"The four structures of AmpC determined with the new docking hits are available from the PDB with accession numbers 6DPZ, 6DPY, 6DPX and 6DPT."
Are we in a situation now where, if I have a bad anti-botic resistant infection I can just order these molecules on the off chance that they will help me?
Can I order a toxin?
Or are these molecules just impractical to use outside of a lab setting
They only determined whether those specific molecules bind to the target. They didn't test if they are toxic, if they are able to actually get to the target, how stable they are under real conditions, into which metabolites they are processed in humans, ...
This is the very first step towards developing potential drugs, it's very, very far from an actual drug. And drug development wasn't the goal of this paper anyway.
And the idea behind this paper was to find molecules that aren't in any catalogue, so you would still have to synthesize them yourself or pay someone to do a custom synthesis for you.
I thought the point of the library of molecules used was the supplier (Enamine) has a systematic method of synthesizing the molecules they've listed. It appears that, in practice, they can supply 90% of the molecules they offer, synthesized on demand:
> Of the 589 molecules selected, 549 (93%) were successfully synthesized (Supplementary Table 10 and Supplementary Data 11, 13)
> Over the past decade, Kiev-based Enamine Ltd has innovated an efficient pipeline to produce any of over a billion never-before-made drug-like compounds on demand — at a cost of about $100 per molecule — by combining any of tens of thousands of standard chemical building blocks with one another using over a hundred established chemical reactions.
'Per molecule' gives no real indication of cost, that appears to be a 'tooling' charge, to design the reaction chain. I doubt $100 will buy you any useful quantity of any research chemical, much less a custom mfg one.
> Are we in a situation now where, if I have a bad anti-botic resistant infection I can just order these molecules on the off chance that they will help me?
Very far from it. I wrote a docking program myself (the most popular one, currently, I believe), and I get emails of this nature sometimes.
Molecular docking only makes predictions about binding. Typically, this gives you what people call an "enrichment", i.e. your guesses are better than random, hopefully much more so. Still, most of these guesses are typically wrong. For example, you might go from 1 in 1000 odds to 1 in 10.
When you do find a drug-like molecule that binds your chosen protein, this is only the beginning of the drug development effort that may end up costing billions.
Hey Oleg, I just wanted to say thanks for doing some really great work. AutoDock Vina made a huge impact on the field, and its results are still on-par with commercial docking codes. Academic software development is a combination of some of the worst parts of open-source software development and the worst parts of academia, but projects like Vina motivated me to be a better coder and scientist in grad school.
When I see a paper like this one of my first thoughts is "What will Derek Lowe say about this?" He is a chemist in the drug industry and excellent writer. On his blog he regularly tears apart overhyped claims for how software searches for targets and automated synthesis of chemicals are going to find perfect cures for everything at the press of a button. His take on this paper is actually positive, though with some important caveats:
I have been out of this space for a few years (transitioned to data science from drug discovery), but from my time doing in silico and in vitro work, a major issue with docking was rank ordering. His comments are right on the mark IMO. Especially this paragraph:
>Another point is that high-middle-low effort on the D4 case. The binding assay results compared to the docking scores are shown at right. You can see that the number of potent compounds (better than 50% displacement, below that dashed line) decreases as the scores get worse; the lowest bin doesn’t have any at all. But at the same time, there are a few false-negative outliers with binding activity at pretty low scores, and at the other end of the scale, the top three bins look basically undistinguishable. So the broad strokes are there, but the details are of course smeared out a bit.
These methods can filter millions of compounds down to hundreds, but as an academic lab, it's still a herculean effort to synthesize hundreds of compounds. And out of those hundreds, you might get a couple that are active. This study is a combination hard work, yes, but also a lot of money and luck.
That being said, good for the team, and good for science. I have nothing but respect for Shoichet and Roth. Didn't ever cross paths with Irwin.
> I have nothing but respect for Shoichet and Roth. Didn't ever cross paths with Irwin.
This makes me laugh since Shoichet was childhood friends with Irwin and they've worked together on almost everything together since 2000 when Irwin went to Northwestern to join his lab.
