Traditional Genome-Wide Association Studies (GWAS) rely on human-defined measurements—scalar values like height, blood pressure, or the volume of a specific brain region. While these metrics are intuitive, they create a "scalar bottleneck." By collapsing a complex, high-dimensional organ like the heart or brain into a single number, we discard the vast majority of biological information.
A new research direction seeks to bypass this bottleneck by using unsupervised deep learning to "discover" phenotypes directly from raw medical data. Instead of telling a model what to measure, researchers allow neural networks to compress images, waveforms, and 3D meshes into low-dimensional latent spaces. These learned features, often called "deep phenotypes" or "endophenotypes," capture subtle, distributed patterns that are highly heritable but invisible to the human eye.
The Scalar Bottleneck and the Latent Solution
In a standard GWAS, researchers look for correlations between genetic variants (SNPs) and a specific trait. If that trait is "lung function," they might use a single value from a spirometry curve. However, as Taedong Yun, Justin Cosentino, and colleagues demonstrate with the REGLE framework, the entire shape of that curve contains genetic signals that a single peak value misses.
By training Variational Autoencoders (VAEs) on high-dimensional clinical data, researchers can extract a "latent representation"—a vector of numbers that summarizes the most important structural features of the input. When these latent vectors are used as the targets for GWAS, the results are striking: researchers find significantly more genetic associations than they do with traditional metrics. For example, Khush Patel and Ziqian Xie used 3D convolutional autoencoders to identify dozens of novel genetic loci associated with brain structure that human-defined anatomical boundaries simply could not capture.
Why Unsupervised Features are More Heritable
It might seem counterintuitive that a model trained without any biological labels would find "better" traits than a trained physician. The reason lies in the objective function of unsupervised and self-supervised learning.
Models like the iGWAS framework, developed by Ziqian Xie and others, use self-supervised learning to identify features that are consistent across different views of the same biological system (such as the left and right eyes of an individual). This forces the model to ignore transient noise and focus on the underlying structural "blueprint."
Because these models are incentivized to capture the fundamental variance in the data, they often "accidentally" discover biological states. Tom M. George showed that VAEs trained on cancer transcriptomes spontaneously encoded drug resistance into their latent spaces without ever being told which patients responded to chemotherapy. The latent space becomes a mirror of the underlying biological drivers.
Beyond Static Anatomy: Capturing Motion and Aging
The power of deep phenotyping extends beyond static images to dynamic processes. Traditional cardiac metrics like "ejection fraction" describe how much blood the heart pumps, but they don't describe how the heart moves.
Researchers like Alejandro F. Frangi, Enzo Ferrante, and Rodrigo Bonazzola have moved toward spatiotemporal graph autoencoders that disentangle the static shape of the heart from its dynamic beating patterns. This approach revealed that the genetics governing heart structure are distinct from those governing heart motion. Similarly, Lyubov Shilova and the REECAP team used contrastive learning to align retinal images along a continuous "aging axis," discovering 178 genetic loci linked to eye health by treating biological age as a high-dimensional latent trait.
Integrating Multiple Modalities
The most recent evolution in this field is the move toward "holistic" representations. Rather than looking at one organ or one data type in isolation, frameworks like the one proposed by Adityanarayanan Radhakrishnan and Puneet Batra use cross-modal autoencoders to fuse electrical signals (ECGs) and structural imaging (MRIs).
By learning a shared latent space between these different views of the heart, the model can identify genetic variants that influence the "cardiovascular state" as a whole. This shift from "measuring parts" to "encoding states" allows tools like JAGWAS to perform joint association studies across hundreds of latent dimensions simultaneously, uncovering a much richer map of the human genome's influence on health and disease.