Monday Lunchtime Discussion
Discussion - 2014 Workshop - Monday lunch
Topic: Overview
Discussion Leaders: Michael Shirts, Clara Christ
What is the current state of free energy calculations in drug design?
Gabe: Are things different now from where they were two years ago?
John van Drie: At first one: David Pearlman said "we are now where we thought we were 20 years ago"
Gabe Rocklin: Number of industrial calculations and comparisons to experimental data are much more than a few years ago
Robert Abel: Two years ago, Mark Murcko threw out a challenge: Why are we all reporting results for a single molecule? We're past that now.
Teng Lin: We are now starting to accumulate enough data to enable people to do something
Sunjin Ho(?) Pfizer: Trying to build up wrong model is very important. Making it easier to use and use consistently is important. Industry wants to try to identify methods that work Pretty much all industry attendees have run some free energy calculations
Bethany: Hasn't had much success in general yet.
Vijay: Understanding why things aren't working is important. Time to be nailing these things down.
John Chodera: Would like to see standard list of problems we look for in free energy calculations so that non-experts can figure out when something has gone wrong.
Christopher Bayly: Last time I tried to apply free energy calculations in industry was a few years ago, and it was a complete train wreck. But industry has alternative methods of scoring that worked pretty well. Previously, free energy methods weren't even in the ring, but now they are contenders. But how do they stack up against other existing methods? Need to compare to good null models. Methods need to be justified against not worst of competing models, but best of competing models. Are there some standard null models we should be comparing to?
Robert Abel: Null model we've had most success with is tracking which compounds have been synthesized directed by FEP vs molecules synthesized directed by all other technologies.
John van Drie: Any computations that can generate insight for chemists are valuable.
JW Feng: Another idea for a null model: Have team vote yes/no on potency (better than 10 nM) with prize being a bottle of wine.
?: Almost about the compounds that are not being made. Bill Jorgensen estimates cost of compound synthesis to be $3K/compound. Always screening out ~100 compounds for every compound we make. $3K x 100 is a significant savings. Considering how many compounds it saves us from making is important in assessing value.
?: Another null model is to look at the size of the change. MW or cLogP.
David Minh: Another null model: Randomize the scores you get from free energies.
Vijay Pande: We're starting to see in specific cases (minimal conformational change) DDG calcs are working reasonably well. Still caveats, but for small changes, things look much better than before.
?: I work with a protein where binding cavity has single aromatic amino acid that changes cavity size.
Chodera: But are all interesting cases caveats?
?: What is the problems created by not including polarization costs in binding free energy calculations?
Bernie Brooks: Just because you add polarization, doesn't mean you're doing better. For example, AMOEBA had trouble with SAMPL4 competition. Now working on doing free energy calculations with high level QM.
David Case: Fixed-charge forcefields have been fit to data which does include polarization effects, so it's not like we're leaving out these effects completely. Not nearly as much effort has gone into polarizable forcefields, so probably all worse (less mature) than fixed-charge forcefields.
?: Let's take the technology we have in this room right now: How many projects in pharma today with sufficient structural info could this be applied to now?
?: If we think of med chem teams, they usually get data one day and start design compounds immediately. Some of those compounds get made within 24h, so FEP can't evaluate all compounds they want to make. The usefulness of FEP is where changes in molecule are not trivial synthetically but trivial by FEP. Nitrogen walks, 5-ring heterocycles, etc. Good for cases where prior SAR has made chemists "gun shy" about making compounds. These are the kinds of problems where FEP can be tested regularly in an industry setting.
Clara: Scaffold changes are also a good opportunity. But we shouldn't forget we also need a crystal structure for your target, which isn't true for all projects.
?: We probably have 40-50 projects (neuroscience targets) that would benefit from FEP.
?: If we were sure prediction is right, chemists would make it. But usually not the case.
Christopher Bayly: What might be a good thing for community is to take overall problem of lead optimization and identify certain kinds of problems that it will be applicable/useful to. We still don't imagine applying FEP across the board for the next few years. How can we formalize the ways in which FEP can best be used in lead optimization now? e.g. high-confidence prediction for synthetically costly transformation, or when we can only make 4 / 20 difficult compounds---try to maximize success. But hard for FEP to be competitive for easy transformations where synthesis can be immediate.
JW Feng: We have a model where vast majority of compounds are made by CRO in China. We give them a list of 20-30 compounds in their queue, but not all can be made at once. Could FEP prioritize list of 20-30 compounds to speed up the rate at which good leads are made/tested?
Review of discussion topics
- Mon afternoon: Pipelines and automation
- Tue lunch: Calculation standards and health
- Tue afternoon: Force fields and parametrization
- Wed lunch: Experimental datasets and constructing validation sets
- Wed afternoon: What do we do next?
Chipot: What happened to community efforts from previous workshops?
?: How are we going to decide that free energy methods are strategically important for drug design?