Pfizer Researcher Paints Industry ‘Pain Points’ on HBS Panel

The ostensible panel topic was “Bioinformatics: Is There a Winning Business Model?” and Bio-IT World will cover the full discussion in its March issue. But Huang answered the question by drawing a portrait of pharma’s challenge and recent progress in its use of informatics tools.

To some extent, Huang’s fellow panelists – Peter Barrett, senior partner, Atlas Venture; Keith Elliston, CEO, Genstruct (biosimulation); Ramon Felciano, CTO, Ingenuity Systems (pathways); John Russell, executive editor, Bio-IT World – rely on pharma for their survival. Huang joked he felt he had a “bull’s-eye” painted on his chest. His unscripted answer, presented here, is a fascinating glimpse into how pharma sees itself.

Huang: The way I’ll answer is I’ll tell you about the pain points we face in pharma and you can draw your own conclusions and I’ll lead you through a few areas. The number one enemy facing our industry isn’t so much Canadian importation or possible regulation on price. It’s our own drug candidate attrition. Within Pfizer, and I think it’s representative of the industry, the odds of a clinical candidate – that’s when you’re done making the molecule and you’re sending it off to see whether it works and is safe and can make approval – is 1 in 25. That’s something like 96 percent failure rate. It’s staggering number, [when] coupled with the cost of R&D versus the productivity measured in new chemical entities, which is essentially flat.

It’s an unsustainable business model for Pfizer and the industry because you can model [the results] very explicitly. You know what you expect to get through your pipeline, modeling the attrition along the way, and you know what the shareholders are expecting. That’s why the industry is consolidating. That’s why the M&As have been happening. The underlying problem is we cannot solve the attrition problem just by throwing more revenue at it.

The reductionist model of drug discovery where we screen specific molecular targets hoping to get efficacious molecules has seduced us to this path where you oversimplify the problem. If you get enough shots on goal, you get this particular target that’s going to solve this particular human health issue. The odds are really low unless you’re working on another me-too product. How many really can the market bear? We need to be able to look at the in vivo effects and find out the cost of the experiments either in animals or humans.

It’s not about shots on goal for us anymore. The way we see it, if you want to double your productivity, you don’t double the number of your biologists and chemists, but by decreasing your attrition [by] four percent you actually double your productivity. So smarter shots on goal. So that’s why [we’re interested in] predictive models [and informatics tools]. It is to provide that molecular insight, the things that we find out too late or too costly, especially when the compound is already baked, the chemistry has already moved on, and you’re just going to live or die by that molecule.

Taking the Plunge
So that’s a scary moment when you stop chemistry. To the best of your knowledge you’ve baked into the molecule your belief, your knowledge, [and] as far you know, this thing is going to be safe and efficacious. [We know a little] about efficacy sometimes, in a therapeutic area, but [on] the safety, we’re really flying blind. We have an embarrassing number of in vitro assays [trying] to keep up with chemistry before it stops and then you find out the bad news.

On the Genstruct side, we have a molecule that works in animals but we actually didn’t know exactly how it worked. This harkens back to the good old days of drug discovery, you know, 50 years ago, where it was called feed and bleed. You had spontaneously hypertensive rats or whatever the animal model was. You only had a screening file of a few thousands compounds, and you actually just rammed them down the throats of the rats and see whether you had an in vivo effect. And you actually address a lot of things at once [doing this] because it integrates so much. You’ve got in vivo availability, early signs of safety, and so if one of your thousand compounds works, off you go to the clinic.

So then molecular biology came around – high-throughput screening, combinatorial chemistry – now we’re literally throwing two and half million compounds against a molecular target, hoping that it’s going to give you the efficacies that you need, whereas you don’t know the safety of the molecule, you don’t even know the safety of the target in most cases. So back to Genstruct. It’s sort of like back to the future type thing where we have a molecule that works in the animal and we don’t exactly know the mechanism, we don’t know what targets, we don’t exactly know the pathway. Genstruct [comes] back with hypotheses with our data. We give them data. They do their black box thing and they come back with a report and say here are some testable hypotheses you can do that are consistent with the experiment you gave us and our knowledge base. That has value. It means we can follow that up. Obviously there was enough material in there that we’ve extended our collaboration.

With Ingenuity, I think this speaks to the ’omics departments in Pfizer who are [staffed by] generalists, [that is] they’re specialists in the technology but they don’t have a lot of therapeutic area knowledge. If you go to any Pfizer site, you’re going to see therapeutic areas organized by oncology, by inflammation, or whatever. Inside you’ll see biologists who read the literature, who have models in their head(s) about what’s interacting with what. It’s sort of like this Rube Goldberg machine, and they’ll [biologists] see an ‘ah-ha, it’s going to arrest tumor development’ or whatever. It’s all in their heads but they don’t have time to actually codify that into any computational model. Nor are they interested and they’re probably suspicious of the ‘omics data sets to begin with because ‘that’s too newfangled, what does it have to do with me?’ type of thing.

Stemming the Data Flood
So what Ingenuity and tools like Ingenuity allow Pfizer to do is [allow] these technologists who spend their lives generating ‘omics data sets, computational biologists who don’t go into a lab but work on a computer, they can present their experiments in a context that sort of makes sense to therapeutic area biologists. Because it’s able to assimilate you know Affymetrix or Agilent data sets and present [the data] in a more or less biology-friendly way, and I agree with Ramon [Ingenuity] that their platform is extremely user-friendly. In some cases you can actually ask for a canonical pathway. And then [Pfizer researchers] say, ‘Hmm, OK, I might be able to follow up.’

The last thing I’ll say is around safety sciences. I alluded to this 1 in 25 success rate in the clinic, and the majority of [the failures] is because we can’t predict safety in vitro. So, not represented at the table [panel] are companies that [are] safety sciences organizations, which have a bioinformatics framework. So [a company in] in Cambridge, U.K. [is] working with our safety sciences organization building ontology by doing text mining, literature mining, to understand what are the molecular culprits that might lead to this form of toxicity that’s sort of weighing us down.

There’s another collaboration [doing] a very similar sort of ontology building so we can work back, sift through the collective knowledge in the literature, to highlight areas – maybe this is the molecular culprit that’s causing this form of toxicity. You develop an assay for that [and] you can screen alongside in the discovery program while you’re still doing chemistry to get rid of that problem. So I think there are many business models if you can somehow address the [attrition and toxicity] problem. [What] can you do to use the Pfizer in-house data or combine it with external literature to illuminate possible mechanism of toxicity. That’s how you get from the four percent to eight percent success rate [and] you’re actually doubling your productivity.