Protein solubility is a key consideration for the production of proteins at industrial scales. Due to the complex and nonlinear nature of the physical processes involved in protein solubility, existing models exhibit poor predictive performance, necessitating costly trial-and-error experiments. We introduce TopoFormer, a novel combined topological-geometric equivariant neural network model capable of predicting binary protein solubility from OpenFold predicted 3D structural data. To train the TopoFormer model, the 3DProtSolDB structural dataset was created from OpenFold predictions of experimental protein solubility measurements in the literature. Concurrently, we assembled the 3DProtSolDB-Expt dataset from corresponding structures in the RCSB PDB. The 3DProtSolDB-Expt dataset was used to train TopoCoder, a topological embedding model that captures secondary structure shape information. The resultant embeddings and corresponding 3DProtSolDB structures were used to train the TopoFormer model. The trained TopoFormer model achieves state-of-the-art accuracies on a held-out test set. Additionally, model interpretability analyses reveal that the TopoFormer model acquires key physical insights into solvation behavior. These techniques are used to analyze the solubility of carbonic anhydrase mutants as reported in the literature. We find that metrics from the TopoFormer model correlate well with water survival probabilities from mo-lecular dynamics simulations, highlighting the utility of the trained model in guiding the rational design of soluble mutants.
The development of Graph Neural Networks for the task of predicting molecular properties has gained a great deal of attention as it typically allows the correlation of quick to compute atomic and bond descriptors with overall molecular properties. With the raising interest in photochemistry and photocatalysis as sustainable alternatives to thermal reactions, curation of virtual databases of computed photophysical properties for training of machine learning models has become popular. Unfortunately, current efforts fail to consider the exciton localization onto different chromophores of the same molecule, leading to potentially large prediction errors. Here we describe a molecular fragmentation strategy that can be used to overcome this limitation, while also providing a way to compare structural diversity between different libraries. Using a newly generated a database of 46,432 adiabatic S0-T1 energy gaps (ALFAST-DB), we compare its diversity against two datasets from the literature and demonstrate that a fragment-based delta learning approach improves model generalizability while achieving accuracies matching those of traditional message passing graph neural network architectures (MPGNN)
Interpretable protein solubility prediction from predicted protein structures is enabled through the fusion of geometric and topological protein representations.