Nessim Raouraoua

1 Introduction Since its debut with AFsample [1], massive sampling with AlphaFold2 [2, 3] has consistently outperformed alternative approaches, especially for the prediction of protein complexes. In CASP15-CAPRI, the Wallner group's top ranking demonstrated that pushing AF2 for greater sampling and diversity can uncover solutions otherwise missed [4-6]. We then refined this strategy in MassiveFold [7], leveraging maximal parallelization on GPU clusters to reduce turnaround times.

1 Introduction Biomolecular interactions govern a wide range of cellular and intercellular activities. Characterizing these interactions at the atomic level is essential for understanding the mechanisms underlying health and disease and for developing new therapeutics.

1 Introduction Protein structure prediction was recently revolutionized by DeepMind's AlphaFold2 (AF2) deep learning based software [1]. At CASP14-CAPRI in 2020, DeepMind showed predictions of monomeric structures with a quality close to the one of experimental determination [2, 3]. In 2021, they published an AlphaFold-Multimer version adapted to protein complex prediction [4]. Version 1 came with five neural network (NN) models.