Progress and Bottlenecks for Deep Learning in Computational Structure Biology: CASP Round XVI

CASP16 is the most recent in a series of community experiments to rigorously assess the state of the art in areas of computational structural biology. The field has advanced enormously in recent years: in early CASPs, the assessments centered around whether the methods were at all useful. Now they mostly focus on how near we are to not needing experiments. In most areas, deep learning methods dominate, particularly AlphaFold variants and associated technology. In this round, there is no significant change in overall agreement between calculated monomer protein structures and their experimental counterparts, not because of method deficiencies but because, for most proteins, agreement is likely as high as can be obtained given experimental uncertainty. For protein complexes, huge gains in accuracy were made in the previous CASP, but there still appears to be room for further improvement. In contrast to these encouraging results, for RNA structures, the deep learning methods are notably unsuccessful at present and are not superior to traditional approaches. Both approaches still produce very poor results in the absence of structural homology. For macromolecular ensembles, the small CASP target set limits conclusions, but generally, in the absence of structural templates, results tend to be poor and detailed structures of alternative conformations are usually of relatively low accuracy. For organic ligand-protein structures and affinities (important for aspects of drug design), deep learning methods are substantially more successful than traditional ones on the relatively easy CASP target set, though the results often fall short of experimental accuracy. In the less glamorous but essential area of methods for estimating the accuracy, previous results found reliable accuracy estimates at the amino acid level. The present CASP results show that the best methods are also largely effective in selecting models of protein complexes with high interface accuracy. Will upcoming method improvements overcome the remaining barriers to reaching experimental accuracy in all categories? We will have to wait until the next CASP to find out, but there are two promising trends. One is the combination of traditional physics-inspired methods and deep learning, and the other is the expected increase in training data, especially for ligand-protein complexes.