Pharma R&D is transitioning toward an AI/data driven model of drug discovery, and great progress has been made in the last 20 years. To fully exploit the potential of this technology we must think holistically about how it is integrated into R&D. Fit-for-purpose data is essential, as are models that have causality at their heart, says Dr Ben Sidders, Chief Scientific Officer at Biorelate.

Pharma R&D has undergone a number of transitions and is now moving towards a data-driven, AI enabled mode of drug discovery. This is evidenced by the rise of AI first companies such as BenevolentAI, Recursion and Insilico Medicine. Traditional pharmaceutical companies are also making great strides to incorporate AI across their businesses from R&D to sales and manufacturing. 

Overall, the perceived value of AI in drug discovery and development has not matched the hype, however. This is a consequence of several factors, most notably the over-inflated expectation that AI would transform all of R&D in one big bang, simultaneously reducing timelines and costs while also boosting success rates. But over the last five years we have seen multiple successful point solutions that are tangibly transforming specific steps within the R&D pipeline.

Signs of progress

In chemistry, AI models like ChemBERTa, which are based on transformer architectures similar to those used in natural language processing, can efficiently predict small molecule properties. By leveraging these models, researchers can sift through vast chemical libraries to identify promising compounds without needing to perform experimental assays on each one. Additionally, retrosynthetic route planning using tools like AiZynth massively accelerate the otherwise manual process of identifying the building blocks for small molecule drug synthesis. Such tools automate key processes in small molecule drug discovery, significantly reducing costs and accelerating R&D programs. In diagnostics, AI models for computational pathology have made significant strides. Foundation models like UNI1 are pre-trained on vast datasets and fine-tuned for specific tasks, and have advanced how pathologists analyse complex medical images. In fields like cancer diagnosis, AI-powered models are now able to assist in identifying disease markers, assessing tumour progression, and even predicting patient outcomes with greater speed and accuracy. For instance, AI systems can process tissue samples and identify key biomarkers such as PD-L1, which is essential for tailoring immunotherapy treatments.

In target discovery, knowledge graphs (a collection of interlinked descriptions of entities that put data into context and enable data analytics and sharing) are adept at integrating a vast number of data sources into a query-able structure and can be used to make informed, relatively unbiased target prioritisation. A graph containing 84 million relationships derived from 37 separate sources has been successfully applied to the problem of target triage, identifying the most promising targets from hundreds of hits arising from genome-wide functional genomic CRISPR screens using a multiparameter optimisation approach2. The time taken to identify hits for validation was reduced from months to days. In addition, these approaches enable better decisions, leveraging more data than any one scientist could feasibly maintain a working knowledge of.

There remain many problems within the R&D lifecycle that are not yet solvable by AI solutions, however. For example, predicting synergistic drug combinations has been the topic of extensive research, and every flavour of AI model has been assessed with only limited success and almost no translational relevance. Nor are we any nearer to being able to predict the effect of a drug on a given patient without first running a clinical trial to gather data on who does and doesn’t respond to that drug. 

What then have we learnt from recent developments, and how might we better realise the value of AI in R&D? 

It is helpful to think through four cornerstones to transformation: data, model, culture and validation. Below we explore each of those elements in turn.

Data

In general, the areas where AI is finding success are those where the data set is large, complete and in many cases has been generated specifically to solve the problem at hand. For example, the UNI foundation model for computational pathology was trained on >100 million images from 20 tissue types. In contrast one of the largest datasets available to train models for drug combination synergy prediction has 910 combinations of 118 drugs – many orders of magnitude smaller. 

This problem is further exacerbated when we look at data from clinical trial cohorts. Here the data is often sparse, and inconsistent in what is measured. For example, one trial might collect demographics and data for a specific blood-based biomarker; another might also collect genomic data. Further, while two cohorts might both report genomic profiles one might be genome-wide and the other a targeted panel; or one might employ a modern technology with greater sensitivity and the other an older technology with limited power to detect genomic variants. 

Then there are differences in the analysis pipelines applied to all these data. Re-processing and harmonising all of these data types is highly labour intensive, and often only the start of the process. The features used to train models may not be derived directly from the harmonised data and may need significant further manipulation (for example, it is common to look at the expression of level of genes at the pathway or functional level, but not at the level of the individual gene).

The most pressing underlying issue, however, is that pharma’s data, particularly that from clinical trials was not generated for AI; rather the data is a byproduct of business as usual. To exploit data in a meaningful way using AI, companies must develop a data strategy – and possess a willingness to fund and generate data on clinical cohorts if possible – with a view to building valuable data of the scale needed for the intended purpose.

