DeepSCM

DeepSCM predicts high-concentration antibody viscosity using a convolutional neural network surrogate model to enable rapid screening of therapeutic antibody candidates for subcutaneous administration.

Key Features:

• Convolutional neural network surrogate model: Uses a CNN-based surrogate to predict viscosity-related scores and reproduce spatial charge map (SCM) signals from sequence input.
• Sequence-based prediction: Processes antibody sequence information directly without requiring structural data.
• Training dataset: Trained on SCM scores computed by high-throughput computing for 6,596 nonredundant antibody variable regions.
• Performance metrics: Demonstrates a linear correlation coefficient of 0.9 relative to SCM scores on a test set of 1,320 antibody sequences.
• Experimental validation: Applied to screen 38 therapeutic antibodies, with one misclassification reported.
• Computational efficiency: Serves as a surrogate to reduce computational demands compared to molecular dynamics–based SCM calculations.

Scientific Applications:

• Viscosity screening for subcutaneous formulations: Predicts high-concentration antibody viscosity to support selection of candidates suitable for subcutaneous administration.
• Therapeutic antibody candidate triage: Enables rapid prioritization of antibody sequences for further biophysical and formulation studies without requiring structural models.

Methodology:

Trains a convolutional neural network surrogate on SCM scores computed by high-throughput computing for 6,596 nonredundant antibody variable regions, evaluates performance on a 1,320-sequence test set by comparison to SCM scores, and applies the model to screen 38 therapeutic antibodies.