Why drug designers will be at a disadvantage without AI
Posted 16th September 2019 by Liv Sewell
We spoke to John Griffin, CSO at Numerate, to find out how AI is changing drug discovery and development. Numerate is a drug design company applying cutting-edge AI to transform the process of small molecule drug design.
How is Numerate’s approach to drug design innovative?
Our mission at Numerate is to transform the process of small molecule drug discovery. The traditional medicinal approach has been, well, traditional. It might represent the last important industrial process where physical prototypes are made and tested rather than generating and testing them first in silico. Think about engineering projects like building a bridge or making an automobile – they are not built and then tested to see whether they break before the final product is perfected.
The founders of Numerate saw an opportunity to leverage breakthroughs in artificial intelligence, data science and horizontally extensible cloud computing. This new approach allows computational predictions to replace many laboratory assays so fewer compounds need to be made and tested to identify high-quality drug candidates for therapeutic programs.
The approach Numerate has pioneered is based on ligand-based predictive modelling. In the ligand-based approach, predictive models are inferred by applying machine learning algorithms to training sets of small molecules that are known to display or not display a biological effect of interest, such as for example, activity against a target or activity against an off-target, membrane-penetration, or residence in the brain after oral administration. These models can then be applied, as in silico assays, to virtual compound libraries so they do not need to be made prior to testing. This of course saves time and great expense.
We have extended our technological platform to cover the entire discovery pipeline from target identification all the way through optimization of a candidate for IND-enabling preclinical studies. Given the nature of the ligand-based approach, it is not confined to any therapeutic area, be it cardiovascular medicine or cancer. Nor is it confined to a particular class of target – enzymes versus receptors versus ion channels – so long as there exists data, or data can be obtained, to train the predictive models that one needs for a particular program. The ligand-based approach has the further advantage of being able to model the functional effect of interest, such as receptor agonism or antagonism, rather than simply binding.
Given this powerful platform, the primary focus for our therapeutic programs is on emerging targets that offer the potential to treat serious unmet medical needs in new or better ways. We have, to date, sought to generate scarce, first-in-class assets to form the basis for partnerships with larger companies that have expertise in drug development. The next phase in Numerate will be advancing programs ourselves and not just focusing on discovery.
Is there any hesitancy regarding the integration of AI and machine learning into the medicinal chemistry process?
There is some. But I think that people now realise they will be at a disadvantage if they do not understand how to productively integrate AI and machine learning into drug hunting. These techniques support superior performance and its important scientists realise that. It is an evolution of our industry. It is also a response to the underlying inefficiency of the traditional process. Drug discovery programs, on average, return less than the cost of capital, and soon it is projected that the return rate will be less than zero.
AI and machine learning methods are starting to find applications in other aspects of the pharmaceutical industry from manufacturing to supply chain management to clinical trial design, patient selection and clinical trial execution. These are worth learning about rather than being concerned about.
Are there any other barriers for AI integration within the process?
It is not just, “Here’s another added input I can put on a spreadsheet when I’m making decisions”. It is the underlying process that changes the compounds that appear on your spreadsheet and the number of compounds. It is a substantially different process that “thinks” about which molecules get made and tested, and log units more than the traditional process. Using AI we can screen a billion compounds and prioritize the small number that are going to have the greatest probability of success at any given stage of drug discovery.
As many small companies do not have a lot of capital, they cannot engage in huge expenditure to build large laboratories. By necessity the capital footprint of our work is light. For example, we do not need to have an HTS facility or library because our large-scale screening is done in silico and the tens of thousands of computers we use to perform that work do not even belong to us. They are somewhere in the cloud, belonging to Google or Amazon Web Services.
As for large companies, my perception is that they are being very thoughtful and circumspect about process re-engineering around AI; particularly given the breadth of potential applications and facts on the ground of R&D infrastructures that have been established and honed over the last 50 to 100 years just on a whim.
What are your predictions for the next 5 to 10 years in the field?
Ten years ago, you could count the number of companies truly using AI on one hand, and that would have included Numerate. There has been a hundredfold or so increase since then. Before long, there will be a thousand companies you can put under the umbrella of AI in healthcare. Seeing this coming, Numerate’s CTO Brandon Allgood (who is speaking at the 3rd Global Pharma R&D Informatics and AI Congress: London), along with colleagues at several other AI-focused companies and organisations, have founded the Alliance for AI in Healthcare (AAIH) as a forum for the many and varied stakeholders in this arena.
There will be increasing integration and recognition of the current value and potential for these systems across pharma and biotech. Small company start-ups will be successful in delivering product candidates to partners, even if they do not take them into development themselves, which poses the single-asset risk for platform technology companies.
Companies will have success applying AI techniques to deliver valuable product candidates into corporate pipelines. There will be clear indications that one is able to better select and then treat patients in human clinical trials using AI-based methods in the context of drug development. And there will be more acceptance and drive to incorporate those methods because the people who do not will be at a competitive disadvantage.
None of the companies that exist today are at critical mass. As the community grows and learns more about one another, we will find consolidation between companies that have strong synergistic capabilities. For example, companies will identify companion companies with highly complementary technologies spanning from target identification to drug discovery to drug development then commercialisation.
Technologically speaking, it is not just the AI and machine learning algorithms that have changed along with the abilities of the cloud, but the data. Without the right data to apply the advanced algorithms to you cannot do anything.
The very newest machine learning techniques, which are based on deep learning architectures, are typically very data hungry. Some data is not enough. Larger datasets of various kinds are now being generated through government-funded and charity-funded efforts like the Chan-Zuckerberg Initiative. They are putting a lot of money into generating an atlas of all the cells in all the organs in the human body in different age states and eventually in the state of disease. The single cell sequencing and characterisation techniques central to this work are revolutionary from a biology point of view, but the amount of data they are going to be producing that could be utilised and analysed using AI techniques will be breath-taking. It will give us a completely different perspective and set of tools.
John Griffin is CSO at Numerate, based in San Francisco, California.
John will be facilitating a roundtable discussion on AI-driven drug discovery at the upcoming 4th Medicinal Chemistry & Protein Degradation Summit. You can read the agenda here.
Leave a Reply