Screen paired antibody Fv sequences for high-concentration viscosity risk. Learn more
DeepViscosity online server
DeepViscosity classifies paired monoclonal antibody Fv sequences as low viscosity (class 0, at or below 20 cP) or high viscosity (class 1, above 20 cP) at 150 mg/mL. It returns the mean and standard deviation of the 102 model outputs together with the 30 DeepSP descriptors used for prediction. DeepViscosity does not predict an exact viscosity in cP.
ProteinIQ runs the fixed standalone DeepViscosity prediction workflow and retains its prediction table, descriptor table, FASTA files, IMGT numbering tables, aligned model input, diagnostics, submitted CSV, and provenance. For method context and a complete tutorial, see how to use DeepViscosity online.
Pricing
DeepViscosity costs a fixed 32 credits per job across the supported batch range. The exact credit price is shown before submission.
| Antibody rows | Credits |
|---|---|
| 1 | 32 |
| 16 | 32 |
| 100 | 32 |
Batch size does not change the price within the current 1 to 100-row limit. Splitting a larger table creates separate jobs, each charged independently.
Inputs
Submit exactly one comma-separated table by pasting CSV text or uploading a .csv file. The header names are case-sensitive.
| Column | Required | Description |
|---|---|---|
Name | Yes | Text identifier for the antibody pair. It cannot contain commas or whitespace. |
Heavy_Chain | Yes | Heavy-chain variable-domain amino-acid sequence. It must be recognized and numbered as an antibody Fv chain. |
Light_Chain | Yes | Kappa or lambda light-chain variable-domain amino-acid sequence. |
One job accepts 1 to 100 antibody rows and at most 1 MiB of CSV text. Required cells cannot be blank, and all three required columns must remain text when the CSV is parsed. Extra columns are accepted and ignored. Use variable-domain sequences without signal peptides or constant regions.
The built-in Source example antibody supplies one correctly formatted row named mAb1. It is useful for checking the input format before replacing the example with project sequences.
Settings
DeepViscosity has no adjustable scientific settings. The sequences determine the scientific result.
| Item | Fixed behavior |
|---|---|
| Antibody numbering | ANARCI with the IMGT numbering scheme |
| Model representation | Fixed 272-position paired Fv input, with 145 heavy-chain and 127 light-chain positions |
| Spatial descriptors | Three DeepSP models return 10 regional values each, for 30 descriptors |
| Viscosity ensemble | 102 artificial neural network models |
| Classification endpoint | Low at or below 20 cP and high above 20 cP, both at 150 mg/mL |
Job name | Optional job label; defaults to DeepViscosity prediction |
Outputs
The result page contains three views: Viscosity predictions, DeepSP descriptors, and Files. Tables can be copied or downloaded.
| File or result | Description |
|---|---|
DeepViscosity_classes.csv | Native prediction table containing Name, Prob_Mean, Prob_Std, and DeepViscosity_classes. |
DeepSP_descriptors.csv | Native table containing Name and all 30 spatial descriptors. |
seq_H.fasta | Heavy-chain FASTA generated from the submitted table. |
seq_L.fasta | Light-chain FASTA generated from the submitted table. |
seq_aligned_H.csv | ANARCI heavy-chain IMGT numbering table. |
seq_aligned_KL.csv | ANARCI kappa or lambda light-chain IMGT numbering table. |
seq_aligned_HL.txt | Fixed 272-position paired input used by the DeepSP models. |
DeepViscosity_input.csv | Exact CSV text submitted to the prediction program. |
deepviscosity_provenance.json | Source revision, model and runtime identity, input digest, execution timing, and output-file digests. |
deepviscosity_source.log | DeepViscosity, ANARCI, TensorFlow, and Keras diagnostics. |
run.log | Human-readable execution summary with row count, model count, descriptor count, timing, and completion status. |
The prediction rows and descriptor rows can also be passed to compatible workflows as structured table-row outputs. The native CSV, FASTA, numbering, and aligned-input files remain available as file outputs.
Understanding results
| Prediction column | Interpretation |
|---|---|
Name | Identifier copied from the submitted row. |
Prob_Mean | Mean of the 102 neural-network probability outputs. Values closer to 1 support the high-viscosity class; values closer to 0 support the low-viscosity class. |
Prob_Std | Standard deviation across the 102 model outputs. It describes model-to-model dispersion, not experimental uncertainty. |
DeepViscosity_classes | Binary result: 0 for at or below 20 cP and 1 for above 20 cP at 150 mg/mL. |
The source assigns class 1 when Prob_Mean is at least 0.5; otherwise it assigns class 0. Prob_Mean is not an exact viscosity, a calibrated probability of development success, or a substitute for a measurement under the intended formulation conditions. Prob_Std has no published universal cutoff for confidence.
The descriptor table contains three families across 10 antibody regions:
| Descriptor family | Meaning in the DeepSP model |
|---|---|
SAP_pos | Positive spatial aggregation propensity |
SCM_neg | Negative spatial charge map |
SCM_pos | Positive spatial charge map |
Each family is reported for CDRH1, CDRH2, CDRH3, CDRL1, CDRL2, CDRL3, combined CDR, heavy variable domain Hv, light variable domain Lv, and paired variable fragment Fv. These are model-derived features without published physical units or universal pass/fail thresholds.
The model uses Fv sequence only. It does not encode constant-region isotype or subclass, formulation composition, pH, excipients, temperature, concentration changes, or experimental handling. Confirm important decisions with viscosity measurements under the intended conditions.
