Improved multimer prediction using massive sampling with AlphaFold in CASP15

Abstract

AlphaFold2 has revolutionized structure prediction by achieving high accuracy comparable to experimentally determined structures. However, there is still room for improvement, especially for challenging cases like multimers. A key to the success of AlphaFold is its ability to assess and rank its own predictions. Our basic idea for the Wallner group in CASP15 was to exploit this excellent scoring function in AlphaFold by massive sampling. To achieve this goal, we conducted AlphaFold runs using six different settings, using templates, without templates, and with an increased number of recycles for both multimer v1 and v2 weights. In all instances, we enabled dropout layers during inference, allowing for sampling of uncertainty and enhancing the diversity of the generated models. In total, 274 289 models were generated for the 38 targets in CASP15, with a median of 4810 models per target. Of these 38 targets, 10 were high quality, 11 were medium quality, 11 were acceptable, and only 6 were incorrect. The improvement over the baseline method, NBIS-AF2-multimer, is substantial, with the mean DockQ increasing from 0.43 to 0.56, with several targets showing a DockQ score increase of +0.6 units. Remarkable, considering Wallner and NBIS-AF2-multimer were using identical input data. The success can be attributed to the diversified sampling using dropout with different settings and, in particular, the use of multimer v1, which is much more susceptible to sampling compared with v2. The method is available here: http://wallnerlab.org/AFsample/.

Keywords: ensembles; interactions; machine learning; multimer; protein structure prediction; sampling; scoring.