GenShin

When a lipid nanoparticle enters the bloodstream, plasma proteins adsorb onto its surface within seconds, forming a protein corona that governs where the particle goes, how long it circulates, and whether the immune system clears it. Characterising that corona today means synthesising the liposome and running mass spectrometry on it, which rules the measurement out as a pre-synthesis screen over a large candidate lipid space. GenShin reframes the problem: instead of simulating adsorption, it scores every lipid-protein pair and ranks the scores within each lipid, treating the resulting ordering as a proxy for relative corona composition.

The obstacle to doing this at scale is that most structure-based affinity models expect a docked intermolecular pose, and docking a lipid against thousands of plasma proteins is both expensive and unreliable. GenShin is pose-free by construction: it encodes the protein subgraph and the lipid graph independently, then builds cross-graph interaction representations directly, never requiring a bound complex. It sits between scoring functions such as GatorAffinity, which consume a complex structure, and sequence-level predictors like AQAffinity, which discard geometry entirely.

The model comes from Pingfei Zhu and colleagues at Jilin University's Key Laboratory of Symbolic Computation and Knowledge Engineering, working with the State Key Laboratory of Natural and Biomimetic Drugs at Peking University.

#Key Features

• Docking-pose-free scoring: cross-graph interaction representations are constructed and refined from the two independently encoded graphs, so no intermolecular pose is needed at training or inference.
• Geometry-enhanced encoders: the protein encoder uses fixed edge-angle embeddings over directed edge representations, and the lipid encoder applies GIN-style message passing augmented with bond-angle features.
• Leakage-controlled pretraining: PDBbind ligand SDF files are converted to SMILES and re-embedded with RDKit conformers decoupled from protein coordinates, preventing the pretraining set from smuggling in pose information.
• Distance-map auxiliary supervision: a node-level inter-side distance-map prediction task runs alongside affinity regression, improving geometric consistency without requiring a pose at inference.
• Rank fine-tuning on measured coronas: pairwise ranking supervision within each liposome teaches the model to place experimentally abundant hard-corona proteins above undetected ones.

#Technical Details

Protein inputs are subgraphs centred on the compound region and on P2Rank-predicted pocket sites, with subgraph-specific margin-based supervision; where several subgraphs exist for one pair, the highest-scoring one is taken. Both encoders share a hidden dimension of 128, and the pretrained model totals roughly 1.9M parameters, scoring a protein-compound pair in 0.68 s on CPU and 0.35 s on GPU, or 0.0147 s per subgraph when the distance-map branch is disabled.

On the standard PDBbind v2016 split, pretrained GenShin reaches an RMSE of 1.189 and Pearson's R of 0.839 without a pose, against 1.217 and 0.8305 for the pose-dependent CurvAGN. CASF-2016 perturbation tests apply 90-degree rotations, 5 Å translations, and 0-5 Å coordinate noise: GenShin holds at RMSE 1.220 while the pose-dependent comparator degrades by 104%. Ranking fine-tuning uses hard-corona abundances measured on ten synthesised liposomes spanning ionisable, phospholipid, and nucleotide-lipid classes, split 8/1/1 at the liposome level and yielding about 24,450 lipid-protein pairs against a 2,445-protein mouse plasma panel. On the held-out test liposome, predicted and measured corona rankings agree at Spearman's rho of 0.919.

#Applications

The deployed workflow is de novo liposome design. A combinatorial library of 34,164 candidate lipids, built by recombining 73 head, 9 linker, and 52 tail motifs and geometry-optimised with MMFF or UFF, is scored against every protein in the plasma panel. Proteins are ranked within each lipid and assessed against criteria covering protective, phagocytosis-marking, targeting-related, clearance-associated, and stability-supportive protein classes, which narrowed the library to 1,654 candidates with favourable predicted corona profiles. The same pose-free scoring is directly applicable to conventional small-molecule virtual screening where docked poses are unavailable or untrustworthy.

#Impact

GenShin's useful contribution is showing that a compact graph model can score protein-ligand pairs competitively without an intermolecular pose, and that pose-dependent scoring degrades sharply when poses are perturbed while pose-free scoring does not. That robustness is what makes corona-scale screening tractable at all. The evaluation is entirely computational: the corona rankings are validated against previously measured mass spectrometry data, and none of the 1,654 selected lipids has been synthesised or tested. The named GitHub repository contains only a placeholder README, with no code, weights, or license released, so the results cannot currently be reproduced or the model reused. The work remains a preprint and has not been peer reviewed.