Fully digitised, AI-directed, in-silico drug discovery and development programs are not just a pipe dream – but what’s standing in their way?
AI in drug development: where are we?
Ever-increasing drug development costs and high attrition rates put a sizeable dent in pharmaceutical companies’ returns on investment. This isn’t just a problem for the biopharma industry and its investors; patients depend on a viable pharma model for innovative and affordable treatments.
Artificial intelligence (AI) has been advocated as a solution. But is it more than hype? Will it be forever on a distant horizon, promising great things for the future? In a 2020 GlobalData survey, AI was identified by the pharma industry as the key disruptive technology in years to come1 – and yet, when it comes to drug development, AI has still not delivered on its promise in any tangible way.
The sheer breadth of possibilities for AI usage is huge. And yet examples of the full realisation of AI aiding the multi-step, drug-to-patient process – by, for example, fostering innovation, streamlining processes, cutting costs and enhancing probabilities of success – are only now beginning to emerge. Certainly, pharma has lagged behind other industries – such as the financial sector, but also the more adjacent healthcare industry – in its adoption of AI applications. And yet a flurry of activity on multiple levels hints at better things to come:
- Big pharma players such as Bayer, Pfizer, and AstraZeneca striking deals with AI biotechs2
- Companies such as GSK, creating internal AI departments3
- AI biotechs hauling in almost $2bn in venture capital in 2020 – including a few hefty late rounds, such as XtalPi’s $319m Series C and Recursion Pharmaceuticals’ $239m Series D2
- Tech giants such as IBM, Apple and Google investing in the pharma and healthcare sector
- Pharmaceuticals companies and academic institutes engaging in private-public partnerships, including drug discovery alliances such as Pharmaceutical Discovery and Synthesis (MLPDS)4
The COVID-19 pandemic has accelerated digitisation across the board, and the development and adoption of AI approaches in the biopharma industry is no exception. Although disruption to supply chains and key processes, such as clinical trials5, has hampered the industry, it has proved increasingly attractive to investors, recording a massive influx of VC capital. And it’s not just the usual biotech investors getting involved; non-biotech internet tech investors such as Andreesen Horowitz (a16z) are also trying to carve out a stake. All this creates a fertile breeding ground for innovation in the AI space.
A vision of AI-enabled drug development
How long will it take to realise AI’s potential for applications in the biopharma industry? And how much benefit will it bring in terms of time, cost savings and innovation?
On a macro-level, it is only a question of when – and not if – biology will become computable (or predictable) on a practically relevant scale. While our understanding of biology has moved at a considerable pace during the past decades, aided by the decoding of the human genome and novel methods such as CRISPR/ CAS9, AI will help us harness the full, predictive power of those discoveries for understanding and treating human diseases. But when?
AI’s significant impact in the field is within sight. Don’t expect one definitive Eureka moment, though – AI’s influence will grow gradually as a result of algorithms making decisions better and/or more efficient across a vast array of highly specific niches along the drug development pipeline. Structurally, its effect will be similar to the influence computers have had (and AI is beginning to have) in other industries, such as accounting or logistics.
In accounting, where information processing is absolutely key, the introduction of computers led to a dramatic change, with whole swathes of jobs/departments becoming suddenly obsolete. Logistics is at the other end of the spectrum: the essence of the business – that stuff needs to get from A to B – remains a manual task. And so computers took hold slowly, in niche parts of the business rather than at its core; nevertheless, the industry has been changed.
Although drug development’s heavy reliance on information processing hints at an overnight, accounting-style transformation, the nature of the information being processed is too complex for that. Change will likely come slower – more akin to logistics – but with a sudden dramatic impact once critical mass is reached.
For now, hunches and intuition still play a big role in biotech decision making – and the culture around storing, mining and trusting data has remained weirdly artisanal. Short-term gains will likely be found in those two areas, from seemingly simple database matching or NLP-based mining of research to augment drug discovery. The big prize – predicting biology – will be the culmination of many relatively simple developments.
And what might that fully AI-driven future model of drug development look like?
AI will innovate, digitise and rationalise target identification and drug discovery
The discovery of new drug targets and lead molecules will rely on sophisticated models that predict the structure and properties of drug targets, small molecules and biologics, by mining the accumulated scientific data that is available in standardised databases. Based on this information, drugs will either be designed de novo, using deep learning methods, or selected from huge libraries of chemical and biological entities, shared in public-private consortia of drug developers and academic institutions. The process of identifying new hits in this scenario will take no longer than a few months, rather than the five-plus years spent using traditional methods6.
Predictive models based on structural, interaction and expression data, as well as published drug profiles, can determine a compound’s properties, such as potential on-and off target toxicities and pharmacological profiles. In silico experiments will become ever more important, reducing – or completely eradicating – the need for years of wet-lab research and sparing animal lives. Moreover, AI-driven repurposing efforts will pinpoint new applications for already approved drugs. We saw a glimpse of this recently, when Eli Lilly’s rheumatoid arthritis drug, baricitinib, was approved for COVID-19. Given the already accumulated data, additional testing for repurposed drugs will be minimal; their path to market will be swift and low-risk.
