Deep Learning Based Optimization of Immunoblotting Reaction Conditions

Immunoblotting is a biochemical technique used to quantify proteins in a sample. It is present in 9% of all protein-related publications and is one of the most used biochemical techniques. However, its methods have remained stagnant since its emergence in the 1980s, with 41% of scientists reporting their blot as unsuccessful more than 25% of the time. In this research, we propose a method for assisting scientists with an optimal primary antibody dilution for their blot. Our research is five-fold: (1) perform a background survey to determine important independent variables, (2) use Large Language Models to scrape open-source articles to obtain variables, (3) extract key structural details, (4) train a multi-modal deep learning model to take protein structure image input and compiled tabular inputs to predict an antibody dilution, and (5) test model using cross validation, heatmaps, and prediction on out-of-sample data. The model achieved a validation accuracy of 94% and an out-of-sample testing accuracy of 89.09%, indicating its ability to reduce iterative processes a substantial amount of the time. Model reliability was assessed by backing regions of importance with scientific findings through saliency maps. Receiver Operating Characteristic (ROC) curves with an average AUC of 0.9317 indicate that the incorporation of multi-modal data increases overall accuracy and reduces both the false positive and false negative rate. The resulting model could reduce the time and resources associated with immunoblotting, allowing scientists to channel their resources into discovering new biological interactions that could contribute to cures and diagnostic tests for debilitating diseases.