e202642: SmartSoluPred: A novel two-stage graph-language modeling framework for intrinsic and…

Student video submission for BioTreks manuscript e202642, 'SmartSoluPred: A novel two-stage graph-language modeling framework for intrinsic and mutation-induced protein solubility.' Learn more at https://biotreks.org/e202642/ Abstract: Protein solubility is a major bottleneck in biologic drug development, particularly for subcutaneous delivery, where poor solubility causes precipitation, inflammation, and dosing limitations. Computational prediction of solubility could reduce the cost and time of experimental screening, but remains difficult because solubility depends on complex interactions between amino acid sequence, structure, and local aggregation. While molecular dynamics simulations can model these interactions directly, they are computationally expensive, motivating the development of faster machine learning approaches. We present SmartSoluPred, a two-step machine learning pipeline: (1) predicting intrinsic protein solubility directly from sequence and structure (continuous percentage values), and (2) modeling mutation-induced changes (solubilizing or desolubilizing) in solubility to guide rational protein engineering. For intrinsic solubility prediction, we construct AlphaFold-derived residue graphs from the eSOL dataset, incorporating Aggrescan3D aggregation features, a novel 400-dimensional function embedding, and 640-dimensional pre-trained ESM-2 (Evolutionary Scale Modeling) model, and fine-tuned them using an Feed Forward Network on the eeSOL dataset. Our graph neural network (GNN) is based on the DeepMind graph network, a 3-fold update function in each layer to update global, node, and edge features. Fine-tuning ESM-2 achieves our best intrinsic solubility predictions (R²=0.479). . To our knowledge, this work is the first to use edge, node, and global embeddings to accurately predict protein solubility. For mutation-induced solubility changes, we fine-tune protein language embeddings to predict the effect of mutations on solubility using the SoluProtMutDB dataset. We fine-tune ESM-2 (35 M Parameters) using a Siamese architecture with a shared encoder for wild-type and mutant sequences, comparing embeddings through a binary classifier with focal loss to address the desolubilizing-skewed class imbalance. We benchmark against the current state-of-the-art Siamese graph convolutional neural networks on the same-protein evaluation using 5-fold CV. Our 1D-sequence-only model achieves 34.6% mean informedness, outperforming the current IEConv 3D-based state-of-the-art (33.2%) while eliminating dependence on structural data , and achieves 14.6% informedness for cross-protein generalization. However, the low cross-protein specificity of 26.5% indicates the model is best suited as a triage tool to prioritize candidates for further experimental validation, rather than as a standalone filter. These results demonstrate how our model matches or exceeds graph neural networks at mutation solubility prediction, while avoiding the computational and accessibility constraints of 3D structural data. Together, our work advances scalable, structure-aware, and mutation-sensitive solubility modeling, bridging 1D sequence embeddings and 3D structural representations for rational protein engineering.