DeepViscosity

Ensemble deep learning model for classifying monoclonal antibody viscosity at high concentration (150 mg/mL) as low or high from VH/VL Fv sequences.

License: MIT · Molecules: protein · Tasks: property_prediction

Actions: predict · Variants: 1

At a glance

Use it when

• You are screening a panel of therapeutic antibody candidates and need to identify which ones are likely to have high viscosity at subcutaneous injection concentrations (100-200 mg/mL)
• You are in early-stage antibody discovery and want to de-risk viscosity liability before any protein expression or experimental viscosity measurement
• You want to understand which antibody regions (CDR-H3, framework) are driving predicted viscosity through DeepSP spatial property analysis
• You need rapid, inexpensive viscosity pre-screening to prioritize which candidates warrant experimental cone-and-plate rheometry
• You are developing subcutaneous antibody formulations and need to flag candidates likely to exceed the 20 cP injectability threshold

Strengths

• Only open-source, sequence-only viscosity classifier specifically designed for monoclonal antibody high-concentration formulations (150 mg/mL threshold at 20 cP)
• Ensemble of 102 ANN models provides uncertainty estimates (probability_std) that flag borderline predictions for experimental follow-up, enabling confidence-based triage
• 87.5% accuracy on the independent Lai_mAb_16 test set (14/16 correctly classified), demonstrating reliable binary viscosity classification
• CPU-only inference with minimal resources (1 CPU, 2 GB RAM) -- one of the lightest models in this catalog, enabling rapid screening of large antibody panels
• Optional DeepSP spatial property output (30 features across 10 antibody domains) reveals which CDR or framework regions drive high viscosity, guiding targeted engineering
• Requires only VH and VL amino acid sequences -- no 3D structure, no experimental data, no expression needed for initial screening
• MIT license enables unrestricted commercial use in pharmaceutical antibody development pipelines

Limitations

• Antibody-only model -- ANARCI alignment requires antibody variable domains; non-antibody proteins (enzymes, receptors) cannot be processed and will fail
• Single Fv pair input only -- cannot model bispecific antibodies, nanobodies (VHH), or IgG formats beyond the Fv region; Fc effects on viscosity are not captured
• Training data from AstraZeneca humanized/human IgGs -- performance may degrade on antibodies from non-human sources or unusual germlines
• Binary classification (low vs high viscosity at 20 cP threshold) -- does not predict continuous viscosity values for more nuanced formulation decisions
• Cannot suggest specific mutations to reduce viscosity -- serves as a scoring function only, not a design tool
• Formulation effects (excipients, buffer composition, ionic strength) are not modeled; predictions reflect sequence-intrinsic viscosity risk only

Reach for something else when

• Your molecule is not a monoclonal antibody with paired VH and VL chains -- DeepViscosity requires IMGT-numberable antibody variable domains
• You are working with nanobodies, single-domain antibodies, or bispecific formats -- these are not supported by the paired VH/VL input requirement
• You need to predict general protein solubility rather than antibody-specific viscosity -- DeepViscosity is antibody-specific; use a general solubility predictor such as SoluProt instead
• You need continuous viscosity values rather than binary low/high classification -- DeepViscosity does not output viscosity in centipoise
• You need to account for formulation conditions (pH, excipients, concentration) in your viscosity prediction -- the model predicts sequence-intrinsic risk only
• You need predictions for non-human antibody formats that may not align well with IMGT numbering

API & schema

Call an action with POST /api/v1/{slug}/{action} — the request envelope is {"items": [...], "params": {...}} and a success returns {"results": [...]}. See the HTTP API page for the base URL, error shape, and full contract.

predict

Call it

curl -X POST http://127.0.0.1:8000/api/v1/deepviscosity/predict \
  -H "Content-Type: application/json" \
  -d '{
  "items": [
    {
      "heavy_chain": "MKTAYIAKQRQISFVKSHFSRQLEERLGLIE",
      "light_chain": "MKTAYIAKQRQISFVKSHFSRQLEERLGLIE"
    }
  ]
}'

Request — DeepViscosityPredictRequest

Field	Type	Required	Constraints	Description
params	DeepViscosityPredictRequestParams | null	no		Optional parameters controlling this action (defaults are used when omitted).
items	list[DeepViscosityPredictRequestItem]	yes	items 1–10	Batch of inputs to process in a single request. Up to 10 antibodies per request.

