WDL-RF (weighted deep learning and random forest) is a novel pipeline for bioactivity prediction of GPCR-associated ligand molecules. In commercial drug design, virtual screening is acceptable only when the prediction accuracy is high. One of the outstanding issues with the bioactivity modeling is that the input to the model, a ligand, can be of arbitrary size, but most of the current predictors can only handle inputs of a fixed size. WDL-RF builds on a novel two-stage algorithm, with molecular fingerprint generated through a weighted deep learning method, followed by random forest based bioactivity assignments. The pipelins allows high-accuracy end-to-end learning of prediction pipelines whose inputs are of arbitrary size. The large-scale benchmark tests showed that the WDL-RF model has an average root-mean square error 1.42 and correlation coefficient 0.78, compared to the experimental measurements.
yangzhanglab
umich.edu
| (734) 647-1549 | 100 Washtenaw Avenue, Ann Arbor, MI 48109-2218