Prot2Prop

What is Prot2Prop?

Protein developability is rarely controlled by one property. A sequence can look stable yet express poorly, remain soluble but aggregate under another condition, or fail during protein material production. Prot2Prop screens six such outcomes from amino acid sequence in one run: aggregation propensity, expression yield, folding stability, material production, solubility, and stability at 65 °C.

The model combines the structure-aware sequence representations learned by ProstT5 with a multitask prediction architecture. It returns three continuous ProteinGym-derived scores and three binary classifications. All predictions describe the complete submitted sequence. They do not identify individual residues or replace measurements under a specific expression, purification, formulation, or storage condition.

How to use Prot2Prop online

Prot2Prop runs online from one raw amino acid sequence or a multi-record FASTA file. ProteinIQ submits the sequences to the multitask ProstT5 model and returns one row per sequence, including three relative regression scores, three binary classifications, per-class probabilities, and a downloadable prediction CSV.

Input

Prot2Prop accepts the 20 standard amino acid codes plus U, Z, O, B, and X. Input is converted to uppercase and whitespace is removed. Before ProstT5 encoding, U, Z, O, and B are represented as X, so variants that differ only at these ambiguous or rare symbols receive the same encoded residue at those positions. Stop symbols such as * are rejected.

There are no scientific settings to configure. The same checkpoint and six task heads run for every sequence.

Results

The result table begins with Name and Sequence, followed by the six tasks in the model's native order.

Regression columns

Each regression task has two columns:

ProteinGym orients DMS_score so that higher values represent higher measured fitness in the source assay. Prot2Prop retains that convention. For aggregation-related data, a higher score should not automatically be translated as "more aggregation"; it means better fitness under the assay's directionality. The three regression columns also use different task distributions, so an aggregation score of 1.0 cannot be compared numerically with an expression or folding score of 1.0.

These values are most useful for ranking related candidates, such as variants of one scaffold scored together. Absolute cutoffs such as "stable above 0" or "high yield above 1" are not supported by the training labels.

Classification columns

Each classification task returns four columns:

The two class probabilities sum to approximately 1. Values near 0.5 indicate a borderline call, while values closer to 0 or 1 indicate stronger model separation. They are model probabilities, not measured frequencies of success or failure. The predicted class uses the larger of the two probabilities, equivalent to a 0.5 boundary for class 1.

The material-production label deserves special attention. Its positive class is failure, unlike the positive solubility and temperature-stability classes. A sequence with material_production_positive_probability = 0.90 is predicted to have a high probability of failure at that stage, not a 90% probability of successful production.

Downloadable output

prot2prop_predictions.csv contains the same 20 columns shown in the spreadsheet:

• name and sequence
• Raw-scale and normalized values for the three regression tasks
• Predicted class, positive probability, and class 0 and class 1 probabilities for the three classification tasks

The CSV preserves FASTA input order and can be used to rank a batch, compare designed variants, or join predictions back to an experiment table.

How Prot2Prop works

Prot2Prop uses the encoder from ProstT5, a protein language model trained to connect amino acid sequences with Foldseek 3Di structural tokens. The frozen encoder converts each submitted sequence into residue embeddings. Prot2Prop then applies:

• A shared adapter that learns signals useful across all six properties
• A small residual adapter for each property
• Attention pooling that reduces residue embeddings to one task-specific sequence representation
• Separate multilayer prediction heads for regression and binary classification

The three regression heads were trained on ProteinGym substitution assays grouped as aggregation propensity, expression yield, and folding stability. The model learns the direction-normalized DMS_score for each group. The classification heads use material-production, DeepSol, and 65 °C temperature-stability datasets.

This shared architecture makes Prot2Prop efficient for broad candidate screening, but it does not mean that one property causes another. Each task has its own adapter and output head, and correlations between predictions can reflect shared sequence features or shared biases in the training data.

Choosing Prot2Prop or a specialized predictor

Prot2Prop works well as an early funnel: rank complete sequences first, then apply a specialized predictor to the property that limits a candidate. For example, a batch with weak aggregation-related scores can be narrowed with Aggrescan3D once structures are available.

Practical limitations

• Sequence-level output: Prot2Prop does not locate problematic residues, model mutations explicitly, or account for a known three-dimensional conformation.
• Assay-relative regression scales: The three continuous values inherit heterogeneous DMS assay scales. They are prioritization scores, not laboratory units.
• Condition-specific labels: Solubility, expression, aggregation, and stability depend on host, construct boundaries, tags, buffer, concentration, pH, and temperature. Most of that context is absent from a sequence-only prediction.
• Material-production semantics: The positive class represents failure at the protein material stage. It is not a general estimate of manufacturability or production yield.
• Ambiguous residues: U, Z, O, and B are converted to X before encoding, reducing residue-specific information.
• Long proteins: The training pipeline was configured for sequences up to 2,048 tokens. Predictions for substantially longer proteins are outside that training envelope.
• No experimental uncertainty: The output has class probabilities but no confidence interval, replicate variance, or uncertainty estimate for the regression values.

The safest use is comparative screening within a related sequence family, followed by condition-matched experimental measurements. Large sequence changes, membrane proteins, multidomain constructs, and proteins that depend on cofactors or post-translational modifications require extra caution.