DDW Editor Reece Armstrong explores how AI is being utilised in the lab and what the technology can really do within pharma.
Artificial intelligence (AI), a branch of computer science encompassing machine learning and deep learning, and which technologies such as automation, natural language processing (NLP) and image recognition all fall into, offers significant opportunity for the life sciences industry to change how it operates. The drug discovery and development process is costly and arduous, offering no certainty of success which often rests on whether a company has chosen the correct drug molecule and target. AI has long been touted as a kind of new frontier in life sciences, something that when applied to its full potential, could fundamentally change the way pharmaceutical companies develop medicines.
We’re still in the early days of the sector though, with the use of AI growing, but still sparse and largely represented by biotechs and technology companies looking to collaborate with pharma. With its use scattered across the industry, just how are researchers utilising AI to improve drug discovery?
The challenge of data
The pharmaceutical industry is built on data. Every successful therapy brings with it the necessary supporting data that confirms its efficacy and place on the market. In the drug discovery stage, researchers are burdened with the task of combing through vast amounts of datasets in order to find potential targets for their molecule. Technologies such as high-throughput screening (HTS), which can analyse vast amounts of chemical and biological compounds in order to expedite target identification, still produce hundreds of gigabytes worth of data, which researchers need to sift through in order to understand their compound.
Mark Davies, SVP of Informatics and Data at BenevolentAI describes the challenge scientists face. “Under the current model of discovering new medicines, scientists have to manually explore vast quantities of information. They must navigate different dimensions of disease and dysregulated mechanisms and pathways, reason through potential safety concerns and assess approaches that could lead to success in clinical trials. Even the most talented scientists cannot access and analyse the sheer volume of available data and biomedical literature needed to do this, all whilst grappling with the inherent complexity of human biology,” Davies says.
According to Davies, one of the benefits AI brings to researchers is its ability to help them tackle the explosion of data that technologies produce throughout drug discovery.
“AI approaches focused on information extraction and inference allows us to directly tackle this data volume challenge by pulling out the potential relationships between those entities, and in doing so, create new connections that allow scientists to better understand complex biology and point to new treatment approaches. As a result, AI can help address major drug development pain points, reducing costs and time to market,” he says.
Tom Sharrock, lead of AI Engagement at Lifebit agrees: “AI enables researchers to analyse and understand volumes of data on an order of magnitude greater than the same researchers would be capable of on their own. This enables conclusions to be drawn on available data faster with improved certainty, which in turn can improve experimentation and drawing conclusions from a hypothesis.”
Having this improved certainty as Sharrock describes is vital for drug discovery, especially when success rates are so low in the industry and costs are so high. With the average cost of bringing a new drug to market estimated between $985 million – $1.3 billion1, it makes sense that companies will be looking for solutions to both reduce costs and improve their chances of success.
For Sanjay Saraf, Head of Data and Analytics Product Management at Benchling, one of the main benefits that AI and informatics is bringing to researchers is “the increased speed of therapeutic production.”
“AI and informatics play a large role in driving efficiencies for researchers and lab technicians with faster decision making, as the data can reduce the number of unnecessary experiments carried out.”
“Another is enabling better scientific choices and being able to discover novel therapeutics that wouldn’t have been hypothesised previously. Finally, it allows for improved visibility into investment choices. Scientists no longer ‘throw something at the wall and hope it sticks.’ Instead, researchers can make calculated investment decisions with a rough likelihood of success,” Saraf adds.
Speaking of choices, the decision a company makes in terms of what indication its molecule will focus on can be make or break.
BenevolentAI’s Mark Davies says that “one of the main reasons why drugs fail is that they actually don’t work in the patients they are tested in.”
“This might be because the wrong target is picked, which will lead to the compound failing in the clinic. Improving success rates by selecting the right target will have a huge impact on the development of new medicines for patients and we are already starting to witness the impact of AI-driven approaches in this field of drug discovery,” he says.
“This is because AI is really good at spotting hidden patterns that are otherwise buried in vast quantities of data. Equally, AI might also uncover information that tells scientists that a target is not worth pursuing.”
For a growing number of years now there has been significant interest in the application of AI within pharma. Investment levels, alongside mergers and acquisitions have been steadily rising, pointing to the ongoing interest in the sector. Indeed, the market for AI applications in drug development Is estimated to rise to $5.1 billion by 2025, according to data from Emersion Insights2.
