DeepSP

DeepSP online server

The DeepSP online server predicts 30 antibody spatial descriptors from paired heavy-chain and light-chain variable-domain sequences. ProteinIQ accepts a CSV of antibody pairs and returns spatial aggregation propensity and spatial charge map descriptors, numbering tables, aligned model inputs, FASTA files, logs, and provenance. No local installation or model setup is required.

The published models were trained using 20,530 antibody sequences. Across the 30 descriptors, the study reported an average correlation of 0.87 between DeepSP predictions and descriptors calculated from molecular dynamics trajectories, with individual correlations ranging from 0.76 to 0.96. This validation applies to the modeled spatial descriptors, not to direct experimental stability measurements.

Pricing

DeepSP jobs start at 20 credits. The calculator scales with the number of antibody pairs in the submitted CSV, with larger batches receiving sublinear pricing. The exact credit price is calculated before submission.

Inputs

The CSV header is case-sensitive. Extra columns are accepted and ignored.

Name,Heavy_Chain,Light_Chain
antibody_1,EVQLVESGGGLVQPGGSLRLSC...,DIQMTQSPSSLSASVGDRVTITC...

One job accepts up to 100 antibody rows and 1 MiB of CSV text. Larger datasets can be divided across multiple jobs. Sequence fields must remain valid text values when parsed from CSV, and blank required values are rejected before submission.

Settings

DeepSP has no adjustable scientific settings in ProteinIQ. It uses the published fixed inference workflow:

Outputs

The numbering and alignment files are especially useful when a sequence produces unexpected descriptor values. They show how the submitted variable domains were mapped into the fixed model representation. ANARCII can be used separately when a more focused numbering analysis is needed.

How DeepSP works

DeepSP first numbers the paired variable domains with ANARCII using the IMGT scheme. The numbered heavy and light chains are mapped into a fixed 272-position representation and one-hot encoded with 21 residue symbols. Three convolutional neural networks then predict 10 regional values each, producing 30 descriptors in total.

SAP_pos is the source model's positive spatial aggregation propensity descriptor family. SCM_neg and SCM_pos are its negative and positive spatial charge map descriptor families. The regional suffix identifies the source-defined antibody region summarized by each value. Hv, Lv, and Fv refer to the heavy variable domain, light variable domain, and paired variable fragment.

Interpreting DeepSP descriptors

DeepSP values are model-derived spatial features. The source does not publish physical units, universal thresholds, confidence intervals, or a general higher-is-better direction for the 30 outputs. Values are most useful as consistent features for comparison or downstream modeling when the same input preparation and model edition are used across the compared antibodies.

The publication demonstrates that these sequence-based predictions reproduce molecular-dynamics-derived spatial descriptors and uses them as features in downstream aggregation modeling. It does not establish each descriptor as a direct measurement of experimental stability, aggregation rate, viscosity, or solubility. Interpretation should include sequence context, numbering quality, formulation conditions, and relevant experimental evidence.

Important limitations

• Predictions depend on successful ANARCII numbering and the fixed 272-position representation. Unusual or incomplete variable-domain sequences should be reviewed with the numbering tables and source diagnostics.

• Duplicate values in the Name column can collapse into one result row because DeepSP groups sequence records by name. Unique identifiers are required when every submitted row must remain distinct.

• The hosted job limit is 100 antibody pairs and 1 MiB of CSV text. This is a ProteinIQ execution limit, not a scientific limit of the DeepSP models.

• DeepSP does not return per-prediction confidence scores or uncertainty estimates.

• Results from different model editions, numbering implementations, or preprocessing pipelines should not be assumed to be directly interchangeable without validation.