Nested types

DeepViscosityPredictRequestItem

Field	Type	Required	Constraints	Description
heavy_chain	string	yes	len 50–200	Heavy chain variable region (VH) Fv sequence.
light_chain	string	yes	len 50–200	Light chain variable region (VL) Fv sequence.

DeepViscosityPredictRequestParams

Field	Type	Required	Constraints	Description
include_deepsp_features	boolean	no	default False	Include 30 DeepSP spatial property features in response.

Raw JSON Schema

{
  "$defs": {
    "DeepViscosityPredictRequestItem": {
      "additionalProperties": false,
      "description": "Single antibody item for DeepViscosity prediction.",
      "properties": {
        "heavy_chain": {
          "description": "Heavy chain variable region (VH) Fv sequence.",
          "maxLength": 200,
          "minLength": 50,
          "title": "Heavy Chain",
          "type": "string"
        },
        "light_chain": {
          "description": "Light chain variable region (VL) Fv sequence.",
          "maxLength": 200,
          "minLength": 50,
          "title": "Light Chain",
          "type": "string"
        }
      },
      "required": [
        "heavy_chain",
        "light_chain"
      ],
      "title": "DeepViscosityPredictRequestItem",
      "type": "object"
    },
    "DeepViscosityPredictRequestParams": {
      "additionalProperties": false,
      "description": "Parameters for DeepViscosity prediction request.",
      "properties": {
        "include_deepsp_features": {
          "default": false,
          "description": "Include 30 DeepSP spatial property features in response.",
          "title": "Include Deepsp Features",
          "type": "boolean"
        }
      },
      "title": "DeepViscosityPredictRequestParams",
      "type": "object"
    }
  },
  "additionalProperties": false,
  "description": "DeepViscosity prediction request.",
  "properties": {
    "params": {
      "anyOf": [
        {
          "$ref": "#/$defs/DeepViscosityPredictRequestParams"
        },
        {
          "type": "null"
        }
      ],
      "default": null,
      "description": "Optional parameters controlling this action (defaults are used when omitted)."
    },
    "items": {
      "description": "Batch of inputs to process in a single request. Up to 10 antibodies per request.",
      "items": {
        "$ref": "#/$defs/DeepViscosityPredictRequestItem"
      },
      "maxItems": 10,
      "minItems": 1,
      "title": "Items",
      "type": "array"
    }
  },
  "required": [
    "items"
  ],
  "title": "DeepViscosityPredictRequest",
  "type": "object"
}

Response — DeepViscosityPredictResponse

Field	Type	Required	Constraints	Description
results	list[DeepViscosityPredictResponseResult]	yes		Per-input results, returned in the same order as the request items.

Nested types

DeepViscosityPredictResponseResult

Field	Type	Required	Constraints	Description
viscosity_class	string	yes	one of low, high	Predicted viscosity class: 'low' (<=20 cP) or 'high' (>20 cP).
probability_mean	number	yes	≥0.0; ≤1.0	Mean predicted probability across 102 ensemble models.
probability_std	number	yes	≥0.0	Standard deviation of predictions across ensemble.
is_high_viscosity	boolean	yes		True if probability_mean >= 0.5 (high viscosity predicted).
deepsp_features	object | null	no		DeepSP spatial properties (30 features) if requested.

Raw JSON Schema

{
  "$defs": {
    "DeepViscosityPredictResponseResult": {
      "description": "Single result from DeepViscosity prediction.",
      "properties": {
        "viscosity_class": {
          "description": "Predicted viscosity class: 'low' (<=20 cP) or 'high' (>20 cP).",
          "enum": [
            "low",
            "high"
          ],
          "title": "Viscosity Class",
          "type": "string"
        },
        "probability_mean": {
          "description": "Mean predicted probability across 102 ensemble models.",
          "maximum": 1.0,
          "minimum": 0.0,
          "title": "Probability Mean",
          "type": "number"
        },
        "probability_std": {
          "description": "Standard deviation of predictions across ensemble.",
          "minimum": 0.0,
          "title": "Probability Std",
          "type": "number"
        },
        "is_high_viscosity": {
          "description": "True if probability_mean >= 0.5 (high viscosity predicted).",
          "title": "Is High Viscosity",
          "type": "boolean"
        },
        "deepsp_features": {
          "anyOf": [
            {
              "additionalProperties": {
                "type": "number"
              },
              "type": "object"
            },
            {
              "type": "null"
            }
          ],
          "default": null,
          "description": "DeepSP spatial properties (30 features) if requested.",
          "title": "Deepsp Features"
        }
      },
      "required": [
        "viscosity_class",
        "probability_mean",
        "probability_std",
        "is_high_viscosity"
      ],
      "title": "DeepViscosityPredictResponseResult",
      "type": "object"
    }
  },
  "description": "DeepViscosity prediction response.",
  "properties": {
    "results": {
      "description": "Per-input results, returned in the same order as the request items.",
      "items": {
        "$ref": "#/$defs/DeepViscosityPredictResponseResult"
      },
      "title": "Results",
      "type": "array"
    }
  },
  "required": [
    "results"
  ],
  "title": "DeepViscosityPredictResponse",
  "type": "object"
}

