A Zika drug discovery collaboration

Today a nice collaboration has lead to submission of an opinion article to F1000Research. As an experiment I have posted the manuscript and some of the supplemental data online at FigShare. I hope to add more as time permits. This is definitely a work in progress.

This project has consumed lots of time from all concerned, but we hope the goal will lead to a molecule for Zika and we think we have shown how it can be done. The experimental part has begun. During the past week and a half there have been hundreds of tweets and emails and many scientists have offered help and suggested papers to read etc. A big thank you to all. Much of the writing was done collaboratively on GoogleDocs and that made for a compulsive effort which is very different to the relative immersion in my own thoughts before shipping a version to collaborators.

What this effort also showed was that a ground up effort can work, that social media can be used to connect scientists and help foster these efforts. While several of the scientists are long standing collaborators we have brought in a few new ones. Several scientists while too busy to contribute to the writing made useful suggestions and helped get the ball rolling and that for me was a huge motivation to do this.

Several of the scientific groups involved are in Brazil, they are working on this and we need the world to realize there are certainly other efforts besides those focused on vaccines. From communications with the NIH NIAID program managers it is clear we will not have an in vitro assay available until mid march. Brazil may have one in days. I urge more collaboration and discussion at higher levels so that funding can be prevented from reinventing the wheel.

The response time to get drug discovery off the ground needs improvement. Scientists have the know how and ideas. Now please provide them with the resources to do it and government can deal with the regulatory and incentives to get pharma involved in manufacturing and distribution etc.



Zika Virus – getting some research moving on drug discovery

Following my last post / plea I have been a little preoccupied trying to find scientists that could : 1. contribute to an open Zika virus drug discovery effort 2. Help write a perspective that summarizes potential drug discovery efforts .

Fortunately several folks connected by email or twitter to offer ideas and assistance. I have a student that kindly signed up for drawing some molecules from Dengue papers. Also I am aware of labs with expertise in setting up assays and screening that could all help.

In addition I have started on the computational side to build a homology model of a protein for which there is a Dengue virus crystal structure. Fortunately there are tools like SwissModel that are pretty straightforward to use, though I have never used it before. I figure no time like the present. Following this I tried docking a few compounds of interest, notably quinacrine which is known to inhibit Dengue virus in vitro .

Now I have a library of compounds docking using Discovery Studio. So will see what  kinds of molecules score well and hopefully will then post the results on figshare.

In parallel some emails to program officers at the NIH have lead to a few more contacts to see if there are any funds available in a shorter time frame that the notification NIAID sent out previously. I have also tweeted to several journals to see if they could open up the Dengue literature and also make repositories for Zika information.

All of this makes me wonder are there others doing the same and how do we prevent duplication such as screening the same libraries of drugs. So how can we ensure sharing of reagents and resources? Its not like there is anything other than altruism motivating us at this point. The last thing we probably need is a competition to develop a drug for Zika, but what we do need however is some funding to cover materials and staff costs.

To date Zika research has probably received zero dollars from the NIH. That may have to change pretty quickly – but what will suffer in return? Will anyone else step up to the plate like BMGF, WELLCOME TRUST etc….a quick search of their sites returned zero hits for Zika. For many its still not on the radar. At what point will it be?

As scientists many of us see how we can help, we have the ideas, some resources and yet the pieces missing are what will grease the wheels (we have to drop everything to work on this) , and where will we put the information we create? If we discover a drug how can we quickly get it to the clinic, what mechanisms for clinical testing will be in place? So many questions but does anyone have answers.

We hope to quickly put out a perspective on what we think could be done on the drug discovery side..hopefully.

So many questions. BTW Like many we have day jobs to do. Whose day job is it to respond to threats like Zika and develop drugs for emerging pathogens, because I would like to meet these people and pass over the baton.


Creating a rapid response – relevant to Zika virus

I have been following the news on the Zika virus over the last few weeks as it has spread with travelers returning from trips and women giving birth to children with microcephaly. Along with it is a spreading underlying fear of where it will lead. Its not so long ago that Ebola, was grabbing headlines. The world was generally very slow to react and that lead to large numbers of fatalities and lingering economic effects. It reminds me of something Christopher Southan, Megan Coffee and myself had written in 2014-2015 on finding small molecules for the next Ebola.  Although Zika virus is a flavivirus that is transmitted by mosquitos rather than an ebolavirus the steps we proposed are entirely transferable to this or any other pathogen. I will not go into detail as the article is open. It can be summarized here as bullets:

Text mine the literature

Patent assignees and/or inventors should openly declare their relevant filings

Reagents and assays could be commoditized

Using manual curation to enhance database links

Engage database and curation teams

Consider Open Science approaches

Adapt the “box” model for shareable reference compounds

Involve the physician’s perspective

Of course its impossible to predict which virus is going to grab global attention next but it does raise the question will we be ready and if so how will nations respond.  Currently most of the press is talking about vaccines and companies are accessing if their platforms will be useful. The USA is meant to have a plan but what is currently being done? And what about the rest of the world?

