Pro4S Regression predicts quantitative protein solubility (0–1 scale) from protein sequence plus 3D structure, combining ESM-3B sequence embeddings, AlphaFold-derived structure graphs, and MaSIF surface descriptors in a unified GNN. The API accepts up to 4 items per request, each with an amino acid sequence and either PDB or CIF content for a single chain. The service returns continuous solubility scores (R² ≈ 0.56 on eSOL, AUC ≈ 0.90 on internal tests) and optional labels for screening, E. coli expression support, and de novo design.
Predict¶
Predict solubility scores for given protein sequences using either PDB or CIF structure data.
- POST /api/v3/pro4s-regression/predict/¶
Predict endpoint for Pro4S Regression.
- Request Headers:
Content-Type – application/json
Authorization – Token YOUR_API_KEY
Request
params (object, optional) — Configuration parameters:
batch_size (int, fixed: 4) — Maximum number of items processed per request
max_sequence_len (int, fixed: 2048) — Maximum allowed sequence length
items (array of objects, min: 1, max: 4, required) — Input proteins:
sequence (string, min length: 1, max length: 2048, required) — Amino-acid sequence with unambiguous standard residues
pdb (string, min length: 1, max length: 5000000, optional) — PDB-formatted structure content for the specified chain; either pdb or cif must be provided, but not both
cif (string, min length: 1, max length: 5000000, optional) — PDB-formatted structure content in mmCIF text form for the specified chain; either pdb or cif must be provided, but not both
chain_id (string, min length: 1, max length: 1, default: “A”) — Chain identifier within the provided structure
Example request:
- Status Codes:
200 OK – Successful response
400 Bad Request – Invalid input
500 Internal Server Error – Internal server error
Response
results (array of objects) — One result per input item, in the order requested:
solubility_score (float, range: 0.0–1.0) — Predicted normalized solubility score for the input protein
solubility_label (string, optional) — Predicted discrete solubility class label for the input protein
Example response:
Performance¶
Numerical behavior of the regression head:
solubility_scoreis a calibrated, monotonic transform of the underlying quantitative regression output, constrained to[0.0, 1.0]Higher values indicate higher predicted soluble fraction / expressibility; values near 0.5 usually correspond to borderline or uncertain solubility
Predictive accuracy on quantitative solubility benchmarks (internal BioLM evaluation reproducing Pro4S paper settings):
eSOL regression (E. coli PURE, N=660): R² ≈ 0.56, RMSE ≈ 0.214, exceeding structure-aware GNN baselines such as GATSol (R² ≈ 0.52, RMSE ≈ 0.223) and GraphSolEnsemble (R² ≈ 0.50)
Yeast PURE-system external regression set (N=108): R² ≈ 0.43 vs. ≈ 0.36–0.42 for strong graph baselines, indicating improved cross-organism generalization
De novo EGFR binder expressibility (binary labels, N=147) when thresholding the regression score: AUC ≈ 0.78 and accuracy ≈ 0.78, outperforming GATSol (AUC ≈ 0.72, accuracy ≈ 0.75) and sequence-only tools (e.g., SoluProt, NetSolP) by roughly 8–15 percentage points in AUC
Comparative performance vs. other BioLM solubility / developability models:
Versus sequence-only models (SoluProt, BioLMSol): Pro4S Regression provides more accurate ranking whenever a reasonable structure (e.g., AlphaFold2, ESMFold) is available; in internal enzyme-engineering panels with AlphaFold2 models, it typically increases Spearman correlation to wet-lab solubility by 0.10–0.20 absolute over BioLMSol and yields higher enrichment of top-soluble designs in the top 10–20% score bin
Versus structure-only GNN solubility models (e.g., GATSol deployments): multimodal fusion of ESM-3B sequence embeddings with 3D structure and MaSIF-style surface features gives a consistent edge of ≈0.04–0.05 R² on eSOL-like datasets, with ablations showing that removing surface information or contrastive alignment degrades both R² and AUC
Score calibration and use in experimental design:
On PURE-system assays used for training, scores ≥≈0.7 are typically associated with high expression probability, whereas scores ≤≈0.3 are enriched for non-expressing / poorly soluble constructs
In internal BioLM campaigns, triaging libraries by retaining only the top 30–40% of Pro4S Regression scores yields ≈1.3–1.5× enrichment for experimentally soluble / expressible designs relative to SoluProt-based filters at comparable recall, and substantially reduces the fraction of completely non-expressing de novo designs while retaining most highly expressed variants
Applications¶
Upfront triage of de novo designed binders and other recombinant proteins for expression campaigns by ranking candidates on predicted solubility from the regression variant, helping reduce non-expressing designs while retaining most highly expressing variants and thereby cutting wet-lab screening load in high-throughput discovery programs
Optimization of recombinant protein constructs for E. coli expression by comparing regression scores across N-/C-terminal truncations, domain boundaries, and tag variants (e.g. His-tag, solubility tags, linkers), using Pro4S’s fused sequence–structure–surface representation to prioritize constructs with higher predicted solubility in the target host system before cloning and purification