I can tell you one key sentence from his post is "reports what I believe is the largest publicly disclosed effort of this type", where the key words are "publicly disclosed". From talking to pharma companies, they all have done virtual screening with larger libraries than what is reported in this paper. I'm not familiar with details of Novartis's efforts, but based on talks with other companies, I would not be surprised if Derek is familiar with virtual screens at Novartis that have been 1-2 orders of magnitude larger.
Many of the questions in virtual screening of large libraries are not answered in the paper as the authors did not look at them. As mentioned in the comments, they are using Enamine RealDB which has been around for a while, and Enamine has an even larger virtual library called RealSpace.
One other very interesting takeaway that Derek points out is that humans eyeballing the candidate molecule-protein interaction are somehow better at selecting good candidates than the computer:
"""
There’s also a human-versus-machine comparison in evaluating the hits. The authors took the top 1,000 compounds and selected 124 of them by eyeballing them for what looked like good interactions in the docking pose (not looking at the scoring), and took 114 molecules on the basis of docking scores alone. The hit rates for the two sets were almost identical (about 24%), but the human-selected ones were disproportionately potent – and indeed, in the two campaigns, the human-selected compounds were quite over-represented in the lists of potent compounds.
as a computational chemist I must say end-use might differ significantly from these screening studies, it is more of a statement on current computing capabilities and a little bit of science icing
Computing capabilities inside this field. You can throw as much cpu time as you want at a bad model and it won’t mean anything. The value here is that it’s not just bruteforcing every concievable compound but instead goes through a sequence of potentially synthesisable drugs which are interesting according to docking simulations.
They could likely simulate tons more, but there is no guarentee that they are; Actually synthesizable, actually hit the target, aren’t toxic.
So at some point the current computing capabilites fall shot. But not because we can’t throw more cpu hours at the problem, simply because we don’t have the computational tools to cross those bery important barriers available at all.
I developed the Exacycle idle computing system at Google and we ran extensive docking, with very good models of synthesizable drugs. Based on my in-depth knowledge of this field, along with extensive experience running a system that produced 1M CPU hours per hour 24/7 and made significant improvements in protein design and structure modelling.
Using more CPU time in this would have helped a lot.
Co-author here (AMA)--A large-scale docking screen of 116M molecules takes ~1100 cpu days on our cluster, working out to about 1 mol/sec, which is very fast for virtual screening. What this doesn't account for is this requires about 30 minutes per compound to precompute information (conformations, partial charges, etc.). So this works out to ~6M cpu/hours to prepare the library for screening, which is a substantial amount of computation. We're loading about 1M molecules a day and have a 2-3 year backlog of compounds to load from Enamine.
The good news is that once the library is prepared, it is quick to screen at more targets--and we make the pre-computed library available at zinc15.docking.org.
Interestingly, as the library grows a limiting factor is storing the library on disk. It is now ~20T. We've set up several mirrors around the world for groups that are actively using it. An interesting problem will be to see if preparing compounds for screening on the fly (e.g. with machine learning models) can overcome this limitation to keep up with library growth.
A big question for us is what will the return on investment in screening larger and larger libraries be? One of the take aways from this work is if docking has moderate enrichment, than screening larger libraries not only gives more hits but actually can increase the hit-rate for the top scoring compounds.
I know that docking using GPU is about an order of magnitude faster than CPU (see today's Schrodinger 2019-1 release notes, https://youtu.be/K4AYdBvuOe4?t=90). Is there a way of doing GPU accelerated precomputation though?
Hey Chris--We're right now using a mix of commercial and open source software like Omega, Corina, AMSOL, and Mol2DB. Probably the slowest step is generating the partial charges for each conformer with a reasonably high quality semi-empirical forcefield. I'm not sure if there are competitive (in terms of quality) GPU based methods, but if there were methods that were ~1000 times faster as can be the case for GPU based methods, it would definitely speed up the pre-computation or make on-the-fly prep feasible. Do you have any ideas of where we should look?
This is exactly why I typically have all sorts of prior rejection criterion to trim down my (relatively humble) set of 17 million trial molecules. I don't have a proper cluster like you-all do. So,'beggars' need to do the easy rejections early (e.g. >12 rotatable bonds? 11 H-bond donors? partition coefficient of 7.3? therefore no further consideration is needed) ADME before an expensive docking.