Perhaps what the world needs is a an international rapid response group that can identify what data and labs can be best suited to tackling a pathogen, rather than these large organizations which are not moving fast. For Example WHO came under attack for the glacial pace of their response to Ebola.

In the case of Ebola there were several high throughput screens of FDA approved drugs in the preceding few years before the outbreak in 2014 which lead to leads that were largely ignored. For Zika virus the closest screening is for compounds against Dengue virus, yellow fever, hepatitis C etc which are in the same family. This  kind of information and the molecules need to be collated so that those involved in screening can go straight to these.

The NIH NIAID responded by sending out a notice of interest for their regular grant programs. So by the time scientists apply, go through peer review and stand a slim chance of funding they may actually start their work within a year if lucky if there are no government close downs. So realistically these grant mechanisms will not bear fruit for several years. There has to be another way? There are probably few labs suited for working on Zika because of the biohazard, (fewer still work on just this family) but it would mean setting aside current research and trying to develop drugs or vaccines instead of say for Dengue or other viruses.

Which brings me back to why we cannot repurpose compounds we already have from say Hepatitis C or with activity against other viruses, parasites or pathogens. As we found with the potent antimalarial pyronaridine this compound was active against Chagas disease and Ebola virus. There must be other examples of molecules like this that can be used across pathogens. But are we even looking?

How can we expedite the research? I think scientists can certainly volunteer any time and resources and perhaps generate the preliminary work that’s needed here. If anyone is interested perhaps we can coordinate it online. The NIH needs faster and more flexible funding and peer review mechanisms to actually resource the work that is needed. When there are challenges like this and Ebola, there has to be some leeway to think outside the box before its too late.

So lets get going, there is no time to waste.




Lets do it all again updating predictions for the ‘real’ BIA 10-2474

After todays unmasking of the ‘real’ BIA 10-2474 I ran it through some of the online tools used earlier and now I have found software issues – that darn pyridine 1 -oxide is not liked by some of the websites probably because they use the same SMILES to structure software plugin?

SwissTarget Prediction identifies FAAH1 top followed by histone deacetylases. SEA and Molinspiration have issues with the N-O and produces a N=O so that predictions are spurious.  MetaPrint2D-React proposes hydroxylation on the cyclohexyl, HitPick suggests a lipase is the target(?), and similarity against a library of synthetic cannabinoids suggests that this molecule is not very similar with a maximal Tanimoto similarity of 0.6.

So no real clues – frustrating that SEA is not able to convert to the right structure and generate a prediction. SwissTargetPrediction still suggests the compound may hit other targets besides FAAH1. Which makes one ask is the N-oxide forming a protein adduct also? Or is something else responsible for the toxicity? Why did they use an pyridine 1 -oxide as a potential drug?

++Thanks to Alex Clark I can now add the predictions with the Open Bayesian Models. Here is a screen shot along with an image of the clustering for the highest predicted human target macrophage-stimulating protein receptor.

new 10-2474 and swisstargetpred

New 10-2474 and SEA prob with structureNew 10-2474 and molinspiration note struct issuenew 10-2474 metaprint 2D reactcorrected 10-2474 hit pickcorrectewd 10-2474 and similarity to cannabinoidsNew 10-2474 and open Bayesianscluster new


is the structure of BIA 10-2474 actually BIA 10-2474

Well the protocol released by Bial for the clinical trial lists a molecule 3-(1-(cyclohexyl(methyl)carbamoyl)-1H-imidazol-4-yl)pyridine 1-oxide that when run through online software to convert it looks nothing like the one used previously and posted on wikipedia etc. I used 2 websites to convert name to structure Opsin and Openmolecules.org

This looks like a puzzle that could run and run – which structure is correct? – what is the molecule that was actually used in the clinic? There definitely needs to be more eyes on this?

All bets are off as to the structure used??

