APO fine-tunes atomic structure models without measured coordinates by assessing consistency among their own predictions
APO fine-tunes atomic structure models without measured coordinates by assessing consistency among their own predictions
On July 30, Shentong Mo and Yatao Bian posted a preprint on arXiv describing Atomic Policy Optimization, or APO. The method fine-tunes a model using candidate 3D structures of crystals and antibodies without comparing them with experimentally measured coordinates.
Designing a crystal or antibody requires specifying the positions of its atoms in three dimensions. A generative model proposes a configuration and is then fine-tuned to select promising candidates. In FlowDPO, candidates are ranked by their distance from an already measured structure. These coordinates are not yet available for new crystal phases or proteins designed from scratch. They are obtained only after the material has been synthesized, the crystal has been grown, or the protein structure has been measured.
In the APO paper, the model generates several candidates for a single chemical composition or sequence. A spectral score identifies a group of similar structures among them. Candidates that are close to the most common structural type receive greater weight. A second reward uses a physical proxy: regularity of atomic packing for crystals and the absence of atomic clashes for antibodies. The model then updates its parameters based on a comparative assessment of the entire group.
The authors tested the method on three crystal datasets and on antibody structures. On MPTS-52, the Match Rate, which is the proportion of structures that matched the reference under the benchmark rules, was 21.14% for APO, compared with 20.27% for FlowDPO. For the antibody H3 loop, a region that often participates in antigen binding, the positional error for Cα atoms decreased from 3.32 to 3.21 angstroms. These are computational benchmark results reported in a preprint.
APO changes the source of feedback used during fine-tuning: instead of relying on a measured structure, the model uses consistency among its own candidates and a physical proxy. The experimental structure remains the reference for evaluating the result. In antibody discovery, this cycle moves part of the selection process to a stage before coordinate annotations become available and expands the computational component of developing biological interventions.