SolvCLIP

Most structure-based drug discovery models learn from a single crystallographic ligand pose, as if a small molecule had one shape. In reality a ligand samples an ensemble of conformations whose populations shift with its chemical environment, and the pose that binds a pocket in water is not necessarily the one favored in a screening buffer or an organic co-solvent. SolvCLIP, developed at The Hong Kong Polytechnic University with collaborators at Lingnan University and Hong Kong Sanatorium & Hospital, makes that environmental dependence an explicit part of pretraining: conformer ensembles generated under 39 different organic solvents serve as augmented views of the same molecule, and the model is trained to produce representations that stay consistent across solvents while still encoding the geometric flexibility each one reveals.

The framework sits on top of the SE(3)-equivariant transformer encoders introduced by Uni-Mol. Separate pocket and ligand encoders are kept frozen throughout pretraining and inference, and a trainable SE(3)-equivariant interaction module learns on top of their representations under three simultaneous objectives: masked molecular reconstruction, interatomic distance matrix prediction, and contrastive alignment of solvent-perturbed conformers. The result is a single pretrained interaction model that is then applied to three separate downstream problems — binding affinity regression, ligand docking, and large-scale virtual screening — rather than being refit for each.

SolvCLIP is a preprint that has not yet been peer reviewed. Training and data-preparation code is released under the MIT license, but no pretrained checkpoint has been published, so reproducing the model requires running the full pretraining pipeline and obtaining the upstream Uni-Mol encoder weights.

#Key Features

• Solvent-aware conformer augmentation: Ligand conformational ensembles spanning 39 organic solvents act as physically grounded augmentations rather than random geometric noise, teaching the model how chemical environment reshapes an accessible pose.
• Three complementary pretraining objectives: Masked molecular reconstruction with a Gaussian cross-entropy loss captures local geometry, interatomic distance matrix prediction with an L2 regression loss captures spatial relationships, and a SimCLR-style contrastive term enforces solvent invariance.
• Frozen encoders with a trainable interaction module: Only the interaction layers are updated, which keeps compute modest and isolates the contribution of solvent-aware training from the underlying molecular representation.
• One pretrained model across three task families: The same checkpoint serves affinity regression, pose prediction, and enrichment-based screening, which is the evidence that the learned representation transfers rather than overfits a single benchmark.

#Technical Details

Pretraining uses 348,970 protein-ligand complexes from the January 2025 BioLiP release, filtered to 311,123 solvent-augmented complexes covering 16,037,848 generated conformations. Training runs for 20 epochs on four NVIDIA A800 GPUs with Adam, a learning rate and weight decay of 1e-4, a global batch size of 32, and gradient clipping at norm 1.0; masking probability is 0.15 with a masking and noise ratio of 0.8. On the Atom3D ligand binding affinity split derived from PDBbind (3,507 train / 466 validation / 490 test), SolvCLIP reaches 1.239 RMSE with Pearson 0.793 and Spearman 0.791, against 1.357 / 0.753 for Uni-Mol and 1.343 / 0.765 for ProNet. For docking, a model trained on 18,404 PDBbind v2020 complexes hits 76.5% of poses under 2.0 Å RMSD on CASF-2016, 64% on the 428-complex PoseBusters set that has become the standard stress test for co-folding systems such as Boltz-2, and 82% on the 85-complex Astex set, with a case study on the azafagomine-glucosidase complex (PDB 2J7H) reproducing the crystal pose to 0.157 Å. On DUD-E virtual screening across 102 targets under three-fold cross-validation, it records 97.1% AUC and an enrichment factor of 109.20 at 0.5%, the strongest early enrichment among the compared methods.

#Applications

The intended users are computational chemistry and structure-based design groups that need one representation to serve several stages of a campaign: ranking hits from a screening library, scoring analogs during lead optimization, and generating plausible binding poses for medicinal chemistry review. The solvent-aware framing is most relevant where assay and formulation conditions differ from crystallization conditions, or where a ligand is conformationally flexible enough that a single modeled pose is a poor summary of its behavior.

#Impact

SolvCLIP is an argument that physically meaningful augmentation beats generic perturbation for molecular pretraining — solvent-driven conformer ensembles carry real thermodynamic information, and using them as contrastive views yields gains without touching the underlying encoder. The evidence is honest about where it stands: Uni-Mol Docking still leads on the CASF-2016 sub-2 Å success rate (80.4% versus 76.5%), and DUD-E AUC is effectively saturated across recent methods, so the meaningful separation shows up in early enrichment rather than headline AUC. The absence of a released checkpoint is the main practical barrier to adoption, and full reproducibility also depends on the upstream Uni-Mol pretrained encoders that SolvCLIP freezes.