Deep Learning Model for Fast Determination of Equilibrium Dissociation Constants Using Biolayer Interferometry Sensorgrams

This paper explores the fusion model of deep learning with Bio-Layer Interferometry (BLI), a key detection method for biomolecular interactions. We constructed a convolutional neural network model capable of quickly predicting the value of the equilibrium dissociation constant (<i>K</i>_d). The model introduces SE and CBA modules, and was also fine-tuned. These optimizations further enhance the model's prediction accuracy and generalization ability. Based on a combination of dry-semidry-wet strategy, we innovatively employed the YOLOv5 model to automatically identify and extract a total of 3812 BLI curves from relevant literature in the PubMed database, 525 curves (wet data) obtained by our lab and 1303 curves (dry data) generated with deep generative adversarial network, expanding the data set to 5640 sensorgrams, which were separated into training set and validation set. The wet data was employed to verify the accuracy of the model in predicting the K_d values, with total accuracy of 60%. The as-proposed deep learning model is able to accurately predict the K_d value soly based on a single concentration BLI sensorgram, suggesting that the method is particularly suitable for the rapid analysis when it is difficult or impossible to perform a multiconcentration BLI assay. With the accumulation of more high-quality BLI sensorgrams in the future, the prediction accuracy and robustness of the model will be further improved.