Usage

One-line summary: Ensemble deep learning model for classifying monoclonal antibody viscosity at high concentration (150 mg/mL) as low or high from VH/VL Fv sequences.

Overview

DeepViscosity is a two-stage ensemble deep learning pipeline developed by Kalejaye et al. (mAbs, 2025) for predicting whether a monoclonal antibody will exhibit low (<=20 cP) or high (>20 cP) viscosity at therapeutic concentration. It uses DeepSP CNN models to extract 30 spatial property features from sequence, followed by 102 ensemble ANN classifiers.

The model addresses a critical bottleneck in biopharmaceutical development: identifying antibody candidates with viscosity liabilities before expensive experimental characterization. It is the first open-source, sequence-only viscosity classifier with built-in ensemble uncertainty estimates.

Architecture

Property	Value
Architecture	3 Conv1D CNNs (feature extraction) + 102 ANN ensemble (classification)
Input	Paired VH + VL Fv sequences
Feature representation	(272, 21) one-hot encoded IMGT-aligned sequence
DeepSP features	30 spatial properties (SAP_pos, SCM_neg, SCM_pos x 10 domains)
ANN hidden layers	4 (128 -> 64 -> 32 -> 16, tanh activation)
Training data	229 mAbs (AstraZeneca, 150 mg/mL, 20 mM histidine-HCl pH 6.0)
Max sequence length	200 residues per chain
Min sequence length	50 residues per chain

See MODEL.md for detailed architecture specifications.

Capabilities & Limitations

CAN be used for: - Early-stage screening of therapeutic antibody candidates for manufacturability - Identifying high-viscosity mAbs before expensive formulation development - Guiding antibody engineering to reduce viscosity via DeepSP feature analysis - Batch prediction of up to 10 antibodies per request

CANNOT be used for: - Continuous viscosity value prediction (binary classification only) - Non-antibody proteins (requires ANARCI-alignable VH/VL sequences) - Sequences longer than 200 residues or shorter than 50 residues per chain - Predicting viscosity at concentrations other than 150 mg/mL - Predicting viscosity in buffer conditions other than 20 mM histidine-HCl pH 6.0 - Nanobodies or single-domain antibodies (requires paired VH/VL)

Other considerations: - Only predicts binary class (low/high), not continuous viscosity values - Trained on mAbs at 150 mg/mL in 20mM histidine-HCl pH 6.0 buffer - Does not capture effects of different isotypes (IgG1, IgG2, IgG4) or Fc region - Performance may degrade for mAbs with high sequence homology (as seen with Apgar_mAb_38 test set) - Requires valid antibody Fv sequences; will fail on non-antibody proteins

Usage Examples

from models.deepviscosity.schema import (
    DeepViscosityPredictRequest,
    DeepViscosityPredictRequestItem,
    DeepViscosityPredictRequestParams,
)

# Single antibody prediction
request = DeepViscosityPredictRequest(
    items=[
        DeepViscosityPredictRequestItem(
            heavy_chain="EVQLVESGGGLVQPGRSLRLSCAASGFTFDDYAMHWVRQAPGKGLEWVSAITWNSGHIDYADSVEGRFTISRDNAKNSLYLQMNSLRAEDTAVYYCAKVSYLSTASSLDYWGQGTLVTVSS",
            light_chain="DIQMTQSPSSLSASVGDRVTITCRASQGIRNYLAWYQQKPGKAPKLLIYAASTLQSGVPSRFSGSGSGTDFTLTISSLQPEDVATYYCQRYNRAPYTFGQGTKVEIK",
        )
    ]
)

