TL;DR: docking is much slower than any ML approach, but the ML approach can be constrained by pharmacophores dictated by the active site.
Side note: Scale
The scale for ligand space exploration is generally is generally several orders of magnitude higher than "hundreds": Zinc DB lists 750 million enumerated compounds, GDB13 has enumerate all possible compounds up to 13 atoms (970 million), Enamine Real is 1.2 billion etc. In an actual experiment, one would use a more restrictive set. As an example of a real project, in Covid19 moonshot project there are 2,800 user suggested follow-ups.
Side note: More than just binding
There are a few requirements for a drug to be effective, and binding is only one of them:
- Must follow Lipinksi's rule of five, i.e. it must be membrane permeable and small.
- Must bind.
- Must be synthesisable. Whereas the experience of someone learnèd is by far the best away to tell, there are several machine leaning approaches that can filter a huge dataset. IBM actually have a fun online tool.
- Must not have side effects
So pre-filtering the list of compounds, regardless of docked poses, will happened, be it manually, by cut-offs or by advanced machine learning.
CPU time: docking vs. machine learning
These various steps are generally followed in that order.
Calculating properties (MW, logP etc.) for a given SMILES string to filter based on Lipinski's rule takes a fraction of a second, so that is generally always a first step.
Whether the compound can synthesised or bought is a general left last and done manually. Pharmacology is complicated so is ignored.
Docking a ligand takes time and different approaches have different time scales. A decent algorithm that uses implicit solvent model takes about a minute per core. A short MD simulation takes an 1 hour per core or more. There are docking programs for screening that can be faster but these dock only one conformer of the small molecule against a rigid protein and the results are not of much value. Also setting up docking properly is not easy (cf. steps required for Autodock 4, each with their own caveats). So to recap what needs to be considered that results in the slowness:
- Hundreds of conformers of the ligand
- Some flexibility of the protein, side-chains repacking only or even backbone changes.
- thousands of poses explored.
- Optionally. Implicit water is not as good as a TIP3 water model
Machine learning have a much smaller time requirement and is way more attractive for publications. Generally these work by using ligand only information, such as properties, splitting it into pharmacophores, colour etc. This requires a decent dataset, such as empirical results, say from a fragment screen, what compounds bound and which didn't. Often there is a lot of hype with machine learning and the result is something close to something already seen, as well put in this blog post by Pat Walters. This is a side effect of the starting data. The recent Stokes et al. 2020 Cell paper uses AI perfectly because of their staggering automated empirical data collection pipeline.
ML after docking
However, ML can actually be used after docking to improve the quality of the score.
The Boyles, Deane, Morris 2019 Bioinformatics paper investigated ligand-only machine learning to see why do ML methods that rely solely on ligand somehow often work. It is a good paper from a sceptical computational biochemistry viewpoint.
The scorefunction used in docking by the sampler can be classified as forcefield-based, empirical or hybrid. In some cases empirical cases, many terms are ligand based properties (/features) that would normally go into a ML search (say one trained on failed hits and successful hits).
For example Autodock Vina (not 4) has regression derived factors in the final weights of the score
Structural features in machine learning: HotSpots
Information of the protein active site can be used in ranking compounds in a non-docking machine learning approach. Namely, finding out what kind of chemical groups bind preferentially to different parts of the protein's active site, i.e. making a "fragment hotspot map". Specifically, a hotspot map allows the determination of a pharmacophore, which is an abstract generalisation of a group of molecules with both both electronic and steric features —a coarse-grain model, if you will. This approach is also used to detect where the active site actually is in the extreme cases where it is not know (primarily a problem for pipelines, not humans). There are several implementations as described here, the first being FTMap (site, paper), which docks 16 different types of compounds. Waters can play a role and severely complicate matters, hence approaches can be taken to counter that. However, unlike docking protein rigidity is not a big nuisance: this is because a hotspot map is most often used for filtering/weighing by the "color" not the shape of the molecules. But shape based searching is possible.
Shape based scoring
In terms of shape-based searching, this is actually a question of having pre-computed conformers, which can be problematic (but can and is often done on the fly), and shape based scoring —for which there are different metrics, such as the classical RMSD to more complexly weighted scores such as SuCOS.