Boosting Protein−Ligand Binding Pose Prediction and Virtual Screening Based on Residue−Atom Distance Likelihood Potential and Graph Transformer
We have an interest in MedChemica in following developments in virtual screening. The paradigm of protein sequence identification, computer folding to yield a 3D structure, pocket identification, then virtual screening yielding small testing sets of compounds could have an enormous impact on the industry. If the screen sets were sub 100 thousand compounds and these yielded highly developable leads 30 to 50% of time, then we have a game changer. Remarkably protein folding is improving all the time, so with improved docking algorithms the pieces of this puzzle could be coming together. I say ‘could’ be coming together.
This particularly paper is another approach using Deep Learning with a different compound and protein descriptor. It is highly interesting and a key reason for highlighting the paper is the code is available from GitHub. When performing docking studies and exploring a new method there is an intrinsic problem of validating the results. The approach normally taken is to compare docking results with other docking programs, which is done extensively in this work. It all looks great. Non of this tells us if this is going to work on the next virtual screen we perform. This is problem, and indeed a problem to define what the ideal validation experiment would be.
…anyway enjoy the paper, and we might try the code in the ASAPproject for new anti-virals for the next pandemic – we will let you know.