Mark Davies sees a “significant uptick in interest from pharma and biotechs,” which is being driven by a need to “ensure the quickest speed to market at the lowest possible cost.”
In turn, companies utilising AI in successful ways should make a return on investment, according to Davies, “as the likelihood of the correct target being discovered increases dramatically.”
“Suboptimal targets are a primary driver behind the failure of a drug at the clinical stage, and the use of AI should result in higher success rates and a quick return on investment,” he adds.
In terms of resources, Professor Roland Wiest & Dr Richard McKinley, lecturers in CAS AI in Medical Imaging at the Swiss Institute for Translational and Entrepreneurial Medicine, started out with a very modest set-up. When asked about the levels of investment required for AI, Wiest & McKinley say it depends entirely on the “kind of question being answered” and “how many people are working in the lab on AI.”
“For imaging we started in 2016 with a single gaming laptop, which was sufficient to make substantial headway in the problem of segmenting multiple sclerosis lesions. We have subsequently made substantial investments, which mean we can run multiple AI projects concurrently, but our infrastructure is of course far below what would be available in industrial labs.”
It’s important to remember though that AI, machine learning and informatics are all part of the same puzzle when it comes to expediting drug discovery and development. High volumes of data caused by advanced instrumentation, robotic automation, and new scientific techniques, are things that researchers must apply AI to in order to gain deeper understanding.
“Machine Learning cannot be effective without accurate and standardised data,” Sanjay Saraf says. The challenge however, according to Saraf, is having data in a central location and in a standardised format, especially when it is being stored is “disparate places.”
“For example, some R&D teams don’t yet have their systems in the cloud, and as a result they lack the computational horsepower or the skill set of data scientists to work effectively with data.
“To overcome this, many mechanisms are needed – for systemising the data that goes into the system, for identifying which features of the data are important, for training the ML model, for users to provide input and feedback on the output from the model, as well as other monitoring and management tools.”
Investing in these types of tools is something that Saraf says “every company looking to leverage AI and Machine Learning will have to make.”
“Without making the substantial investments into building both the teams and data architecture to support Machine Learning, businesses will fail to reap the benefits of these technologies,” Saraf explains.
For Dr Arne Kusserow of Merck KGaA,(Merck Group), the investment angle boils down to the simple point of you get what you put into it.
“To keep it simple, digitalisation, automatisation, Machine Learning and AI are intended to increase efficiency. If prices can be kept, this in turn leads to higher cross-margins and revenues. The history of industrialisation proves this. It’s not a reasonable question of how much money you have to invest for an invention that increases your efficiency, it is relevant if it pays off. If it pays off and others do it, while you don’t, your competitors gain efficiency and will outcompete you. It’s a matter of survival,” Dr Kusserow tells DDW.
The incredible speeds at which Covid-19 vaccines were successfully developed and brought to market highlighted to the world and the industry that lengthy development timelines aren’t always assured. The pandemic also highlighted the need for life sciences companies to adopt technologies that can help them work more efficiently. Whether this being remote technologies to enable at-home working, cloud platforms for data, or software for clinical trials so they could continue to operate.
BenevolentAI was able to witness this level of fast-tracked development when it used its platform to help discover a potential Covid-19 treatment. The company used its platform to identify an existing rheumatoid arthritis drug owned by Eli Lilly and Company and repurpose it into a treatment to prevent Covid-19 patients progressing to being placed on a ventilator. The drug, which was used in combination with remdesivir, was granted Emergency use Authorisation (EUA) by the FDA within a year of data being published.
“I think the pandemic has pushed the industry to recognise that a change is needed in the way drugs are discovered and developed. To have an AI-enabled prediction robustly validated in large, randomised control trials has really helped build trust, confidence and credibility for our tech and approach at BenevolentAI,” the company’s Mark Davies says.
Dr Arne Kusserow of Merck Group thinks the pandemic has absolutely had an effect on the AI sector.
According to Dr Kusserow, pharmaceutical companies are searching for solutions that enable them to reduce the “time needed for acquiring scientific data”, to share with a contract research organisation, and also to reduce the time a “regulating authority needs to understand the data.” If found, these solutions can remarkably reduce the time needed to bring a drug successfully through clinical trials and approval.
“This is the lesson learned: it is possible to bring drugs faster to market and it make a huge difference. We see a strongly increased demand for our solutions,” Dr Kusserow says.