# With DeepSP features included
request_with_features = DeepViscosityPredictRequest(
    params=DeepViscosityPredictRequestParams(include_deepsp_features=True),
    items=[
        DeepViscosityPredictRequestItem(
            heavy_chain="EVQLVESGGGLVQPGRSLRLSCAASGFTFDDYAMHWVRQAPGKGLEWVSAITWNSGHIDYADSVEGRFTISRDNAKNSLYLQMNSLRAEDTAVYYCAKVSYLSTASSLDYWGQGTLVTVSS",
            light_chain="DIQMTQSPSSLSASVGDRVTITCRASQGIRNYLAWYQQKPGKAPKLLIYAASTLQSGVPSRFSGSGSGTDFTLTISSLQPEDVATYYCQRYNRAPYTFGQGTKVEIK",
        )
    ],
)

# Batch prediction (up to 10 antibodies)
batch_request = DeepViscosityPredictRequest(
    items=[
        DeepViscosityPredictRequestItem(
            heavy_chain="EVQLVESGGGLVQPGRSLRLSCAASGFTFDDYAMH...",
            light_chain="DIQMTQSPSSLSASVGDRVTITCRASQGIRNYLA...",
        ),
        DeepViscosityPredictRequestItem(
            heavy_chain="EVQLVESGGGLVQPGGSLRLSCAASGFTFSDSWIH...",
            light_chain="DIQMTQSPSSLSASVGDRVTITCRASQDVSTAVA...",
        ),
    ]
)

Confidence Metrics

Metric	Range	Interpretation
probability_mean	0-1	<0.3: confident low viscosity; 0.3-0.7: uncertain; >0.7: confident high viscosity
probability_std	>=0	<0.1: high ensemble agreement; 0.1-0.2: moderate; >0.2: low confidence

Predictions with probability_mean near 0.5 and high probability_std should be validated experimentally.

Technical Glossary

Fv (Variable Fragment): The antigen-binding portion of an antibody, consisting of the VH and VL domains. This is the minimal unit required for antigen recognition.

CDR (Complementarity-Determining Region): Hypervariable loops within VH/VL that form the antigen-binding surface. There are 3 CDRs per variable domain (CDR H1-H3, CDR L1-L3).

IMGT Numbering: A standardized residue numbering scheme for immunoglobulins maintained by the International ImMunoGeneTics information system. Enables comparison of antibodies with different sequence lengths.

SAP (Spatial Aggregation Propensity): A measure of hydrophobic surface exposure calculated over antibody structural domains. Higher SAP_pos values indicate more exposed hydrophobic patches.

SCM (Spatial Charge Map): A measure of surface charge distribution. SCM_neg and SCM_pos capture negative and positive charge patches, respectively.

cP (centipoise): Unit of dynamic viscosity. Water at 20 C is ~1 cP. The 20 cP threshold represents a practical upper limit for subcutaneous injection through standard gauge needles.

Architecture & training

Architecture

Model Type & Innovation

DeepViscosity is a two-stage ensemble deep learning pipeline for binary classification of monoclonal antibody (mAb) viscosity at high concentration (150 mg/mL). The key innovation is combining DeepSP-derived spatial property features with a large ensemble of simple neural networks to achieve robust viscosity classification without requiring expensive molecular dynamics simulations or experimental measurements.

Stage 1 -- DeepSP CNN Feature Extraction: Three 1D convolutional neural networks (Conv1D) predict 30 spatial aggregation propensity (SAP) and spatial charge map (SCM) features from IMGT-aligned, one-hot-encoded antibody Fv sequences. Each CNN predicts 10 features across antibody structural domains (CDR H1-H3, CDR L1-L3, combined CDR, Hv, Lv, Fv).

Stage 2 -- Ensemble ANN Classification: 102 independently trained artificial neural networks (ANNs) each predict the probability of high viscosity from the 30 scaled DeepSP features. The final prediction is the mean probability across the ensemble, thresholded at 0.5.

This approach differs from prior viscosity models (e.g., CamSol, TAP) by learning spatial features directly from sequence rather than relying on 3D structure calculations or hand-crafted descriptors.

Parameters & Layers

DeepSP CNN Models (3 models):

Component	Details
Architecture	Conv1D regression
Input shape	(272, 21) -- one-hot encoded aligned Fv
Models	SAPpos, SCMneg, SCMpos
Output per model	10 spatial property values
Activation	ReLU (convolutional layers)
Loss function	Mean absolute error (MAE)

Ensemble ANN Models (102 models):

Component	Details
Architecture	Fully connected feedforward
Hidden layers	4 (128 -> 64 -> 32 -> 16 neurons)
Activation	tanh
Input	30 scaled DeepSP features
Output	1 sigmoid probability
Training method	Leave-one-group-out (LOGO) cross-validation
Total ensemble models	102

Total parameter count: Approximately 50K per ANN model, ~5.1M across all 102 ensemble models. CNN models are additional but smaller.

