Figure 6 Superposition of the experimental (green) and
closest calculated structures (blue) for two protein complexes. T1113 is
a phage shell homodimer with intertwined polypeptide chains at the
interface. The left-hand subunit is shown in grey, to clarify the
intertwined region. The interface score (ICS) is 93%. H1140 is a
nanobody (right)-antigen (left) complex (ICS 81%).
Figure 6 shows two examples of high agreement with the corresponding
experimental structures, each representing an assembly challenge class
that was problematic for older methods: T1113 is a small bacterial
homodimer with no homologous structures available. The two polypeptide
chains intertwine across the interface in a domain-swap like manner, a
feature that defeats classical docking methods. Deep learning methods
treat the whole complex as a single entity, circumventing that
difficulty. H1140 is a nanobody /protein antigen complex. Earlier
benchmarking (29) had shown that immune complexes defeat at least the
standard AlphaFold2-Multimer procedure. In CASP15 there are a total of
eight immune complex targets (five nanobody complexes and three antibody
complexes). Of these, two had homologous experimental complexes
available, and so were easy targets. Three others have the lowest
accuracy interfaces in this CASP (ICS 0.12, 0.30, and 0.45). But for the
remaining three, high quality (ICS 0.74, 0.80, 0.81) models were
produced by a small number of participating groups. Standard AlphaFold
Multimer with default parameters was not effective on any of these, in
accordance with the general observation that for many targets, enhanced
sampling is necessary to obtain the best results.
As with the single protein structure category, the most effective
methods in assembly modeling are based on AlphaFold2, usually the newer
AlphaFold-multimer (31), a version of AlphaFold where training included
data for protein complexes. Three of the methods are also in the top
five performers in the single protein category. Again, as with the
single protein category, the most successful groups used modifications
of standard AlphaFold procedures, including much more extensive sampling
through variations on MSA construction, the use of multiple seeds, an
increased number of recycles and extensive network dropout. In addition,
one group (32) devised a machine learning/Voronoi polyhedral interface
scoring function which evidently aided in selection of accurate models.
Details of methods can be found in the CASP15 assemblies assessment
paper (11) and papers by some of the best performing groups.
Although the improvement in accuracy is enormous, there are still a
substantial fraction of poor scoring interfaces. There are multiple
possible reasons for the lag in performance. This is the first-time deep
learning methods have been extensively used for protein complexes
whereas this was the third CASP where deep learning has been used for
single proteins. Thus, we may see substantial improvement next time as
lessons are learned from CASP15. More fundamentally, there are many
fewer structures of complexes in the PDB than single proteins, so that
training set is inherently smaller. It may be possible to use the
current methods to generate additional synthetic training data (33).
Analogously to single proteins, interface accuracy probably falls off
with the depth of the multiple sequence alignment spanning the interface
(although one leading group reported omitting these data (34)). That may
explain the generally weak performance for immune complexes, so it is
encouraging to see partial success there.