Computational chemistry for drug discovery: A flight plan

Over the years as I have applied both computational and other technology platforms to problems in drug discovery and design, a few guiding principles have crystallized in my mind as signposts and caveats. Calling the list a "checklist" is a bit reductive. Instead these principles appear to me as part of a "flight plan", rules of thumb and guidelines that don't just enable you to check a set of boxes but help keep your eyes on the ultimate destination (in our case a lead, IND or marketed drug). Here's what I have in mind:

1. Means and Ends: Technology in drug discovery is the means to the end. The end is developing drugs that improve, extend and save people’s lives. One can be passionate about technology and still need to be able to subsume it to the ultimate goal of helping patients. And like a person climbing a ladder who is willing to throw away its rungs once he climbs them, practitioners of technology should be able to discard it when it doesn’t work for them, or when other technology better serves their purpose.
2. Swiss Army Knife: Computational chemistry is a set of tools, a Swiss Army Knife rather than a gleaming dagger, a utility belt that can adapt to the problem at hand rather than the proverbial hammer looking for a nail. Every problem calls for a different tool; you can have preferences for or training in specific tools, but you should not pick favorites and use whatever best works for a problem.
3. Synergy: Computational chemistry tools work best when they piggyback on each other. The imperfect output from an AI tool can be improved by funneling it into a structure-based pipeline. The output from a structure-based pipeline can complement that from a ligand-based workflow. This synergy applies even more when combining computational with experimental tools. High-throughput and virtual screens complement each other. Data from NMR informs structure prediction. Computational chemistry tools form an open society, building on top of each other and helping tools from other open societies do what they do best.
4. Complexity: Drug development benefits from a diverse array of tools precisely because it is so complex and uncertain. Human biology is the most unpredictable and complicated system we have tried to understand. The laws of complexity say that if you want to tame complex systems, you need to be flexible and leave enough wiggle room for addition, elimination and accretion of technologies. You need to be humble in recognizing that no one piece of technology is privileged.
5. Models: Remember that everything in drug discovery is a model. Protein complexes on a computer screen and patient-derived xenografts are obvious models, but so are x-ray crystal structures, SPR and dose-response curves, cell images showing apoptosis, PK/PD data, and clinical trials. Every piece of data comes to us filtered through the lens of a particular tool, with its attendant statistics and curve-fitting. There is no such thing as true “raw data”, only ever-increasing approximations to it. Every piece of data therefore must be taken with a grain of salt, seen through a glass darkly.
6. Simplicity: Never use a complicated method when a simple method works, especially when the latter is interpretable and directly informs molecular design. Never declare a method to be the “best” until it runs the gauntlet of validation and comparison with simpler techniques. But do not worship simplicity as an end in itself. Nature does not always shave with Occam’s Razor, and sometimes what’s too complex for the human mind to imagine (for instance multiple conformations fitting NMR data piecemeal)? is a better approximation than the alternative (for instance a single structure fit to NMR data).
7. Foxes and Hedgehogs: Because of its need to fit specific tools to specific problems on demand, computational chemistry is the ultimate foxes’ game. Its practitioners need to be nimble and flexible. But without the hedgehogs who drill deep developing molecular dynamics or cryo-EM or fragment-based design, the foxes would not have a playground to practice their art. Both foxes and hedgehogs will remain necessary for the growth of computational drug design.
8. Experiment and Theory: The most important thing about theory is experiment. Computational chemists are most effective when they know what the experiments are, what makes them tick, what the error bars are. Without incomplete knowledge of the experimental details, they may model the wrong thing and go down a rabbit hole. Details of protein crystallization constructs, compound stereochemistry and ionization states, cofactors and binding partners are not just optional knowledge; they are the bedrock on which accurate models are built.
9. Biases: Every drug discovery scientist infers and calculates based on a slice of their favored reality. Every computational chemist operates with mental models that are informed by anecdotal success. There is always a chemotype or a specific technique that seemed to have worked spectacularly well in a project. But anecdotal success is not data. All scientists suffer from cognitive biases, most notably confirmation bias. Inconvenient data should not be ignored, negative results should occupy a special place in training data for models.
10. Collaboration: If computational techniques should make up an open society, the larger community in which computational chemists live and breathe should make up a democracy, one which reflects the democratic nature of science in which anyone, irrespective of title or position, should be free to question data or suggest ideas. Knowing that every tool is an imperfect tool applied to complex, uncertain, emergent biological systems, ideas from every corner should be equally welcome and equally stringently evaluated. If it takes a village to raise a child, it takes a world to develop a lifesaving drug.