Training Data

Property	Details
Dataset	AstraZeneca proprietary mAb viscosity dataset
Size	229 monoclonal antibodies
Measurements	Viscosity at 150 mg/mL in 20 mM histidine-HCl pH 6.0
Class distribution	Low viscosity (<=20 cP) and high viscosity (>20 cP)
Split method	Leave-one-group-out (LOGO) cross-validation across 102 groups
External test sets	Lai_mAb_16 (n=16), Apgar_mAb_38 (n=38)

Known biases: - Training data is exclusively from AstraZeneca's mAb pipeline, which may over-represent certain antibody germline families - All measurements at a single buffer condition (20 mM histidine-HCl pH 6.0) - Binary classification threshold (20 cP) is specific to this dataset's distribution - Only Fv (variable) regions are used; Fc region contributions to viscosity are not captured

Loss Function & Objective

Stage 1 (DeepSP CNNs): Mean absolute error (MAE) for regression of spatial property values:

L_CNN = (1/n) * sum(|y_true - y_pred|)

Stage 2 (Ensemble ANNs): Binary cross-entropy for viscosity classification:

L_ANN = -(1/n) * sum(y*log(p) + (1-y)*log(1-p))

where y is the binary viscosity label (0=low, 1=high) and p is the predicted probability.

Tokenization / Input Processing

1. ANARCI alignment: VH and VL sequences are aligned to IMGT numbering using ANARCI with HMMER
2. Fixed-length extraction: 145 positions from heavy chain + 127 positions from light chain = 272 total positions. Positions not present in the alignment are filled with gap characters
3. One-hot encoding: Each position is encoded as a 21-dimensional vector (20 standard amino acids + gap), producing a (272, 21) matrix
4. Sequence constraints: VH/VL Fv regions must be 50-200 residues each

Performance & Benchmarks

Published Benchmarks

LOGO Cross-Validation (Training Set, n=229)

Model	Accuracy	AUC	Dataset
DeepViscosity (Ensemble)	85.2%	0.901	LOGO CV (n=229)
Single best ANN	78.6%	0.852	LOGO CV (n=229)
Random Forest (DeepSP features)	80.3%	0.871	LOGO CV (n=229)
Logistic Regression (DeepSP features)	76.4%	0.841	LOGO CV (n=229)

Independent Test Sets

Test Set	n	Accuracy	Misclassifications
Lai_mAb_16	16	87.5%	2/16
Apgar_mAb_38	38	68.4%	12/38
Combined external	54	74.1%	14/54

The Apgar_mAb_38 performance drop is attributed to high sequence homology among mAbs in that set, which the model struggles with.

BioLM Verification Results

Metric	Published	BioLM	Difference	Status
Lai_mAb_16 accuracy	87.5%	deferred	--	deferred (requires deployment)
Output probability range	[0, 1]	[0, 1]	0	PASS
Ensemble size	102 models	102 models	0	PASS
DeepSP features	30 features	30 features	0	PASS

Comparison to Alternatives

There are no direct antibody viscosity prediction models currently in this catalog. DeepViscosity occupies a unique niche.

Model	Task	Metric	When to prefer
DeepViscosity	Viscosity classification	87.5% accuracy (Lai test)	Screening mAb candidates for manufacturability
CamSol (not in this catalog)	Solubility prediction	Spearman rho ~0.5 vs viscosity	General protein solubility (not viscosity-specific)
TAP (not in this catalog)	Developability scoring	Correlated with viscosity	Broader developability assessment

Error Bars & Confidence

The 102-model ensemble provides built-in uncertainty quantification:

• probability_std: Standard deviation across ensemble predictions. Higher values indicate greater model disagreement and lower confidence
• Typical probability_std range: 0.05-0.25
• Predictions near the 0.5 threshold with high probability_std should be treated with caution
• The ensemble approach mitigates individual model overfitting but cannot account for systematic biases in the training data

Strengths & Limitations

Pros

• Sequence-only input: Requires only VH/VL Fv sequences, no 3D structure needed
• Fast inference: CPU-only, completes in seconds per antibody
• Uncertainty quantification: 102-model ensemble provides meaningful confidence estimates via probability_std
• Spatial features available: Optional DeepSP feature output enables downstream analysis of which regions drive viscosity
• Reproducible: Deterministic seeds ensure identical outputs for identical inputs

