The state-of-the-art to assess the structural quality of docking models is currently based on three related yet independent quality measures: F_nat, LRMS, and iRMS as proposed and standardized by CAPRI. These quality measures quantify different aspects of the quality of a particular docking model and need to be viewed together to reveal the true quality, e.g. a model with relatively poor LRMS (>10Å) might still qualify as 'acceptable' with a descent F_nat (>0.50) and iRMS (<3.0Å). This is also the reason why the so called CAPRI criteria for assessing the quality of docking models is defined by applying various ad-hoc cutoffs on these measures to classify a docking model into the four classes: Incorrect, Acceptable, Medium, or High quality. This classification has been useful in CAPRI, but since models are grouped in only four bins it is also rather limiting, making it difficult to rank models, correlate with scoring functions or use it as target function in machine learning algorithms. Here, we present DockQ, a continuous protein-protein docking model quality measure derived by combining F_nat, LRMS, and iRMS to a single score in the range [0, 1] that can be used to assess the quality of protein docking models. By using DockQ on CAPRI models it is possible to almost completely reproduce the original CAPRI classification into Incorrect, Acceptable, Medium and High quality. An average PPV of 94% at 90% Recall demonstrating that there is no need to apply predefined ad-hoc cutoffs to classify docking models. Since DockQ recapitulates the CAPRI classification almost perfectly, it can be viewed as a higher resolution version of the CAPRI classification, making it possible to estimate model quality in a more quantitative way using Z-scores or sum of top ranked models, which has been so valuable for the CASP community. The possibility to directly correlate a quality measure to a scoring function has been crucial for the development of scoring functions for protein structure prediction, and DockQ should be useful in a similar development in the protein docking field. DockQ is available at http://github.com/bjornwallner/DockQ/

Summary: The AlphaFold2 neural network model has revolutionized structural biology with unprecedented performance. We demonstrate that by stochastically perturbing the neural network by enabling dropout at inference combined with massive sampling, it is possible to improve the quality of the generated models. We generated ∼6000 models per target compared with 25 default for AlphaFold-Multimer, with v1 and v2 multimer network models, with and without templates, and increased the number of recycles within the network. The method was benchmarked in CASP15, and compared with AlphaFold-Multimer v2 it improved the average DockQ from 0.41 to 0.55 using identical input and was ranked at the very top in the protein assembly category when compared with all other groups participating in CASP15. The simplicity of the method should facilitate the adaptation by the field, and the method should be useful for anyone interested in modeling multimeric structures, alternate conformations, or flexible structures. Availability and implementation: AFsample is available online at http://wallnerlab.org/AFsample.

Background: Employing methods to assess the quality of modeled protein structures is now standard practice in bioinformatics. In a broad sense, the techniques can be divided into methods relying on consensus prediction on the one hand, and single-model methods on the other. Consensus methods frequently perform very well when there is a clear consensus, but this is not always the case. In particular, they frequently fail in selecting the best possible model in the hard cases (lacking consensus) or in the easy cases where models are very similar. In contrast, single-model methods do not suffer from these drawbacks and could potentially be applied on any protein of interest to assess quality or as a scoring function for sampling-based refinement. Results: Here, we present a new single-model method, ProQ2, based on ideas from its predecessor, ProQ. ProQ2 is a model quality assessment algorithm that uses support vector machines to predict local as well as global quality of protein models. Improved performance is obtained by combining previously used features with updated structural and predicted features. The most important contribution can be attributed to the use of profile weighting of the residue specific features and the use features averaged over the whole model even though the prediction is still local. Conclusions: ProQ2 is significantly better than its predecessors at detecting high quality models, improving the sum of Z-scores for the selected first-ranked models by 20% and 32% compared to the second-best single-model method in CASP8 and CASP9, respectively. The absolute quality assessment of the models at both local and global level is also improved. The Pearson’s correlation between the correct and local predicted score is improved from 0.59 to 0.70 on CASP8 and from 0.62 to 0.68 on CASP9; for global score to the correct GDT_TS from 0.75 to 0.80 and from 0.77 to 0.80 again compared to the second-best single methods in CASP8 and CASP9, respectively. ProQ2 is available at http://proq2.wallnerlab.org.

Protein interactions are key in vital biological processes. In many cases, particularly in regulation, this interaction is between a protein and a shorter peptide fragment. Such peptides are often part of larger disordered regions in other proteins. The flexible nature of peptides enables the rapid yet specific regulation of important functions in cells, such as their life cycle. Consequently, knowledge of the molecular details of peptide-protein interactions is crucial for understanding and altering their function, and many specialized computational methods have been developed to study them. The recent release of AlphaFold and AlphaFold-Multimer has led to a leap in accuracy for the computational modeling of proteins. In this study, the ability of AlphaFold to predict which peptides and proteins interact, as well as its accuracy in modeling the resulting interaction complexes, are benchmarked against established methods. We find that AlphaFold-Multimer predicts the structure of peptide-protein complexes with acceptable or better quality (DockQ ≥0.23) for 66 of the 112 complexes investigated—25 of which were high quality (DockQ ≥0.8). This is a massive improvement on previous methods with 23 or 47 acceptable models and only four or eight high quality models, when using energy-based docking or interaction templates, respectively. In addition, AlphaFold-Multimer can be used to predict whether a peptide and a protein will interact. At 1% false positives, AlphaFold-Multimer found 26% of the possible interactions with a precision of 85%, the best among the methods benchmarked. However, the most interesting result is the possibility of improving AlphaFold by randomly perturbing the neural network weights to force the network to sample more of the conformational space. This increases the number of acceptable models from 66 to 75 and improves the median DockQ from 0.47 to 0.55 (17%) for first ranked models. The best possible DockQ improves from 0.58 to 0.72 (24%), indicating that selecting the best possible model is still a challenge. This scheme of generating more structures with AlphaFold should be generally useful for many applications involving multiple states, flexible regions, and disorder.