Developability risk assessment for therapeutic and industrial protein leads by comparing regression solubility scores across variants or mutational scans to flag sequences or regions with reduced predicted solubility, supporting lead selection with better formulation and manufacturability prospects; not a replacement for dedicated aggregation, viscosity, or immunogenicity studies but an early, solubility-focused filter
In silico library pruning for protein engineering campaigns (e.g. stability, binding, or activity optimization) by applying regression scores to remove variants predicted to be poorly soluble while retaining most high-solubility candidates, increasing the fraction of expressible proteins in downstream assays without biasing toward a particular function and avoiding screens on sequences unlikely to express
Transfer-learning–inspired workflows via choosing different Pro4S API variants (classification, regression, finetune) to match available data and design goals, using the finetuned model when E. coli expression-like behavior is desired; note that all API variants are trained on published datasets and may be less reliable for proteins, hosts, or conditions far outside those domains
Limitations¶
Batch Size: Each request to the Pro4S Regression API must include between 1 and 4 items in
items(Pro4SPredictRequest.itemshasmin_length=1andmax_length=4). Larger batches must be split client-side; responses are returned in the same order as the request.Maximum Sequence Length: Each
sequencemust be a non-empty amino-acid string of length 1–2048 characters (max_sequence_len=2048). Longer protein sequences require truncation or tiling strategies and may not be suitable for direct use with this endpoint.Required Structure and Chain Handling: Every
Pro4SPredictRequestItemmust include exactly one ofpdborcifcontaining a single structure (min_length=1,max_length=5000000), plus a one-characterchain_id(default"A"). Multi-chain complexes, ambiguous chain IDs, or mismatchedsequence/structure pairs are not supported and can lead to invalid or misleadingsolubility_scorepredictions.Model Scope and Training Domain: Pro4S Regression is trained primarily on soluble/insoluble expression in E. coli and solubility measured in cell-free systems (e.g., eSOL, PURE). The continuous
solubility_score(0.0–1.0) reflects relative aqueous solubility under those conditions only; it does not model formulation behavior, aggregation kinetics, long-term stability, or solubility in other hosts (CHO, yeast, plants, etc.). Use on membrane proteins, large multi-domain complexes, highly disordered proteins, or non-protein polymers falls outside the intended domain.Dependence on 3D Structure Quality: Predictions rely on structural and surface features. Low-confidence or inconsistent input structures (e.g., coarse models, missing loops, truncated termini, or inconsistent protonation/charge states) can degrade prediction quality. For early-stage high-throughput design with uncertain structures, sequence-only solubility models may be more efficient, with this endpoint better suited to later-stage ranking of a smaller set of high-confidence designs.
Use Cases Where Another Model May Be Preferable: This regression variant is optimized for continuous solubility scoring. If you only need binary solubility/expressibility labels, the
classificationorfinetunePro4S variants are generally more appropriate. For very large sequence libraries without reliable structures, faster sequence-only predictors or simple sequence heuristics (e.g., charge/composition filters) are typically better for first-pass filtering, with Pro4S Regression reserved for structure-annotated candidates where more expensive multimodal inference is justified.
How We Use It¶
Pro4S regression enables teams to rank protein variants by predicted quantitative solubility and expected expression behavior, so it functions as a central developability filter alongside generative design models, stability and aggregation predictors, and structure-based metrics. In practice, Pro4S scores are fed into multi-objective ranking pipelines for antibody and enzyme campaigns, combined with language-model embeddings and 3D descriptors to down-select libraries before synthesis, then updated with experimental solubility/expression data to refine subsequent design rounds. Standardized, batch-oriented API access allows solubility scoring to be embedded directly into automated workflows (for example, during sequence generation with paired structures, structure triage, or portfolio-level risk assessment), which reduces the fraction of non-expressing or poorly behaving candidates and accelerates progression of viable leads into experimental validation.
Integrates with generative design, stability/aggregation predictors, and structure-based property models in multi-parameter developability optimization.
Supports scalable, batched evaluation of up to 4 sequence–structure variants per request so expression resources focus on candidates whose solubility profiles align with downstream formulation and manufacturing goals.
References¶
Qian, J., Yang, L., Wang, R., & Qi, Y. Pro4S: prediction of protein solubility by fusing sequence, structure, and surface. Manuscript in preparation. No journal, DOI, or public preprint are available at this time.