OPSINname to structure for cpd from protocol


Differences between similar FAAH inhibitors and their in silico target predictions

**UPDATE As of 21st Jan 2016 please note the actual structure of BIA 10-2474 changed to this.

The continuing follow up to the clinical trial disaster.

Chris Southan yesterday suggested I try CID 57880883 as the comparator to CID 54576693 (JNJ-42165279).  These have IC50 75 nM vs 350 nM, respectively according to Jannsen’s WO2011139951. Yesterday I ran predictions for JNJ-42165279 using SEA and SwissTargetPrediction.Today I grabbed the SMILES for CID 57880883 and ran the predictions with results below.

Interestingly the addition of chlorine dramatically changes the predictions with SEA suggesting an interaction with 5HT-2B ranking above FAAH. SwissTargetPrediction does not seem as sensitive producing very similar results to CID 54576693 (clinical trial with this compound cancelled yesterday) at the top of the ranking. Lower down there are differences in targets suggested and ranked.

It would be interesting to see if J&J mentioned issues with 5HT-2B for CID 57880883. Again both compounds have a very different predicted profile to BIA 10-2474. A thorough profiling perhaps across these and other targets is warranted.

Again this underlies perhaps off-target issues and how some molecules in the same class may behave differently – obviously there are structural differences between the Bial compound and the J&J compounds, let alone metabolism differences that should also be considered.

CID 57880883 and SEACID 57880883 and Swiss target predictionimages of molecules


BIA 10-2474 is not that similar to synthetic cannabinoids and comparison to another FAAH compound

**UPDATE As of 21st Jan 2016 please note the actual structure of BIA 10-2474 changed to this.

Lauren commented on an earlier post about synthetic cannabinoid toxicities and stroke potential. I took the structure of BIA 10-2474 and looked at the similarity to  a virtual library of synthetic cannabinoids which Matthew Krasowski and I had published on in 2014. When I compare this compound to another previous FAAH clinical compound from J&J JNJ-42165279  based on MDL fingerprints and Tanimoto similarity, the latter compound has a slightly higher maximal similarity (0.66 vs 0.62)  to compounds in this library (see second table below). I am just showing the top 34 compounds for clarity. These similarity values are not that high (1 would be identical) so perhaps comparison to synthetic cannabinoids and their potential toxicity issues is a bit of a stretch.

For comparison to earlier predictions I have now run the same J&J compound through some of the predictive tools used previously to suggest targets / off targets..Unlike with BIA 10-2474 FAAH1 is right at the top of the list with SEA and Swiss target prediction  sites. BIA 10-2474 had very different target rankings in both SEA and Swiss Target prediction and also stood out in our first analysis which drove this analysis initially. These in silico target predictions point to a different target profile for BIA 10-2474 when compared to the J&J compound. This could obviously get quite complex if we start predicting possible metabolites as well.

10-2474 similarity to cannabinoidsJnJ cpd similarity to cannabinoids

JnJ cpd in SEAJNJ with Swiss target pred


Where will BIA 10-2474 lead us?

**UPDATE As of 21st Jan 2016 please note the actual structure of BIA 10-2474 changed to this.

My “Armchair science” / speculative contributions on this blog for the clinical trial disaster with BIA 10-2474 got a bit of notice over the past few days at Forbes and C&E News. There have also been discussions with reporters at the Boston Globe and Nature, so lets see where they lead, if anywhere. But what is interesting from my perspective is each time I chat with someone outside it puts a slightly different perspective on what motivated a role in this crowdsourced evaluation. From one side its what I do anyway. Try to understand what molecules are doing computationally. Normally I would use models built for a specific target or an activity measured in whole cells. Then I would use the model to find new molecules that might be worth testing. I also have a nearly 20 year interest in predicting toxicity so its not as if I am new to this. But I do not think any of the reporters even Google me or search in PubMed, or even know what computational toxicology is. People are perhaps too busy.

The challenge here is that BIA 10-2474 is a black box, we do not even have 100% proof its the compound or some human error that caused the disaster. The timeline on figuring this out is fluid, what can be done in the meantime?

So to me this is not armchair science, its gone beyond that, it is my attempt at trying to contribute to identify what this molecule can do, what it might have done in the human clinical trial subjects. I think the foundation for making predictions using published data is real science. Some might argue that drawing a structure or a SMILES string into a computer  and pressing a ‘button’ to make predictions is science or a career but I would beg to differ.

