary statistics. By mathematically correcting for the noise and sample overlaps that inherently deflate traditional polygenic scores, this framework revealed a significantly stronger genetic overlap between complex traits than previously recognized.
Furthermore, traditional heritability models often fail when applied across diverse groups. As Honglei Chen et al. demonstrated, polygenic scores derived from single-ancestry cohorts rapidly lose their predictive power when applied across different racial and ethnic backgrounds. This drop highlights that missing heritability is also a consequence of incomplete global mapping; capturing the full spectrum of human genetic variance requires models that account for diverse ancestral backgrounds.
By moving away from the reductive assumption that complex traits can be explained by a handful of isolated, universal DNA markers, modern genomics is steadily closing the heritability gap. Through signal aggregation, multi-omic integration, and population-specific sequencing, the "missing" pieces of the genetic puzzle are finally being put back into place.