Driving adoption of new methods

October 5, 2008

Image via WikipediaVery often you hear about new methods, often more computationally expensive, that are pegged as improvements to existing commonly used techniques. Many a “Google killer” comes to mind as do methods for predicting protein-ligand interactions, something I have a little more experience with. All such methods face a common challenge - they have to overcome both a mental block to trying out anything new when existing methods work well, as well as a need to demonstrate that a change will be worth the effort.

Take virtual screening for example. While methods for docking and scoring, esp the high throughput variety, are hardly limited to one big dominant player, a la Google in search, the concepts underlying docking and scoring fall into familiar territory for many people. Some years ago, scientists began experimenting with methods like MM-PBSA and LIE, hoping to come up with results that were based on more physical models of molecular recognition and using better sampling methods. This is a complex problem, but the hope was that you could improve upon existing techniques.

I would argue that such higher order methods have seen some level of acceptance, partly due to frustration with the incumbents, but not quite at the rate that many, including myself, would have hoped. Why is that?

I think there are a few reasons. Better underlying engines for molecular mechanics calculations and molecular dynamics simulations would help. But perhaps the biggest reason is the same old one. You need to leapfrog not just the incumbents in performance and accuracy, but you have to leapfrog chemists, who can whip out 50 compounds a week (that was the number someone from a big pharma company once told me we’d have to beat). Plus I think the approach should be different. With our current limitations we are limited to hierarchical approaches, gradually increasing the expense of our methods, till the improvement is not worth it. Hopefully we can come up with methods that allow us to stop thinking about enrichment and more about whether we can truly evaluate molecular recognition and find the best binding molecules.

So what brought this on? A paper by Julien Michel and Jonathan Essex on Hit Identiﬁcation and Binding Mode Predictions by Rigorous Free Energy Simulations. The paper, in addition to being from Jonathan Essex takes a look at the ability to select the best binding modes from conformations generated by a docking program across a set of structurally diverse ligands.

Docking is inexpensive. Scoring is more expensive, especially given the poor quality of results. If you can get a number of good starting points, and then evaluate the best binding modes, you have already done a good job of identifying potential candidates. If you can also differentiate between a set of diverse compounds you are really achieving that next step that methods must take. The paper is pragmatic and realizes the challenges. In my opinion, we are still not there from the performance point of view. I think there is a lot of promise and as part of a hierarchical approach we can definitely start deploying these methods, somewhat carefully. For lead optimization, where time is not that much of an issue, there is significantly more promise.

I am still somewhat concerned that we have not taken that next quantum leap in performance and capabilities. Perhaps we just can’t, but I’d sure like to see something different try and fail.

Your Monday evening simulations

Posted by Deepak Singh | Filed Under Drug Development, Modeling & Simulation

Amazon Recommends

On FriendFeed, this post was liked by 7 people and commented on 4 times show

Comments

Trackbacks

(Trackback URL)

Semantic randomness : business|bytes|genes|molecules
October 7, 2008 at 6:01 am

[...] site, which is not as compelling a proposition, especially since you don’t get the kind of quantum leap in ...