Some of the models I have highlighted make use of ChEMBL which has carefully compiled molecules and their binding data to targets for > 1 M molecules. This is big data that s been used to create machine learning models. Leveraging data like this seems obvious to try to figure out what a novel compound might ‘hit’. There are many other tools that members of the scientific community have developed and could apply, so it will be interesting to see some of the papers that come out in the future.

Other questions arose during discussions and after, like why should I do this? I don’t work for the company involved. For my own personal interest in using the various tools and models that are out there I want to see what they come up with. When push comes to shove can any of these tools even work?

Also the way this unfolded over a holiday weekend in the US, made me think that the traditional way of peer reviewed science would not cope in this instance until weeks/months after the event. As the various scientists were trying to identify the molecule and then we run it through different apps and online tools prior to making the data available, moves this outside of the traditional paper / publication scenario. Its like an open lab notebook in that the blog posts are showing you the thinking, results and ideas as they happen (well at least after they are typed). The others in the community have free access to this information which they can link to or not as they come up with their own posts.

So all in all, its a definite change from what one would normally do in keeping your ideas private or sharing with close collaborators, building up the results and experimental verification and then doing the analysis, write up and publishing. In this case its flipped. Once the structure of the candidate compound was postulated and put in Wikipedia and on Chris Southan’s blog it was open season.

The closest I have come to this scenario was the work on trying to understand the targets for the FDA drugs that were active against Ebola, as the initial work was posted on FigShare etc. and ultimately resulted in a publication. In the case of BIA 10-2474, unless there is compound that is made available or synthesized we cannot do the experiments to verify the predictions. As of today I do not see a clear path to publishing these experiences in a ‘journal’ but then this might change.

So where will BIA 10-2474 lead? For one it might make people think about what other tools can be used to get insights on a molecule before human testing. Perhaps we should be exploiting Big Data to predict toxicity, off targets as well as potential metabolites that might not have been predicted in liver microsomes. I will be giving a talk next Monday at SLAS 2016 in San Diego and this experience fits nicely into the title ‘Ensuring Chemical Structure, Biological Data and Computational Model Quality’. Between now and then perhaps there will be other information that will become available. In addition there are certainly many more computational models and tools that could be evaluated. The question is will this tragedy persist in the news cycle or will it fade away until it happens again with another trial. What can we learn in the interim? Because that is were our efforts will have most value, preventing it from happening again for a very long time.




Using more online software to make predictions for BIA 10-2474

**UPDATE As of 21st Jan 2016 please note the actual structure of BIA 10-2474 changed to this.


To continue the theme of the last few posts on the BIA 10-2474 compound in which I have used over 2000 Bayesian models and an App as well as another groups target prediction tool. To take this a bit further I went to the Click2Drug site to find other examples of software I might be able to use to predict potential interactions for this compound. I tried several of the tools under target prediction that did not require registration.

One of the first I tried was SwissTargetPrediction – this one predicted FAAH as the target and was followed by a list of others such as Sterol O-acyltransferases,  histone deacetylases etc.

Another tool called HitPick selected only the serotonin 2A receptor. Several other tools either failed to predict any targets or just seemed to hang. Overall efforts with these tools was pretty disappointing which was unexpected as virtually all of the approaches have resulted in papers.

Next up I went to an old favourite in MetaPrint2D-React to predict sites of metabolism based on a database of known metabolism data for many molecules. David Kroll’s recent blog was concerned about anilines as impurities. When the BIA 10-2474 compound is input (as SMILES) into MetaPrint2D it highlights a red area as being a high probability site of metabolism. If the C=O is dealkylated then it will form an aniline. So could it be possible that the molecule is metabolized to this metabolite and that this is contributing to the toxicity?

Some screenshots of the freely available software is shown below. Certainly there are many other methods that could be used including docking tools, and I have not tried those yet.

swiss target predictionhitpickmetapredictdealkylation


BIA 10-2474 – more predictions

**UPDATE As of 21st Jan 2016 please note the actual structure of BIA 10-2474 changed to this.


I thought it perhaps worth trying other computational approaches with BIA 10-2474. First up is the similarity ensemble analysis (SEA)  from the Shoichet lab. This also makes use of ChEMBL (version 16 data).SEA scores

This seems to agree to some extent with the previous analysis with the Bayesian models from Alex Clark and the Polypharma app small subset, in that Ephrin type B receptor 4 is ranked high as well as Cytochrome P450 17A1 followed by an array of kinases. Notice that in SEA that it provides an E-value. There may be some differences in data between version 16 and 20 which could account for differences.

Older posts «