ADL: ADFR

Dr. Yaron Steinberg ymsteinberg at gmail.com
Sun Mar 6 11:00:26 PST 2016


Dear Prof. Sanner

Why would you highlight the CDK2 result, when your overall success
rate was 5/35 or 14%?

You started with 35 systems, gave up on half of them, and for for the
remaining 17, ADFR predicted the binding mode of 5 correctly.

What are we to use a method with a 14% overall success rate for?

To my mind, the most salient result of this study is that it shows
that one can not use docking methods to model receptor movement, as
many probably already knew.

A more nuanced approach should be used instead.

Ron


On Mon, Feb 29, 2016 at 2:27 PM, Michel Sanner <sanner at scripps.edu> wrote:
> Hello Ron
>
> On 2/29/16 12:50 AM, autodock-request at scripps.edu wrote:
>>
>> Message: 3
>> Date: Sun, 28 Feb 2016 15:03:14 -0800
>> From: "Dr. Yaron Steinberg"<ymsteinberg at gmail.com>
>> Subject: Re: ADL: Software Announcement
>> To:"autodock at scripps.edu"  <autodock at scripps.edu>
>> Message-ID:
>>
>> <CAFf7Xv8FEQu5R41913pnp5J0XZfG4_jQ6wsWH0EaM7LGGVKQ5w at mail.gmail.com>
>> Content-Type: text/plain; charset=UTF-8
>>
>> Interesting paper, but your reported Astex Diverse Set accuracy (74%)
>> is actually worse than the previously published results for rDock
>> (76%) and Vina (81%), although better than Glide's (68%).
>
> I agree, however I would like to clarify that ADFR is using the AutoDock4
> scoring function. Not surprisingly it shows identical performance to
> AutoDock4 for this test. The points we are making in this part of the paper
> was not about docking accuracy but rather about the higher efficiency of the
> genetic algorithm implemented in ADFR.
>>
>>
>> http://journals.plos.org/ploscompbiol/article?id=10.1371/journal.pcbi.1003571
>>
>> I am surprised by some of your conclusions though. ADFR uses 50 GA
>> runs at 4.2 hours each, or 210 hours for each ligand. This is 1/7000th
>> as fast as your own Vina results.
>>
>> When there are 12 residues, ADFR needs 430 hours for each ligand. I am
>> surprised by your conclusion that this means that ADFR scales well. It
>> is simply about equally slow in both cases.
>>
>> There must be many common components to both rigid and flexible
>> receptor docking, including the calculation of the ligand-receptor
>> energy and of the internal energy. The slower you make them, as ADFR
>> seems to have done here, the longer all experiments will take, and the
>> more equally slow everything will appear.
>
> The version of ADFR described in the paper only had the C implementation of
> the scoring function with every thing else implemented in Python, which
> explains the lengthy run times compared to the highly optimized C++
> implementation of Vina. The new version of ADFR that we just announced has
> unoptimized C++ implementations of the scoring function, the flexibility
> tree, and the local search procedure but still uses a Python implementation
> of the genetic algorithm and clustering procedures. Yet, it  already shows a
> 20-fold speedup over the version from the paper for the CDK2 data where we
> dock ligands into a receptor with 12  flexible side chains.
>
> here are the times (in hours) for the CDK2 data set
>
> Python    C++    exhaus. 8   exhaus. 20  exhaus. 200
>   ADFR     ADFR    Vina        Vina         Vina
> 379.98    19.16   1.85        5.05          46
>
> As you can see new version of ADFR partly ported to C++ is currently about
> 10 times slower than Vina with default exhaustiveness (8). however, Vina
> becomes 2.4 times slower with exhaustiveness 200 while still not achieving
> ADFR's success rate.
>>
>> In your cross-docking experiments, you say that you down-weight some
>> terms in ADFR, but what if you simply optimized the Vina or Autodock
>> force-fields for their best cross-docking accuracy too? Might this not
>> be simpler and more useful than publishing a new docking program
>> that's 1/7000th as fast as what we had before and shows worse Astex
>> Diverse Set accuracy?
>>
>> What happened to the custom and unified force-fields for Vina and
>> Autodock that were discussed in this mailing list during the last
>> fundraising campaign? I posted many suggestions then.
>>
>> Ron
>
>
> best regards
>
> --
>
> -----------------------------------------------------------------------
>    o
>   /   Michel F. Sanner Ph.D.            The Scripps Research Institute
> o     Associate Professor               Department of Molecular Biology
>   \                                     10550 North Torrey Pines Road
>    o  Tel. (858) 784-7742               La Jolla, CA 92037, TPC 26
>   /   Fax. (858) 784-2341
> o     sanner at scripps.edu                http://www.scripps.edu/~sanner
> -----------------------------------------------------------------------
>


More information about the autodock mailing list