VIDRA: genetics as a dose, not a yes/no

On Genetic dose-response modelling predicts drug mechanisms, dosing, and adverse events — Stefanucci et al., medRxiv, 2026.

I just quickly glanced at the interesting paper VIDRA by the Open Targets team that estimates the genetic dose-response curve between gene activity and disease.

When we think of using human genetics as evidence for new drug discovery, we often view the human genetics evidence as binary: either it exists or it doesn't. In reality, it is a continuous spectrum.

There are simple variants that do not change the function, simple variants that slightly change protein function, others that cause more damage and Mendelian variants that can completely shut down the protein. Together, these form a continuum of naturally occurring perturbations. In pharma speak, genetics has a quantitative dose effect like a drug on function/phenotype.

Thinking of naturally occurring genetic variations as a genetic dose and measuring its relationship is extremely useful in drug discovery. One obvious use is asking whether a target should be inhibited or activated, which genes have the strongest effect, and how strongly the disease depends on changes in its activity.

What is interesting is that the team did not do huge perturbation experiments like Perturb-seq or deep mutational scanning. Instead, they used a huge collection of GWAS, eQTL, pQTL, and rare variant datasets across studies and diseases to capture the effect of genetic dose on function.

The challenge is how to combine these disparate data types, each measuring different aspects of gene function. For example, eQTL measures the effect of variants on RNA expression, pQTL measures the effect of variants on protein expression, and ClinVar measures pathogenicity. They all measure slightly different aspects of biology on different scales.

VIDRA addresses the challenge using one of my favorite statistical techniques, partial pooling. A naive approach is complete pooling, i.e. standard regular regression that lumps everything together and does not differentiate the different data types and their scales. The opposite extreme is treating each data type completely independently.

Partial pooling aims to hit the sweet spot through hierarchical modeling. It allows each modality like eQTL or pQTL to have its own parameters, while still borrowing information from the others using a shared underlying distribution. This builds a unified, accurate estimate of genetic dose.

Can't wait to dig more and play with the data.