3D-Jury



3D-Jury is a metaserver that aggregates and compares models from various protein structure prediction servers.

The 3D-Jury algorithm takes in groups of predictions made by a collection of servers and assigns each pair a 3D-Jury score, based on structural similarity. To improve accuracy of the final model, users can select the prediction servers from which to aggregate results. The authors of 3D-Jury designed the system as a meta-predictor because earlier results concluded that the average low-energy protein conformation (by way of aggregation) fit the true conformation better than simply the lowest-energy protein conformation.

The Robetta automatic protein structure prediction server incorporates 3D-Jury into its prediction pipeline.

As of January 2024, the links to 3D-Jury originally hosted by the BioInfoBank Institute are no longer valid.

Algorithm
First, pairwise comparisons are made between every combination of models generated from chosen protein prediction servers. Each comparison is then scored using the MaxSub tool. The score, $$sim( M_{a,b}, M_{i,j} ) $$, is generated by counting the number of Cα atoms in the two predictions within 3.5 Å of each other after being superpositioned.

To get a roughly 90% chance two models are of a similar fold class, the authors set a threshold of 40 as the lowest score possible for a pair of models to be annotated as "similar". The authors admittedly chose this threshold based on unpublished work.

There are two scores 3D-Jury gives: the best-model-mode score using one model from each server ($$3D-Jury-single(M_{a,b})$$) and the all-model-mode score that considers all models from each server ($$3D-Jury-all(M_{a,b})$$).

The best-model-mode score using one model per server, $$3D-Jury-single(M_{a,b})$$, is calculated as,

$$3D-Jury-single (M_{a,b}) = \frac{ \sum^N_i \max^{N_i}_{j,a \ne i \mbox{ OR } b \ne j}   (sim(M_{a,b}, M_{i,j})) }{   1 + N }$$

where $$N$$ is the number of servers and $$N_i$$ is the number of top ranking models (with a maximum of 10) from the server $$i$$, while a pairwise similarity score is calculated between models $$M_{a,b} $$ (model $$b $$ from server $$a  $$) and $$M_{i,j}  $$ (model $$j $$ from server $$i  $$).

While the all-model-mode score considering all models from the servers, $$3D-Jury-all(M_{a,b})$$, is calculated as,

using similar variables as noted with the best-model-mode score.

Note, these meta-predictor scores do not take into account the confidence scores from each of the models from other servers.