29 comments
[ 3.1 ms ] story [ 71.2 ms ] threadhttps://www.nature.com/articles/s41586-019-0917-9
Are we in a situation now where, if I have a bad anti-botic resistant infection I can just order these molecules on the off chance that they will help me?
Can I order a toxin?
Or are these molecules just impractical to use outside of a lab setting
This is the very first step towards developing potential drugs, it's very, very far from an actual drug. And drug development wasn't the goal of this paper anyway.
And the idea behind this paper was to find molecules that aren't in any catalogue, so you would still have to synthesize them yourself or pay someone to do a custom synthesis for you.
> Of the 589 molecules selected, 549 (93%) were successfully synthesized (Supplementary Table 10 and Supplementary Data 11, 13)
I fully agree your main points.
> Over the past decade, Kiev-based Enamine Ltd has innovated an efficient pipeline to produce any of over a billion never-before-made drug-like compounds on demand — at a cost of about $100 per molecule — by combining any of tens of thousands of standard chemical building blocks with one another using over a hundred established chemical reactions.
E.g. here is one of them:
https://www.rcsb.org/structure/6DPZ
Very far from it. I wrote a docking program myself (the most popular one, currently, I believe), and I get emails of this nature sometimes.
Molecular docking only makes predictions about binding. Typically, this gives you what people call an "enrichment", i.e. your guesses are better than random, hopefully much more so. Still, most of these guesses are typically wrong. For example, you might go from 1 in 1000 odds to 1 in 10.
When you do find a drug-like molecule that binds your chosen protein, this is only the beginning of the drug development effort that may end up costing billions.
http://blogs.sciencemag.org/pipeline/archives/2019/02/11/vir...
>Another point is that high-middle-low effort on the D4 case. The binding assay results compared to the docking scores are shown at right. You can see that the number of potent compounds (better than 50% displacement, below that dashed line) decreases as the scores get worse; the lowest bin doesn’t have any at all. But at the same time, there are a few false-negative outliers with binding activity at pretty low scores, and at the other end of the scale, the top three bins look basically undistinguishable. So the broad strokes are there, but the details are of course smeared out a bit.
These methods can filter millions of compounds down to hundreds, but as an academic lab, it's still a herculean effort to synthesize hundreds of compounds. And out of those hundreds, you might get a couple that are active. This study is a combination hard work, yes, but also a lot of money and luck.
That being said, good for the team, and good for science. I have nothing but respect for Shoichet and Roth. Didn't ever cross paths with Irwin.
This makes me laugh since Shoichet was childhood friends with Irwin and they've worked together on almost everything together since 2000 when Irwin went to Northwestern to join his lab.
Many of the questions in virtual screening of large libraries are not answered in the paper as the authors did not look at them. As mentioned in the comments, they are using Enamine RealDB which has been around for a while, and Enamine has an even larger virtual library called RealSpace.
"""
There’s also a human-versus-machine comparison in evaluating the hits. The authors took the top 1,000 compounds and selected 124 of them by eyeballing them for what looked like good interactions in the docking pose (not looking at the scoring), and took 114 molecules on the basis of docking scores alone. The hit rates for the two sets were almost identical (about 24%), but the human-selected ones were disproportionately potent – and indeed, in the two campaigns, the human-selected compounds were quite over-represented in the lists of potent compounds.
"""
They could likely simulate tons more, but there is no guarentee that they are; Actually synthesizable, actually hit the target, aren’t toxic.
So at some point the current computing capabilites fall shot. But not because we can’t throw more cpu hours at the problem, simply because we don’t have the computational tools to cross those bery important barriers available at all.
Using more CPU time in this would have helped a lot.
The good news is that once the library is prepared, it is quick to screen at more targets--and we make the pre-computed library available at zinc15.docking.org.
Interestingly, as the library grows a limiting factor is storing the library on disk. It is now ~20T. We've set up several mirrors around the world for groups that are actively using it. An interesting problem will be to see if preparing compounds for screening on the fly (e.g. with machine learning models) can overcome this limitation to keep up with library growth.
A big question for us is what will the return on investment in screening larger and larger libraries be? One of the take aways from this work is if docking has moderate enrichment, than screening larger libraries not only gives more hits but actually can increase the hit-rate for the top scoring compounds.