GSnet

Many of the properties that determine how a protein behaves in solution — how fast it diffuses, how much surface it exposes to solvent, how strongly it is solvated, the ionization state of a side chain at physiological pH — are calculable from a 3D structure but expensive to calculate. Solving the Poisson-Boltzmann equation for solvation free energy or running constant-pH molecular dynamics for a pKa can take days per protein, which rules those methods out for proteome-scale annotation just as predicted structures have made proteome-scale structural analysis possible.

GSnet, from Michael Feig's laboratory at Michigan State University, is a graph transformer trained to predict six such properties directly from coordinates: solvation free energy, radius of gyration, hydrodynamic radius, translational and rotational diffusion coefficients, and molecular volume. Its more interesting claim is not the speedup but the transfer. Because the network is pretrained on a supervised multi-property objective, its node embeddings encode structural determinants useful for tasks it was never trained on — solvent-accessible surface area, and, via a companion atom-level network called aLCnet, residue-specific pKa values. On both, the GSnet embeddings outperform general-purpose protein representations including ESM-2 and GearNet.

The work was posted in December 2024 and published in the Journal of Chemical Theory and Computation in 2025. Several unrelated models are also called GSNet — in vehicle pose estimation, traffic risk, and image segmentation — and this one's artifact lives in a repository named ProteinStructureEmbedding, which is easy to miss.

#Key Features

• Six physicochemical and geometric targets at once: One network is trained jointly on solvation free energy, radius of gyration, hydrodynamic radius, both diffusion constants, and volume, with a gradient mask handling structures where a reference value is missing.
• Rotation-invariant residue graph: Nodes are residues, edges connect alpha carbons within 15 Å, and features are E(3)-invariant — learned amino acid embeddings, dihedral encodings, and each alpha carbon's distance to the center of mass.
• Embeddings that carry to new tasks: Freezing the pretrained network and training only a small output head predicts solvent-accessible surface area accurately, the paper's central transfer-learning result.
• aLCnet for local electrostatics: A sibling network builds an atom-level graph within 10 Å of the target residue, adding element identity and PDB2PQR partial charges as node features and 5 Å atom-atom edges, giving it the charge sensitivity pKa prediction requires.

#Technical Details

GSnet takes 150-dimensional node features through six transformer message-passing layers, with edge features formed by expanding each alpha-carbon distance over 300 Gaussian basis functions spaced from 0 to 15 Å. Pretraining used AlphaFold2 models for a random 153,513-protein subset of UniProtKB/Swiss-Prot, with reference values from HYDROPRO and MDTraj; solvation free energies came from APBS for a 30,114-protein subset. The training split held 138,290 structures spanning 16 to 1,015 residues, with 15,223 for validation, and BLAST checks confirmed that validation accuracy did not depend on homology to the training set.

Errors on the geometric targets range from 0.65% to 2.47%, below the roughly 4% error HYDROPRO itself shows against experiment; solvation free energy is weaker at 3.89%, and GSnet learned features specific to AlphaFold models when predicting it, since predictions shift systematically on experimental PDB structures. On the transfer task, an output head over frozen GSnet embeddings predicted molecular SASA with an RMSE of 2.87 nm² against 10.79 for ESM-2 and 11.18 for GearNet. For pKa, aLCnet was pretrained on the simulation-derived PHMD549 dataset and fine-tuned on experimental values from PKAD-1 and PKAD-2 under a strict train-test similarity filter; the best predictor approaches 0.9 pKa units of accuracy, whereas heads built on ESM-2 or GearNet embeddings could not beat the null model at all. Ensembles of intrinsically disordered peptides degraded as expected for structures unlike anything in training, which fine-tuning on one 100-structure ensemble repaired.

#Applications

The obvious use is annotating large structure sets — predicted proteomes, simulation ensembles, or design libraries — with properties previously too expensive to compute in bulk: diffusion behavior for interpreting biophysical measurements, hydrodynamic radii for comparison against SAXS, and per-residue pKa values for setting protonation states before a simulation. The repository ships pretrained weights with documented paths for predictions and for extracting embeddings, so the network can also serve as a feature extractor for a downstream model.

#Impact

GSnet makes a specific, testable point about protein representation learning: for properties that depend on shape, surface, and charge distribution, a network pretrained on physics-derived supervised targets carries information that sequence language models and self-supervised structure encoders do not, and the difference is large rather than marginal. The caveats are equally concrete. The solvation free energy head is tied to the structural idiom of AlphaFold predictions, the geometric heads degrade on disordered and very short chains, transfer requires the training domain to cover the application domain, and no license is attached to the code or weights.