Training Machine Learning Prediction of Antibody Thermostability Utilizing PROPERMAB Features and the GDPa4 Dataset

Antibody drug development is a years-long, costly process. Antibody thermostability is one of the key developability features of an antibody drug that influences its chances of success in development. But, experimental testing such as nanoDSF for large libraries of candidates requires substantial resources. Instead, we evaluated an open-source feature-based machine learning workflow to train a model for antibody stability (Tm1) predictions from sequence and structure-based features. We used PROPERMAB to extract features from a selection of 231 different monoclonal and bispecific antibodies from the GDPa4 nanoDSF dataset. We benchmarked regression and classification models that predict the median first thermal transition temperature (Tm1) from PTS-IF in degrees Celsius for antibodies in the GDPa4 data set from Gingko Datapoints. We trained the models on 184 antibodies and evaluated model performance on 47 held-out antibodies. The three different regression models were ElasticNet, RandomForest, and GradientBoosting. RandomForest achieved the highest reported correlations among the regressors (Pearson r = 0.750, R2 = 0.562). Performance was also evaluated relative to 66.29 degrees Celsius as the threshold for thermostability. The classifier models were LogisticRegression, RandomForest, SupportVector, and GradientBoosting, and classification was defined relative to the training set median of 66.29 degrees Celsius. Thus, above-threshold and below-threshold describe position in this data set, not developability. The best performing model, RandomForest, improved on recall but reduced its precision by recovering two above-threshold antibodies in return for two more false positives. False positives are preferred to false negatives and classification models tended to reduce false negative counts. Most test antibodies also share a parent with a training antibody, so the models predict well for new combinations of known parents but generalization to unknown parents remain unseen However, due to the limited data set used, independent dataset validation is needed before drawing concrete conclusions.