Every great scientific breakthrough starts the same way – with a question. This cycle – ask, explore, test, learn – aims to uncover truths from meticulous observation. In the case of drug discovery, it’s like trying to solve a puzzle with invisible pieces, and the process is unpredictable and expensive.
Developing new medicines is a torturously long and complex process, and the global failure rates are at around 90%, says Dr John Woodland, a research officer at UCT’s Holistic Drug Discovery and Development Centre, known as H3D. As the continent’s first and only drug discovery and development centre, H3D is now using AI to develop medicines for infectious diseases as well as building models that will improve the outcomes of treatment for African patients. Dr Jason Hlozek, an AI/ML investigator working in drug discovery at H3D, says the centre is using AI to improve its “wet” lab work, where hands-on biological and chemical experiments are done.
Called CAD, for computer assisted drug/ discovery, the goal is to help the team fail fast and fail cheaply, says Hlozek. The team aims to meet two criteria: does the drug work, and is it safe? “The process is very iterative so that we can identify potential project-killing points sooner rather than later,” he says. “We don't want to get all the way down the road before realising there’s a problem we can’t solve. The further you go on the drug development journey, the more expensive it gets. We’re talking about tens of millions of dollars to run a clinical trial, so you want to find issues as early as possible.”
The first prize would ultimately be for you to arrive at the doctor, they do a quick test and then they can discern immediately what kind of dose would be most effective based on specific biomarkers or your genetic makeup.
John Woodland, H3D
Hlozek says the team is building models based on the data it already has. “These models learn from previous research, what worked and what didn't, and then they make predictions. Drug discovery follows a design, make, test cycle. A team of chemists will come together and propose their ideas – maybe there are 100 compounds they want to test – but as humans, they can only make so many different compounds at a time. The model can tell them which to pursue first, based on previous experience. They’ll make them and test them and, as science goes, the results will probably answer some questions and reveal new questions that we need to ask.”
The goal of the modelling pipeline is to produce models that maximise predictive performance in the most automated way possible, says Hlozek. The pipeline has two main components: featurisation and estimation. Chemical structures need to be converted, or featurised, into numerical vectors for machine learning. Several ML models will then be trained, using Microsoft FLAML, a Python library used for an ML automation.
These ensemble models are then used to train a surrogate model, which is a simplified approximation of the more complex model. Because the model is so lightweight, it's hosted on a small internal VM at UCT. Hlozek says his team built a simple interface that the model plugs into, so the chemists don't have to run the model themselves. “It's a click and run.” This is important because drug discovery brings together people from different scientific disciplines, many without computational expertise, so the tech needs to be easy to use, he says.
The further you go on the drug development journey, the more expensive it gets. We’re talking about tens of millions of dollars to run a clinical trial, so you want to find issues as early as possible.
Jason Hlozek, H3D
Woodland says it wanted its models to be open source and accessible to its partners and scientists on the continent, so that they weren’t too onerous to run. It’s hoped that AI will help researchers make better decisions, speed up the process and “see things we can’t”. But Hlozek says AI won’t replace scientists. “The model is not dictating the science. It's meant to support the science. This is not an excuse to turn your brain off. You still have to be engaged and thinking critically. This is important because I don't think machines can replicate human creativity.” This computational capacity needs to be seen as just another tool in our toolkit, says Woodland. “It shouldn't be thought of as replacing anything.”
Despite bearing over 25% of the global disease burden, only around 4% of global clinical trials take place in Africa, due to a lack of skills and infrastructure. Typical models and algorithms are tailored for patients who are genetically different from African populations. The continent has the most genetic diversity relative to any other population group in the world, says Woodland.
This has important implications for medicines, he adds, because a patient in Africa might metabolise or break down specific enzymes very quickly or very slowly. “If we want a treatment to be successful, we need to be mindful of this and tailor the dose or the treatment regime according to their genetic makeup.” This ties into the idea of creating personalised medicine, but that’s still a long way off. “The first prize would ultimately be for you to arrive at the doctor, they do a quick test and then they can discern immediately what kind of dose would be most effective based on specific biomarkers or your genetic makeup.”
H3D is the only organisation doing drug discovery research on the continent. The centre focuses on developing new medicines for diseases that disproportionately affect people living on the African continent, such as malaria, TB and antibiotic- resistant infections. Since launching in 2010, the closest its come to creating a new medicine was the development of an antimalarial molecule that reached Phase II clinical trials before the project was paused. The compound, MMV048, was the first clinical candidate to come out of Africa with the potential to be used for malaria treatment and as a single-dose cure. This molecule targeted a lipid kinase enzyme, but then interacted with other enzymes in the body, resulting in toxicity issues. Rather than seeing this as a failure, Woodland describes it as a proof-of-concept. If anything, he says, it shows the global community that we can do drug development and discovery in Africa.
THE LAB IN A LOOP
A lab in a loop is a concept in scientific research that involves training AI models with large amounts of data generated from lab experiments and clinical studies. An iterative process that transfers knowledge from one project to the next, these models can deliver predictions and insights that inform what the next set of lab experiments should be. This process generates new data that can again be used to train the models, making them more accurate. The goal is to streamline the traditional trial-and-error approach for novel therapies.
SPEEDY VACCINES TO FIGHT COVID-19
If drug discovery is such a lengthy and complex process, how were Covid-19 vaccines created, tested and brought to market so quickly? H3D’s Dr John Woodland says Covid-19 was unique in that it affected people all over the world and there was tremendous political will and a massive amount of resources were mobilised to address it. “We're working in a disease space where there isn't, perhaps regrettably, that same enthusiasm or urgency to make things happen very quickly.” Woodland believes humanity was fortunate that Covid-19 was a coronavirus because it meant there was already a wealth of research. “That's why it’s so important that basic science is funded and supported because you never know when you could benefit from it or when a use case is going to come up.”
* Article first published on brainstorm.itweb.co.za
Share