Use DeepSP Online
Commercially Available DeepSP No-Code Web Server
DeepSP: Predicting Antibody Spatial Aggregation Propensity and Charge from Sequence
DeepSP (Deep Spatial Properties) is a convolutional neural network surrogate model that predicts spatial aggregation propensity (SAP) and spatial charge map (SCM) scores across antibody variable regions directly from sequence, without running molecular dynamics (MD) simulations. It was developed by Lateefat Kalejaye, I-En Wu, Taylor Terry and Pin-Kuang Lai at the Department of Chemical Engineering and Materials Science, Stevens Institute of Technology, and described in Computational and Structural Biotechnology Journal in 2024. The authors state that it is freely available as an online tool and that code and parameters are freely available on GitHub.
Therapeutic antibody development faces high viscosity and aggregation tendencies, especially for the high-concentration, low-volume formulations needed for subcutaneous administration. SCM and SAP are computational techniques that help predict viscosity and aggregation, respectively, but both rely on structural data from MD simulations, which are computationally demanding.
DeepSP extends the earlier DeepSCM model, which predicted SCM for the whole Fv region only, by predicting SAP and SCM scores in individual domains of the variable region, so they can serve as sequence-only features for stability models.
Attribute | Detail |
|---|---|
Task | Predict 30 SAP and SCM spatial properties from antibody Fv sequence |
Model | Convolutional neural network with three convolutional layers |
Training data | 20,530 antibody Fv sequences from the Observed Antibody Space (OAS), labeled with MD-derived scores |
Regions scored | CDRH1, CDRH2, CDRH3, CDRL1, CDRL2, CDRL3, all CDRs, Hv, Lv and Fv |
Publication | Comput. Struct. Biotechnol. J. 23, 2220-2229 (2024) |
How DeepSP Predicts Antibody Biophysical Properties Without Simulation
The authors built the training set by simulating antibody Fv homology models (generated with ABodyBuilder-ML) and averaging SAP and SCM scores over the MD trajectories. After filtering, 20,530 sequences remained. The model then learns to map sequence to those averages.
Numbering and alignment: sequences are annotated with the Chothia numbering scheme so heavy and light chains have a fixed input size, with gaps filled by dashes.
One-hot encoding: each residue is encoded as a one-hot vector for the CNN.
Three models, ten regions each: separate models predict SAP_pos, SCM_neg and SCM_pos across all ten regions, giving 30 properties. The authors chose this over training 30 separate models because performance was not significantly different.
Descriptors as features: the predicted scores can feed other machine learning models for antibody stability.
DeepSP Accuracy and Antibody Aggregation Prediction Results
Evaluation | Result |
|---|---|
Agreement with MD-derived scores (test set, N = 2053) | Correlation 0.76 to 0.96 across 30 properties, average 0.87 |
Aggregation rate of 21 antibodies, DeepSP features (SVR) | R = 0.97, MSE = 0.03; leave-one-out cross-validation R = 0.75, MSE = 0.18 |
Same task, earlier MD-based features | R = 0.94, MSE = 0.08; leave-one-out cross-validation R = 0.77, MSE = 0.22 |
The authors conclude that DeepSP features give performance similar to MD-derived features while substantially reducing computational time. In both studies, the most important selected feature was a positive charge descriptor (SCM_pos_Fv in the earlier study, SCM_pos_CDRH3 here), and CDRH3 features were often selected, which the authors attribute possibly to its high sequence diversity.
What is Tamarind Bio?
Tamarind Bio is a no-code bioinformatics platform built to give life scientists and researchers access to powerful computational tools. Many cutting-edge machine learning models are hard to deploy and use. Tamarind provides an intuitive, web-based environment that removes the complexity of high-performance computing, software dependencies and command-line interfaces.
The platform is designed for biologists, chemists and other researchers who may not have a background in programming or cloud infrastructure but want to run models on their own data. Key features include:
A user-friendly graphical interface for setting up and launching experiments
A robust API for integration into existing research pipelines
An automated system for managing and scaling computational resources
Tamarind treats information and data security as a top priority, as detailed in its Trust Center and Terms of Service.
Accelerating Discovery with DeepSP on Tamarind Bio
Developability screening: rank antibody candidates by predicted SAP and SCM descriptors before expression.
Aggregation-rate modeling: use the 30 descriptors as features in models of high-concentration aggregation, as the authors demonstrate with 21 antibodies.
Viscosity risk assessment: SCM is described by the authors as a screen for high-concentration viscosity, and SAP for aggregation-prone regions.
Large library triage: score many sequences without MD; the authors note the descriptors can be generated for many sequences at once.
How to Use DeepSP on Tamarind Bio
Log in: go to tamarind.bio and open the DeepSP tool.
Provide sequences: enter the name, heavy chain and light chain Fv sequences of each antibody (the original web form takes these three inputs).
Check inputs: the model was trained on paired Fv sequences with limited CDRH3 length, so use variable-region sequences only.
Run: submit the job and wait for it to finish.
Download outputs: retrieve the predicted SAP_pos, SCM_neg and SCM_pos scores for each region.
Interpret: compare scores across candidates, remembering they approximate MD-derived values.
Use downstream: feed descriptors into aggregation or viscosity models, or use them to prioritize candidates for experiments.
Things to Keep in Mind
DeepSP predicts MD-derived descriptors; correlations ranged from 0.76 to 0.96, so some properties are predicted better than others.
The aggregation-rate validation used only 21 antibodies, and the authors note the selected features might differ with a larger dataset.
Training sequences excluded unusually long CDRH3 regions (over 30 residues) and sequences with insertions in other CDR or framework regions, so inputs outside this range may be less reliable.
Predictions are computational estimates and need experimental confirmation.
Source: Kalejaye L, Wu I-E, Terry T, Lai P-K. DeepSP: Deep learning-based spatial properties to predict monoclonal antibody stability. Computational and Structural Biotechnology Journal 23:2220-2229 (2024). DOI: 10.1016/j.csbj.2024.05.029. Code: https://github.com/Lailabcode/DeepSP.