Refinement of AlphaFold-Multimer structures with single sequence input

Abstract

AlphaFold2, 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 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 refined structures without the help of multiple sequence alignments or templates. The performance of AFM-Refine-G was assessed using four datasets: Ghani_et_al_Benchmark2 and Yin_et_al_Hard using AlphaFold-Multimer version 2.2 outputs, and CASP15_multimer and Yin_and_Pierce_af23 using AlphaFold-Multimer version 2.3 outputs. Of 1925 predicted structures, 203 had DockQ improvement > 0.05 after refinement, demonstrating that our model is useful for the refinement of multimer structures. However, considering the per target success rate, the overall improvement was modest, suggesting that the original AlphaFold-Multimer network had already learned a biophysical energy function independent of MSAs or templates, as proposed by Roney and Ovchinnikov (Roney and Ovchinnikov, 2022). Furthermore, both the default AlphaFold-Multimer and our refinement model showed lower performance for immune-related targets compared to general targets, indicating that room for improvement remains. AvailabilityThe inference scripts are available from https://github.com/t-oda-ic/afm_refiner under the Apache License, Version 2.0. The network parameters are available from https://figshare.com/articles/online_resource/afm_refine_g_20230110_zip/21856407 under the license CC BY 4.0.