OmicsPred catalog brings together 3.3 million models that estimate RNA, protein, and metabolite levels from DNA data
OmicsPred catalog brings together 3.3 million models that estimate RNA, protein, and metabolite levels from DNA data
On September 1, Nature Genetics published an article describing the OmicsPred catalog. Each model is accompanied by information that helps researchers assess whether it is suitable for a new study.
Large studies collect genotypes, meaning sets of DNA variants carried by participants, much more often than they directly measure RNA, protein, and metabolite levels. A model is trained using data from people for whom both genotypes and these molecular measurements are available. It assigns weights to the contributions of DNA variants and predicts the level of a specific molecule. The resulting formula can then be applied to genotypes from another group or to summary results from genetic studies.
A list of coefficients alone is insufficient for this transfer. OmicsPred records the DNA variants and their weights, the genome assembly version that defines the variants' coordinates, the tissue, information about the training and validation samples, participant ancestry, and prediction accuracy. Researchers can therefore compare their own data with the conditions under which a model was trained and validated.
By May 2026, the catalog contained 3 339 469 models. Most predict gene activity in 49 human tissues, while others predict protein and metabolite levels. The first version of OmicsPred, released in 2023, contained 17 227 models. The current work converts models from different studies into a common format for description and search.
The authors applied the catalog to summary results from genetic studies conducted by the Million Veteran Program. They tested associations between predicted RNA and protein levels in blood and plasma and 1 233 health conditions in three ancestry groups: African, Admixed American, and European. This analysis produced about 46 million comparisons. After correction for the number of comparisons, more than 190 thousand associations remained statistically significant.
These associations identify molecules that warrant further study. The catalog shows which formula was used, which data were used to train it, and how accurately it predicts the relevant trait. Other researchers can therefore apply the model again and test these hypotheses using new data.