SoluProtMut

Protein solubility governs how much recombinant protein a fermentation can yield and whether a biologic can be dosed at therapeutic concentration, which makes it a routine engineering target. Predicting how a substitution shifts it has nevertheless lagged behind the neighbouring problem of stability prediction: the labelled data are scarce, split across incompatible assays, and biased toward a handful of proteins, and existing mutational predictors have plateaued near a 70% correct-prediction ratio.

SoluProtMut, posted as a preprint in September 2025 by Loschmidt Laboratories at Masaryk University together with collaborators at TU Wien and the University of Greifswald, attacks that data problem from the opposite direction. Rather than pooling low-throughput measurements from many proteins, it is trained exclusively on deep mutational scanning (DMS) data for three proteins, then applied without refitting to proteins it has never seen. The model is a Siamese pair of geometric graph networks that each score one variant structure; the classification comes from the difference between the two scores, which makes the predictor anti-symmetric by construction — reversing a mutation reverses its predicted sign — instead of coaxing that property out of a regularization term in the loss.

The work also contributes new wet-lab data: solubility effects for 277 single-point and multipoint mutants of haloalkane dehalogenase LinB, measured in-house with a split-GFP assay and combined with 1,037 published myoglobin mutants into an external validation set assembled after training.

#Key Features

• Anti-symmetric Siamese design: Two weight-shared base networks score the wild-type and mutant variants, and the prediction is taken from their difference, enforcing sign consistency architecturally rather than through a loss penalty.
• No mutant structure required: The network consumes the wild-type C-α backbone paired with the mutant sequence, so it needs no side-chain modelling or structure relaxation for the variant.
• Trained only on high-throughput data: The released model is fit exclusively to DMS measurements, avoiding the protein-, position-, and substitution-specific biases that alanine scans and site-directed mutagenesis records introduce.
• Strongest on multipoint mutants: Despite being trained solely on single-point substitutions, it leads the field on combinatorial variants, the case most relevant to design campaigns.
• Cheap inference: Prediction takes seconds per mutation on a CPU, with no GPU required.

#Technical Details

The base network is a graph convolutional network derived from the IEConv architecture for 3D protein structures. Each residue enters as a C-α point carrying an eight-dimensional VHSE physicochemical descriptor; on-the-fly multigraphs connect vertices within a Euclidean radius that grows from 8 to 16 Å across three convolution-and-pooling rounds, with kernels parameterized by both sequential and Euclidean distance, giving SE(3) invariance. Residual bottleneck blocks feed a two-layer perceptron head that emits a scalar solubility score. The network has 2.6 million trainable parameters, and the deployed predictor averages five base-model pairs from five-fold cross-validation. Training data come from SoluProtMutDB (32,992 datapoints across 103 proteins, over 97% of them DMS-derived), narrowed to four DMS assays covering pyrrolidine ketide synthase, levoglucosan kinase, and TEM β-lactamase; the largest split holds 15,741 training points and takes 1.5 days to train on a single GTX 1080 Ti.

Evaluated with informedness (sensitivity + specificity − 1), the best model reaches 13.4% across the full held-out non-DMS set, ahead of OptSolMut (13.0%), CamSol (10.1%), and SoluProt (10.0%). On multipoint mutants it scores 26.5% informedness, 8.8 percentage points above the next-best method, with 80.4% precision on the desolubilizing class. Generalization to the two external proteins is weaker: the best informedness is 13.8% on LinB, and on the filtered myoglobin set the earlier PON-Sol2 predictor leads at 23.7%. Confidence intervals are wide throughout, a direct consequence of how little mutational solubility data exist.

#Applications

The model suits protein engineers triaging variant libraries before expression — flagging substitutions likely to push a construct into inclusion bodies, and ranking combinatorial designs, where its margin over sequence-based tools is largest. Because it scores from a wild-type backbone and runs on a CPU, it slots into enzyme and biologics design loops without specialized hardware. Its precision on the desolubilizing class makes it most useful as a negative filter: avoiding detrimental designs rather than nominating solubility-enhancing ones.

#Impact

SoluProtMut is the first mutational solubility predictor trained solely on deep mutational scanning data, showing that high-throughput proxies can outperform the curated low-throughput records the field has relied on, and its differential Siamese construction is a reusable recipe for any anti-symmetric mutational task. Accuracy nonetheless remains moderate: the authors state plainly that generalization to unseen proteins is limited, performance varies substantially by protein, and yeast-derived training subsets help on yeast-assayed targets while hurting elsewhere. The work is a preprint and has not been peer reviewed. Weights and inference code are distributed through a HuggingFace Space rather than a versioned source repository, and neither the code, the weights, nor the preprint carries a reuse license.