Rigor and Reproducibility Via Virtual Screening in SMSF

On the computational side of SMSF, we have made big strides in rigor and reproducibility with computational procedures. This reflects a broader shift in this direction by the computational chemistry community at-large.

Our main services involve virtual screening, where huge virtual compound libraries can be evaluated using computational models to recommend much smaller sets of compounds for purchasing/synthesis and testing. This enables investigators to find active compounds for their target of interest without having to run expensive, brute-force HTS.

In terms of rigor, we have developed virtual screening methods on campus in collaboration with UW Profs. Anthony Gitter, Michael Newton, and Stephen Wright. These methods are validated extensively through retrospective and prospective tests.

Our structure-based virtual screening, which is based on simulating the interaction between 3D molecular models of protein targets and compounds, uses up to 8 different docking programs and compiles these results to provide a more robust prediction. The vast computation resources at the CHTC have enabled this approach, which requires millions of independent docking calculations to validate this method in retrospective tests against >20 benchmark protein drug targets with known actives and inactives. We evaluate performance based on how the consensus score shifts the actives to the top of the list (see Ericksen et al., 2017).

Our ligand-based virtual screening procedures use machine learning methods that do not require a protein target structure and can even predict phenotypic activities. The models are trained with available experimental data comprising examples of active and inactive compounds. The models learn a relationship between molecule attributes and an experimental read-out like assay activity. We have demonstrated excellent performance in retrospective testing on 128 selected targets with public assay data (PubChem BioAssay) (Liu et al., 2019). Prospective tests on a campus target involving 8 million molecules has shown great promise and we have recently expanded this procedure to virtual screens exceeding 1 billion molecules. Related to this work we have also developed procedures that work in low data contexts where few training data are available. These methods were validated on >200 human protein kinases and enabled effective predictions in prospective tests on microbial kinases (Zhang et al., 2019).

In terms of reproducibility, we use programs or develop our own codes with unrestricted academic licenses to facilitate reproducibility of our work. Our procedures and codes are made publicly available on GitHub and detailed in our publications.



Ericksen SS, Wu H, Zhang H, Michael LA, Hoffmann FM, Wildman SA (2017) Machine Learning Consensus Scoring Improves Performance Across Targets in Structure-Based Virtual Screening. J Chem Inf Model. 57(7):1579-90.

Liu S, Alnammi M, Ericksen SS, Voter AF, Ananiev GE, Keck JL, Hoffmann FM, Wildman SA, and Gitter A. (2019) Practical Model Selection for Prospective Virtual Screening. J Chem Inf Model. 59(1):282-293. https://doi.org/10.1021/acs.jcim.8b00363.

Zhang H*, Ericksen SS*, Lee CP*, Ananiev GE, Wlodarchak N, Yu P, Mitchell JC, Gitter A, Wright SJ, Hoffman FM, Wildman SA, and Newton M. (2019) Predicting kinase inhibitors using bioactivity matrix derived informer sets. PLoS Comput Biol. Aug 5; 15(8):e1006813. doi: 10.1371/jounal.pcbi.1006813. PMID: 31381559. *co-first authors