Cons

• Binary output only: Classifies as low/high rather than predicting continuous viscosity values
• Single buffer condition: Trained on 20 mM histidine-HCl pH 6.0 at 150 mg/mL only
• Small training set: Only 229 mAbs, limiting generalization to diverse antibody families
• No Fc contribution: Uses only Fv region; cannot capture Fc-mediated viscosity effects
• Isotype-agnostic: Does not distinguish IgG1, IgG2, IgG4 subclass effects

Known Failure Modes

• High-homology sets: Performance degrades significantly on antibody sets with high sequence identity (Apgar_mAb_38: 68.4% accuracy vs 87.5% on Lai_mAb_16)
• Non-antibody input: Sequences that cannot be aligned by ANARCI to IMGT numbering will cause a runtime error
• Unusual CDR3 lengths: Very long or very short CDR H3 loops that fall outside the IMGT position inclusion list may lose information
• Buffer-dependent viscosity: Antibodies whose viscosity is dominated by buffer/excipient interactions rather than Fv properties will be poorly predicted
• Boundary cases: Antibodies with true viscosity near 20 cP may be classified either way with high uncertainty

Implementation Details

Inference Pipeline

Request (VH + VL sequences)
  |-- 1. Validate input (Pydantic schema: 50-200 residues, unambiguous AA)
  |-- 2. ANARCI alignment (IMGT numbering, heavy + light)
  |-- 3. Fixed-length extraction (145 H + 127 L = 272 positions)
  |-- 4. One-hot encoding (272 x 21 matrix)
  |-- 5. [CPU] DeepSP CNN inference (3 models -> 30 spatial features)
  |-- 6. StandardScaler normalization (pre-trained on 229 samples)
  |-- 7. [CPU] Ensemble ANN inference (102 models -> 102 probabilities)
  |-- 8. Aggregate: mean probability, std, threshold at 0.5
  +-- 9. Format response (viscosity_class, probability_mean, probability_std)

Memory & Compute Profile

Component	Memory	Time (per antibody)
3 DeepSP CNN models	~50 MB	~100 ms
102 ANN ensemble models	~200 MB	~2 s
ANARCI alignment	~100 MB	~1 s
TensorFlow runtime overhead	~1.5 GB	--
Total	~2 GB	~3 s

No GPU required. CPU-only inference on a single core.

Determinism & Reproducibility

Setting	Value
NumPy seed	42
TensorFlow seed	42
GPU/CUDA	Not used (CPU only)
TF_ENABLE_ONEDNN_OPTS	Disabled

The model is fully deterministic: identical inputs produce identical outputs across runs.

Caching Behavior

• Cache key composition: Determined by input sequences and params (include_deepsp_features flag)
• Response caching is available as an optional, off-by-default gateway feature (BIOLM_CACHE_ENABLED) -- see the gateway docs; it is not handled by the model container.
• Memory snapshots: Enabled -- all 105 models (3 CNN + 102 ANN) are loaded during snapshot creation for fast cold starts

Training Procedures

Training Configuration

DeepSP CNN Training (upstream, not reproducible within BioLM):

Hyperparameter	Value
Optimizer	Adam
Loss	MAE
Training data	Structural properties from antibody 3D models

Ensemble ANN Training (upstream, not reproducible within BioLM):

Hyperparameter	Value
Optimizer	Adam
Learning rate	0.0001
Hidden layers	128 -> 64 -> 32 -> 16
Activation	tanh
Output activation	sigmoid
Validation	Leave-one-group-out (102 folds)
Feature scaling	StandardScaler (trained on 229 samples, 30 features)

Cross-Validation Results

From the paper (LOGO CV on 229 mAbs):

Metric	Value
Accuracy	85.2%
AUC	0.901
Sensitivity	0.83
Specificity	0.87

Versions & Changelog

Version	Date	Changes
v1	2025-03	Initial implementation: 102 ANN ensemble + 3 DeepSP CNNs, ANARCI alignment, embedded scaler

---

Biology

Molecule Coverage

Primary Molecule Type(s)

DeepViscosity is designed exclusively for monoclonal antibodies (mAbs), specifically their variable fragment (Fv) regions. It takes paired heavy chain variable (VH) and light chain variable (VL) sequences as input.

