Protein–ligand scoring stops asking only how much to pass: attention now encodes then decodes, docking success reaches 88%

The attention mechanism learns one more thing: the same source information can be rewritten differently for each target, not merely weighted differently. Standard attention uses a scalar weight to decide how much information a source sends to a target, and the same source value representation is shared across all targets. With the same persistent spectral representations and surrounding transformer architecture, swapping latent interaction attention back to standard self-attention consistently lowers performance across the three CASF benchmarks, for example dropping CASF-2016 PCC from 0.857 to 0.819.

The method first compresses source information into a low-dimensional latent interaction space and then lets the target decode it back into feature space, giving target-dependent transformation without a separate mapping for every source–target pair. Parameterizing an independent transformation for every source–target pair would require explicit dense mappings in the high-dimensional feature space, which is costly and adds unnecessary flexibility. Sweeping the latent dimension from 2 to 64 on CASF-2007, PCC peaks at dimensions 4 and 8 (0.808) and falls to 0.765 at 64, where an out-of-memory failure occurs on a 32 GB NVIDIA Tesla V100S.

The reported evaluation covers scoring and docking on three-dimensional protein–ligand complexes: pretraining used 19,462 unlabeled complexes, scoring was evaluated on the CASF-2007, CASF-2013, and CASF-2016 core sets, docking on the 195 ligands of CASF-2013, and generalization on the LBA30 and LBA60 sequence-identity splits. The authors state that latent interaction attention is not specific to protein–ligand systems and may extend to protein–protein, protein–peptide, protein–nucleic acid, and mutation-induced interaction modeling, and that it can be applied to other large language models involving heterogeneous components.

The authors note that the current formulation primarily models the protein and ligand and captures environmental factors only indirectly, so complexes strongly influenced by structural water molecules, metal ions, cofactors, or other local chemical environment components remain more challenging; the error analysis shows the largest-error outlier ligands have very low similarity to the training set (for example 0.205 and 0.193). The authors also state that an optimal compact latent dimension does not imply protein–ligand interactions are intrinsically low-dimensional, only that information exchange within latent interaction attention can be mediated through a compact space.