ProtT5 and random forests-based viscosity prediction method for therapeutic mAbs.

Area of Science:

• Biotechnology
• Protein Engineering
• Computational Biology

Background:

• Therapeutic antibody viscosity is critical for subcutaneous delivery, impacting formulation concentration and volume.
• Accurate early-stage viscosity prediction is needed to streamline the development of high-concentration antibody formulations.

Purpose of the Study:

• To develop and validate a sequence-based computational method for predicting monoclonal antibody viscosity.
• To facilitate high-throughput screening of antibody candidates for optimal viscosity characteristics.

Main Methods:

• Feature extraction from antibody V-region sequences using the ProtT5 pretrained model.
• Dimensionality reduction of features via Kernel Principal Component Analysis (Kernel-PCA).
• Development of a Random Forests (RF) regression model trained on 40 commercial therapeutic antibodies.

Main Results:

• The RF model achieved high accuracy in predicting viscosity values (Pearson correlation coefficient = 0.928).
• The classification model demonstrated excellent performance for high (>30 cP) and low (<30 cP) viscosity (Accuracy = 0.975, AUC = 1.000).
• The proposed method outperformed five existing state-of-the-art viscosity prediction techniques.

Conclusions:

• The ProtT5 and RF-based method offers a reliable and efficient approach for predicting antibody viscosity from sequence data.
• This predictive tool can significantly accelerate the development of advanced therapeutic antibody formulations.