Leveraging unpaired sequences in all-atom protein co-generation

Develop methods to leverage unpaired protein sequences—sequences without associated structures—within all-atom protein generative models such as Protpardelle and P(all-atom), enabling effective training and generation when only sequence data is available.

Background

All-atom co-generation directly models all backbone and side-chain atoms, implicitly co-generating sequence and structure and offering fine-grained control crucial for tasks like enzyme and antibody design. Recent methods include Protpardelle, which diffuses over a fixed atom73 representation and decodes sequence mid-process, and P(all-atom), which uses an atom14 representation and decodes sequence post-generation.

The authors highlight both computational challenges due to high dimensionality and a key data challenge: current all-atom approaches do not provide a clear pathway to exploit the vast corpus of sequences lacking paired structures, limiting training scale and applicability. Addressing how to incorporate such unpaired sequences is an explicit unresolved question.