Scope: - Therapeutic monoclonal antibodies (IgG format) - VH and VL Fv domains only (excludes constant regions, Fc, hinge) - Sequences must be alignable to IMGT numbering via ANARCI - All organism sources are theoretically supported, but the training data is from humanized/human therapeutic mAbs

Performance considerations: - Best performance on antibodies similar to the AstraZeneca training set (humanized/human IgGs) - Performance degrades on antibody sets with very high internal sequence homology - Non-antibody proteins (enzymes, receptors) cannot be processed -- ANARCI alignment will fail

Cross-Applicability

Molecule Type	Applicability	Evidence	Caveats
Therapeutic mAbs (IgG1/2/4)	High	Primary training domain	Fc effects not captured
Biosimilar antibodies	High	Same Fv architecture	Trained on innovator mAbs; biosimilar-specific behavior untested
Bispecific antibodies	Low	Not tested	Single Fv pair input only; cannot model bispecific geometry
Nanobodies (VHH)	None	Not applicable	Requires paired VH/VL; single-domain antibodies not supported
Fab fragments	Moderate	Fv region is shared	Viscosity of Fab vs full IgG may differ due to missing Fc
Non-antibody proteins	None	Not applicable	ANARCI alignment requires antibody variable domains

Biological Problems Addressed

Problem 1: High-Concentration Antibody Viscosity Screening

Why this matters: Therapeutic antibodies are typically administered as subcutaneous injections at high concentrations (100-200 mg/mL) to deliver sufficient drug in a small injection volume (1-2 mL). At these concentrations, many mAbs exhibit dramatically increased viscosity due to reversible self-association of the antibody molecules. High viscosity (>20 cP) causes:

• Difficulty in syringe-based injection (patient compliance, injection pain)
• Manufacturing challenges (filtration, fill-finish, pump limitations)
• Formulation instability and aggregation risk
• Increased development cost and timeline

Traditional approaches: Viscosity is measured experimentally using cone-and-plate rheometry or microfluidic viscometers, which require purified protein at scale (milligrams to grams). This is expensive and slow, particularly in early-stage discovery when hundreds of candidates may need screening.

How DeepViscosity helps: From just the VH and VL amino acid sequences, DeepViscosity predicts whether an antibody is likely to have low (<=20 cP) or high (>20 cP) viscosity at 150 mg/mL. This enables:

• Early-stage triage: Screen antibody candidates computationally before any protein expression
• Lead optimization guidance: Identify sequence features driving high viscosity via DeepSP spatial properties
• Reduced experimental burden: Focus expensive viscosity measurements on borderline or high-priority candidates

Accuracy: 87.5% on the independent Lai_mAb_16 test set, with ensemble uncertainty estimates (probability_std) indicating prediction confidence.

Problem 2: Antibody Engineering for Reduced Viscosity

Why this matters: When a promising therapeutic antibody exhibits high viscosity, engineers must modify the sequence to reduce viscosity while preserving binding affinity and other functional properties. This is typically done through iterative mutagenesis -- an expensive trial-and-error process.

How DeepViscosity helps: The optional DeepSP feature output (30 spatial properties across 10 antibody domains) reveals which regions contribute most to predicted viscosity:

• SAP_pos (Spatial Aggregation Propensity, positive): Hydrophobic patches that promote self-association
• SCM_neg (Spatial Charge Map, negative): Negatively charged surface patches
• SCM_pos (Spatial Charge Map, positive): Positively charged surface patches

By examining which CDR or framework region has the highest SAP_pos or most extreme charge features, engineers can prioritize mutation sites to reduce viscosity. For example, high SAP_pos in CDR H3 suggests that hydrophobic mutations in that loop may improve viscosity without disrupting the binding paratope in other CDRs.

Limitations: DeepViscosity predicts the effect of the current sequence but cannot directly suggest specific mutations. It should be used as a scoring function within a broader engineering workflow, not as a standalone design tool.

Applied Use Cases

Use Case 1: Pre-Clinical mAb Candidate Screening

Source: Kalejaye et al. "Accelerating high-concentration monoclonal antibody development with large-scale viscosity data and ensemble deep learning." mAbs (2025). DOI

The primary paper demonstrates using DeepViscosity to screen mAb candidates before experimental viscosity measurement. On the Lai_mAb_16 independent test set, the model correctly classified 14 of 16 antibodies, identifying high-viscosity candidates that would otherwise require weeks of experimental characterization. The ensemble approach (102 models) provides uncertainty estimates that flag borderline predictions for experimental follow-up.

Use Case 2: Formulation Development Prioritization (Anticipated)

While not explicitly demonstrated in the published literature, DeepViscosity predictions could guide formulation development prioritization. Antibodies predicted as low-viscosity with high confidence may proceed directly to standard formulation screens, while those predicted as high-viscosity may require specialized formulation strategies (e.g., viscosity-reducing excipients like arginine or NaCl) from the outset.

