CASP8 Results - Meeting summary

The prediction methods from the McGuffin group were successful in a number of categories at CASP8. I was invited to give talks about my predictions of model quality using the ModFOLD server and my predictions of disorder using the DISOclust server, which were both of a high standard relative to those from most other groups. Slides for my talks can be downloaded using these links: ModFOLD talk, DISOclust talk. Below is an initial summary of results, which I have interpreted from the talks given by the official assessors .

Quality Assessment (QA) category - results for the ModFOLD servers (ModFOLDclust, ModFOLD v 2.0, ModFOLD v 1.1)

• The ModFOLDclust server predictions (Group number 031) ranked as the top overall in the QMODE1 category, although there was no significant difference between the top few methods.
• The human predictions (Group number 379 - McGuffin) ranked 1st in terms of the assessment of the per-residue error in our own TS models and ranked amongst the top five groups in terms of QMODE1. For these "human" predictions, I used ModFOLD v 2.0 beta to rank models from all servers.
• The ModFOLD v 1.1 predictions were comparable to the other single model/non-clustering based methods.

Disorder (DR) prediction category - results for the DISOclust server

• Both manual and server predictions using the DISOclust method (Groups 097 and 379) were ranked amongst the top five groups according to all scoring methods. In terms of AUC scores, the DISOclust server ranked 3rd, although the differences between the top few methods are not significant.

Tertiary Structure (TS) prediction category - results for the manual predictions and the nFOLD3 server

• The human TS predictions (Group number 379 - McGuffin) were ranked in the top 20 groups according to cumulative GDT_TS scores, however it seems that only 3 other human groups produced significantly better models according to the paired t-tests.
• The server only TS predictions (Group number 100 - nFOLD3) were ranked in the top 40 according to cumulative GDT_TS scores, however again only the top 20 or so servers appeared to be significantly better and several of these are not publicly available outside the context of the CASP experiment.

Domain prediction (DP) category - results for the DomFOLD server

• Methods similar to our new DomFOLDpdp method appeared to be the best overall, although we did not enter our new method in time for CASP8.
• The original DomFOLD server (Group number 317) ranked amongst similar methods which are less dependent on buliding a good initial model. These methods may be more useful on Free Modelling targets, however this was difficult to assess at CASP8 due to the sparse number of new folds with multiple domains.