Benchling’s Sanjay Saraf describes how at the start of the pandemic, “the majority of companies were slowed down by their lack of ability to do in-person research, but they quickly pivoted to ‘in silico’ design approaches for future therapeutics.” This digital approach enabled them to work faster and more efficiently than if they were in a lab.
Saraf cites examples of setting up computer models to show how an experiment would happen, rather than setting up and running that same experiment in-person. However, Saraf warns that this digital approach hasn’t resulted in a total shift of how pharma companies operate and that “scientists still need to test their therapeutics in the lab.”
“While the pandemic accelerated tech investment in biotech and awareness among consumers on the life-changing work pharma companies did in the fight against Covid, it may have increased time to market for most non-Covid related therapeutics,” Saraf adds.
The human element
As the technology currently stands, the majority of AI and machine learning technologies cannot operate by themselves, relying on data input or oversight from humans to ensure errors are kept to a minimum. This presents an interesting paradigm where both humans and AI systems are reliant on each other to get the best possible outcomes.
For researchers in the lab, Benchling’s Sanjay Saraf says that “while these tools can be used to solve extremely constrained problems, they need to be put into scientific and operational context by humans to be usable in the lab. For most analytical workflows, a human still needs to remain in the driver seat.”
“Over the course of a set of experiments, most researchers will rely on some degree of human intuition. That can be as simple as identifying an obvious outlier, or excluding a class of experiments because the mechanism of action has too much overlap with critical proteins.
While AI and Machine Learning as we know them today, can’t completely replace human intuition, it can analyse data, build knowledge and apply it to future problems. These technologies can also make new suggestions to either incorporate or ignore. To the degree that we consider these deductions ‘obvious’, we could say that AI and Machine Learning helps to reduce human error,” Saraf adds.
One of the other challenges that these technologies present is the potential of bias creeping into any algorithms or predictive systems that pharmaceutical companies are using. Human bias in inevitable and unfortunately in pharma this has led to a lack of diversity in clinical trials, leading to certain medicines not working as effectively for certain patient populations compared to others.
For Kathy Brunner, CEO of Acumen Analytics, bias presents a bigger to challenge to the market uptake of AI technologies.
“The larger challenge to uptake I would say is algorithm bias which AI has to overcome in the future. Bias can find a way to lurk into algorithms in some ways. AI technology utilises training datasets to make predictions. And these datasets happen to include human-based decisions laced with gender, sex, race, or any other personal information,” Brunner says.
Over the coming years, more pharmaceutical companies will begin to embrace the possibilities that AI offers, and there’s no doubt that more drug candidates discovered through AI will enter clinical trials. But what can we expect going forward?
For Lifebit, AI represents an entire change to how the healthcare industry currently cares for patients.
“Our vision is that when an individual enters the clinic, their genomic sequence, electronic health records (EHR) and diagnostic testing results will be integrated into an AI-powered system that can help diagnose their specific condition. Furthermore, once a diagnosis is complete, the system will also recommend the most suitable personalised treatment for the individual,” Tom Sharrock says.
BenevolentAI sees data as being one of the biggest challenges the industry currently faces regarding AI.
“There is a great deal of publicly available data but there is not enough of the right data for more specific applications – i.e. assay or clinical trial data – so within the drug discovery ecosystem, collaborating with the right data providers and generating the right experimental data is key. Another challenge is that the world’s biomedical data unfortunately is not consistently standardised to common formats, so I hope that another key trend we will see is the consistent adoption of standardised data principles,” says the company’s Mark Davies.
Merck Group expects to see “faster digitalisation and automation in labs.”
In comparison with other industries labs are still manufactories and far behind in terms of automatisation. Moving away from manual tasks towards higher automation always increases efficiency and decrease costs per unit. Labs will close-up, simply because they are very important to us as human beings. The main benefit for researchers will be that they can focus on what they decided to do: doing research instead of administrating and documenting their research,” says Dr Arne Kusserow of Merck Group.
Benchling’s Sanjay Saraf believes that now that the biotech industry has seen the benefits of AI and Machine Learning, organisations will need to begin to incorporate technology into their own operations.
“Currently, access to these AI and Machine Learning tools requires biotechs to partner with other organisations in the technology space to provide these solutions. In the next 10 years, we anticipate a shift in ownership of these technologies, with biotechs investing in developing these technologies themselves to become technology innovators. Others will continue to focus on their biological value proposition, but will need to prioritise the sourcing of AI and machine learning capabilities from a third party,” Saraf says.
Included in the DDW AI eBook, sponsored by Benchling.