Enterprise AI Analysis

The model was trained on a dataset of single-chain fragment variable (scFv) antibody sequences and associated binding scores (KD values) against a SARS-CoV-2 HR2 region peptide. This dataset included variants generated by introducing 1-3 amino acid changes into three seed antibodies. KD values were preprocessed by taking the arithmetic mean of the two closest replicates. 71,834 unique antibodies were used for training. Mean Squared Error (MSE) was used as the loss function, and Adam optimizer for parameter optimization. 85% of data for training, 15% for validation. Trained on NVIDIA A100 GPUs with a batch size of 128 for 100 epochs. Fine-tuning of pre-trained ESM-2 was performed, and a model with randomly initiated weights was also trained for comparison. Best performing model achieved highest Pearson correlation on validation set.

Ab-Affinity demonstrated superior performance in predicting binding affinity compared to other LLM-based methods (DG-Affinity, ESM-2, AbLang). T-SNE visualizations showed that Ab-Affinity embeddings produced a smooth gradient of binding affinity, unlike ESM-2. The model achieved the highest Pearson (0.652) and Spearman (0.712) correlation coefficients on the test set. Ab-Affinity embeddings also proved highly effective for downstream classification tasks, such as determining binding affinity classes (High, Medium, Low) and identifying improved binding, with significantly higher AUC values than ESM-2. Attention maps revealed focus on CDRs and adjacent regions for binding prediction. The model also implicitly captured thermostability properties, separating antibodies into clusters based on their thermostability values in t-SNE.