Table 2 ∣.

Hit picking criteria

Docking score	Docking scores are approximations of interaction energies and therefore cannot be used for absolute ranking of hits. Its main purpose is to identify compounds that are likely to bind from the large background of compounds that are unlikely to bind
Broken molecules	Errors in 3D building can result in incorrectly built compounds. Common errors include improper tautomerization or protonation states owing to the difficulty in predicting pKa. Such ‘broken’ molecules should be deprioritized during visual inspection
Internal strain	Often internal strain is not included in energy functions but is an important concern since a ligand may score highly if it adopts a strained conformation that increases its contacts with the receptor (e.g., nonplanar amide). Visual examination can catch these strained compounds, but various tools have been developed to systematically analyze hundreds or thousands of compounds making them useful in this setting127-129. We recently proposed a computational tool that estimates torsion strain energies based on experimentally determined torsional populations in the Cambridge Structural Dataset129,130
Interaction patterns	Specific interactions to key residues are main features of promising candidates. Hydrogen bonds between ligands and side chains can be computationally identified by simple distance cut-offs (<3.5 Å). For example, hydrogen bonding to kinase hinge residues131, or salt bridges to the conserved aspartic acid D3.32 in aminergic GPCRs132, offer reasonable anchor points for interaction filters. In a similar fashion, a pharmacophore filter can be used to search for any atoms in an unexplored subpocket within the larger binding pocket. Molecules that score well but form only one key interaction may be deprioritized. ‘Floating’ molecules that may bury an energetically frustrated water should also be eliminated
Unsatisfied hydrogen bond donors and acceptors	Molecules with unsatisfied hydrogen bond donors or acceptors, especially in hydrophobic pockets of the site, often pay a high desolvation cost; not all scoring functions are designed to filter these out explicitly. Therefore, we suggest keeping the number of unsatisfied acceptors below 3 and the number of unsatisfied donors to 1 or below, as burying a donor can incur a greater penalty than burying an acceptor133
Novelty filter	One of the advantages of large-scale docking with make-on-demand chemical libraries is the ability to screen for novel chemistry. Therefore, if the goal of the project is to identify novel scaffolds for hits, it is useful to filter out docking hits that resemble known actives. This can be accomplished using any number of computational methods that compare the 2D or 3D topologies of hit compounds. In our hands, we typically remove hits that have an ECFP4 Tanimoto cutoff of 0.35 or higher to any known active12,13
Scaffold clustering	From the initial docking screen, a set of diverse scaffolds should be selected for experimental testing to cover a larger subset of chemical space. Clustering metrics to identify diverse hits may include ECFP4134 fingerprints, Bemis-Murcko88 scaffolds, bcl::Cluster similarity135 or Atom Counts136