As well as saving time and costs, AI will foster innovation and improve treatments for patients. Predictive models exploiting multiple data sources, as well as sophisticated screening and design methods, will identify novel targets and drugs addressing those targets: currently untreatable diseases will become treatable; therapies will become more personalised to the individual. Moreover, increasing the quality of molecules that enter clinical development will subsequently increase the probability of clinical success.
AI will modernise clinical trials, improving recruitment, adherence and probabilities of success
In the clinic, AI-designed trial protocols will use methods such as natural language processing (NLP) to mine the available data from previous trials and establish key determinants of trial success, such as exclusion criteria, endpoints and sample sizes. Patient identification will employ sophisticated recruitment algorithms, drawing on genetic screening, electronic health records and other sources to match patients to suitable trials.
Placebos will be a thing of the past, and control arms will be fully virtual, based on historical controls or digital twin approaches; this will slash recruitment time and halve trial sizes. With less to fear from receiving inactive (or less active) comparators, patients will be much more willing to partake in clinical trials. Some types of trials (for example for well-established modes of action and diseases contexts or formulation changes) will, conceivably, be fully digital.
For those trials that still require patients’ active participation, disruption will be minimal. Most will be decentralised, with digital biomarkers measuring study outcomes and adherence. This will mean less travelling to study sites for patients, and more highly valuable real-world data for the trial, enriching assessment of the benefits and challenges of a given drug.
Logistically, sophisticated algorithms will aid the supply chain management for trials and post-marketing demands, further speeding up trials and reducing costs.
Together, these measures will make trials more efficient, saving years in recruitment time and preventing failures due to insufficient enrolment or patient adherence. Costly protocol amendments could be eradicated too. Expect overall clinical development costs to be halved, minimum, as a result.
At the same time, trials will be more ethical, sparing patients from inactive or harmful treatments, while further tailoring the treatment to each patient’s specific needs (based on genetic backgrounds, disease severity, concomitant medication and many more factors).
The end of the beginning: where do we really stand with AI-based drug development?
We are only at the very beginning of the curve towards AI-driven drug development. To gauge how steep its slope will be in the years to come, let’s first take stock of where we stand in AI-based drug development today.
While many biotechs, pharma and tech companies claim AI breakthroughs, those applications have mostly been tried in a couple of smaller studies or within a narrow context. This does not necessarily provide proof of concept, especially given the fact that compatibility and comparability between different methods and vendors is challenging.
We currently have:
- drugs whose discovery was aided by AI now entering clinical development7,8
- a drug whose repurposing was projected by AI on the market9,10
- and AI-aided clinical trials being implemented11,12
These are encouraging signs – and yet you could argue that simply succeeding in their given context is not enough. It will take evidence AI continuously saves time and money, or increases innovation, compared with regular methods, to prove it provides a genuine benefit.
The question remains: what does a proof of concept for AI-aided drug development really look like? Does the question even make sense? In drug development, especially, AI will refer not to a single technology, but a large number of tools, data infrastructures and cultural changes. A few general purpose end-to-end approaches may, one day, emerge from this zoo, but this is a distant dream at present.
It’s important to point out that AI applications don’t exist in a vacuum; they flourish or flounder due to the availability and quality of input data. Access to high-quality data sources, as well as legal and ethical issues around AI data usage, are sure to challenge drug developers in the years to come. And as more and more companies get involved in the collection and mining of fine-grained patient data, questions of privacy and data ownership will become ever more relevant and widespread.
Although many data sources are freely available, others are proprietary. Competition between both data creators and data users can result in a silo culture – a problem that’s only compounded by the fact that biomedical data is often redundant and non-standardised, making the use of NLP technologies challenging. Both data standardisation – of electronic health records, for instance – and data sharing between competing organisations will be key for the success of AI in drug development. This may even be enforced by regulators (or politics in general), prompting new paradigms similar to patent protection.
If developers continue to address those challenges, AI might not merely improve drug discovery and development; it might utterly reshape it into a fully digitised process. It sometimes seems that algorithmic applications are too readily branded “transformative”, but when AI eventually learns to compute and predict biology, its impact on the pharmaceuticals industry – and healthcare as a whole – will truly justify the tag.
Imagery shows quantum computers, which will transform analysis and integration of enormous datasets, vastly enhancing our machine learning and artificial intelligence capabilities.
1. Global Data industry survey published in Nov 2020, highlighting the technologies that pharma representatives consider disruptive for the future
2. Deep Pharma Intelligence report on the AI biopharma landscape, including deal and financing activities, key players and technology breakthroughs.
3. GSK invests in internal AI efforts, planning on hiring 80 AI specialists by the end of 2020.
4. The Machine Learning for Pharmaceutical Discovery and Synthesis Consortium, an MIT-driven drug discovery alliance that includes 15 midsize and big pharma players.
5. Article in The Lancet discussing the impact of COVID-19 on the clinical trial landscape.
6. Deloitte report on AI in drug discovery, 2019.
7. Article discussing the collaboration between Benevolent AI and AstraZeneca, including the first clinical entry for kidney disease derived from their partnership
8. Article discussing the development of DSP-1181, the first AI discovery-aided drug to enter clinical development
10. Article describing the combination action of baricitinib and remdesivir in COVID-19 patients.
11. Article highlighting how AI applications might boost success of clinical trials, including quotes from clinical trial specialists and AI researchers.
12. GlobalData industry survey on decentralized trials.