Refinement of AlphaFold-Multimer structures with single sequence input

Abstract

AbstractAlphaFold2, introduced by DeepMind in CASP14, demonstrated outstanding performance in predicting protein monomer structures. It could model more than 90% of targets with high accuracy, and so the next step would surely be multimer predictions, since many proteins do not act by themselves but with their binding partners. After the publication of After AlphaFold2, DeepMind published AlphaFold-Multimer, which showed excellent performance in predicting multimeric structures. However, its accuracy still has room for improvement compared to that of monomer predictions by AlphaFold2. In this paper, we introduce a fine-tuned version of AlphaFold-Multimer, named AFM-Refine-G, which uses structures predicted by AlphaFold-Multimer as inputs and produces more refined structures without the helps of multiple sequence alignments or templates. The performance of AFM-Refine-G was assessed using two datasets, Ghani_et_al_Benchmark2 and Yin_et_al_Hard, adapted from previous studies by Ghani et al. and Yin et al., respectively. The Ghani_et_al_Benchmark2 dataset consists of 17 recently published heteromers and the Yin_et_al_Hard dataset consists of 133 multimers, including immune-related complexes and repebody-antigen complexes, with several whose correct structure AlphaFold-Multimer could not predict. We predicted five models per target (750 models in total) and analyzed the improvement in the DockQ of each model. Of 750 models, 115 had DockQ improvement > 0.05 after refinement, demonstrating that our model is useful for the refinement of multimer structures. However, 14 structures had degraded DockQ < −0.05 after refinement, and the overall prediction quality for targets in Yin_et_al_Hard was quite low; 97 out of 133 were classified as ‘Incorrect’ with CAPRI criteria, revealing that there is still room for improving multimer predictions.