Related Models

Predecessor Models

• DeepSP (Rawat et al. 2019): The CNN feature extraction component of DeepViscosity. DeepSP predicts spatial aggregation propensity (SAP) and spatial charge map (SCM) features from sequence alone, replacing the need for 3D structure-based calculations. DeepViscosity extends DeepSP by adding an ensemble classification layer on top of these features.

Complementary Models

No models currently in this catalog are direct complements for a viscosity prediction workflow. Potential future integrations:

• Structure prediction (ABodyBuilder3, ESMFold): Generate 3D structures for mAbs predicted as high-viscosity, enabling physics-based analysis of self-association interfaces
• Inverse folding (ProteinMPNN): Design sequence variants that maintain fold stability while reducing predicted viscosity

Alternative Models

Alternative	Advantage over DeepViscosity	Disadvantage
CamSol (not in this catalog)	General solubility predictor; structure-aware	Not viscosity-specific; requires 3D coordinates
TAP (not in this catalog)	Multi-property developability score	Requires 3D homology model; coarser viscosity proxy
Sharma et al. ML models	Trained on larger proprietary datasets	Not publicly available; no ensemble uncertainty

DeepViscosity is currently the only open-source, sequence-only viscosity classifier with ensemble uncertainty estimates.

Biological Background

Antibody Structure

Antibodies (immunoglobulins) are Y-shaped proteins produced by the immune system to recognize foreign molecules (antigens). A typical IgG antibody consists of two heavy chains and two light chains. The antigen-binding region is formed by the variable fragment (Fv), which comprises:

• VH (Variable Heavy): The variable domain of the heavy chain (~110-130 residues)
• VL (Variable Light): The variable domain of the light chain (~110-130 residues)

Within each variable domain, three hypervariable loops called complementarity-determining regions (CDRs) form the antigen-binding site (paratope). CDR H3 is typically the most variable and often dominant in antigen recognition. The remaining residues form the framework regions that maintain the structural scaffold.

IMGT Numbering

The International ImMunoGeneTics (IMGT) numbering system provides a standardized way to number antibody residues across species and germlines. This is critical for DeepViscosity because the model requires a fixed-length representation: by aligning all antibodies to IMGT positions, diverse sequences become directly comparable. ANARCI is the tool that performs this alignment using hidden Markov models (HMMs).

Viscosity at High Concentration

At therapeutic concentrations (100-200 mg/mL), antibody solutions can become highly viscous due to reversible self-association. This is driven by:

• Electrostatic interactions: Charge patches on the antibody surface attract neighboring molecules
• Hydrophobic interactions: Exposed hydrophobic regions (measured by SAP) promote clustering
• Fab-Fab and Fab-Fc interactions: Both intra- and inter-molecular contacts contribute

The 20 cP threshold used by DeepViscosity is a commonly cited upper limit for subcutaneous injectability. Solutions above this viscosity are difficult to push through narrow-gauge needles (typically 25-27 gauge) used for patient self-administration.

Why This Matters for Drug Development

Approximately 80% of approved monoclonal antibody therapies use subcutaneous injection, which requires high concentration formulations. Viscosity is one of the most common developability liabilities, affecting:

• Patient compliance: High-viscosity injections are painful and slow
• Manufacturing: Filtration, filling, and pumping become difficult above 20-50 cP
• Stability: Self-association at high concentration can promote irreversible aggregation
• Cost: Late-stage reformulation or sequence redesign to address viscosity can cost millions and delay timelines by 6-12 months

Early computational screening with tools like DeepViscosity can identify viscosity risks before any protein is expressed, potentially saving significant development time and cost.

---

Sources & license

License: MIT (text)

Papers

• Accelerating high-concentration monoclonal antibody development with large-scale viscosity data and ensemble deep learning — mAbs, 2025 · DOI

Source repositories

• github: https://github.com/Lailabcode/DeepViscosity

Cite

@article{kalejaye2025deepviscosity,
  title={Accelerating high-concentration monoclonal antibody development with large-scale viscosity data and ensemble deep learning},
  author={Kalejaye, Lateefat A. and Chu, Jia-Min and Wu, I-En and others},
  journal={mAbs},
  volume={17},
  number={1},
  pages={2483944},
  year={2025},
  publisher={Taylor \& Francis},
  doi={10.1080/19420862.2025.2483944}
}