The next requirement is to be able to extrapolate from that data, in a meaningful and actionable way.

Model

AI models are fantastic at classification and predictive problems – for example identifying whether an histology slide is from a diseased or normal tissue sample. For AI to truly revolutionise drug discovery, though, the technology must move beyond pattern recognition and predictive analytics to include causality, which is essential for technical and regulatory success in drug discovery. For example, predicting that a drug might work in a new indication is valuable but it is not the same as explaining why the drug will work in that indication. In order to support internal and regulatory decision-making it is essential to have explainable biology that supports a mechanistic understanding of the drug or biology at hand.

A promising approach is emerging now to fill that gap. It calls for a balanced strategy that integrates both prior knowledge and data-driven insights to provide a more robust framework suitable to the needs of drug discovery. In this approach, AI combined with highly accurate cause-and-effect relationships can distil not only a broader array of targets with strong promise, but also a supporting narrative: a mechanistic understanding of the biological processes at work and detailed causal explanations of their role in disease. 

Cause-and-effect relationships can be mined from the literature and created from experimental data. These relationships define the regulatory interactions between two biological entities, be they genes, proteins or cells. These can be combined into structural causal models, a framework to represent and analyse the causal relationships between variables. Such models provide a systematic way to model how changes in one variable can lead to changes in another. These can be used during the training process of more expansive foundation models but also to build specific mechanistic models that further describe the output from an upstream finding. Such actionable insights are essential to guide further experimentation and discovery of drugs with plausible, explainable mechanisms.

Validation

The output from all AI solutions should be validated, experimentally if appropriate. In discovery, and target discovery in particular, an AI-driven R&D system is fundamentally different to the traditional approach. 

Firstly, the R&D function should be set up so that all data feeds back to the AI model. This helps to mitigate some of the challenges laid out above but also ensures that the model can be continually improved. This is synonymous with the continuous integration and continuous deployment practice employed in modern software development, but is often referred to as the “lab in the loop”. For example, every result from the CRISPR screening group should find its way back into the knowledge graph so that future queries can benefit from that data.

Secondly, unlike scientists that utilise their expertise to distil their knowledge down to one target of interest, an AI system can identify hundreds of targets. It is tempting for discovery teams to cherry pick from these the targets they are familiar with, which eventually leads to the ‘we didn’t need the AI model’ circle of flawed logic. Data scientists operating in this environment therefore optimise their models toward ‘goldilocks’ targets. 

Goldilocks targets are novel in the sense that there is no launched drug or significant competitor activity but are not truly novel in that there exists information or knowledge external to the modelling environment that associates the target with a disease in some capacity. Goldilocks targets excite drug discovery teams because they don’t have to fully trust the AI model and the prior knowledge can suggest approaches for experimental validation. However, this fundamentally negates the potential of AI. To prevent this a validation model that represents a triage system is needed. Orthogonal in silico approaches might be used to go from 1,000 to 100 targets, but to go from 100 to 10 the team should adopt the quickest, most high-throughput experiment that will provide the next rung of supporting evidence regardless of other prior evidence.

Culture

Underlying many of the data, model and validation issues is the culture of the organisation and its failure to fully adapt to an AI-based way of thinking or working.  Many pharmaceutical companies have long-established processes and workflows that rely on traditional methods. AI adoption requires a fundamental shift in thinking, and there is often hesitation among scientists and decision-makers to trust AI-driven approaches over conventional practices. This is driven by a prevalent concern about the reliability and interpretability of AI models. This lack of transparency in AI algorithms, often called the “black box” problem, raises doubts about their scientific validity (see above, under ‘model’). Pharmaceutical companies tend to have their data spread across multiple teams, divisions, and even external partners. These data silos prevent seamless integration and hinder and contribute to the data issues already described.

Many researchers and professionals in pharma R&D may lack the expertise to work with AI tools effectively. While there are increasing efforts to bridge this gap, upskilling or recruiting talent with AI expertise is essential. However, it goes both ways – data scientists must be educated in the decision-making process of R&D and understand/develop methods that directly support that. There are also fears over job retention and the narrative needs to be developed that AI can raise the productivity level of all R&D researchers, allowing the same number of staff to achieve more.

Taking action

Projecting 5-10 years into the future, we can expect that every major decision taken along the drug R&D pipeline is accelerated by unparalleled access to knowledge – knowledge that has matured from purely associative, predictive assertions to deep causal linkages between targets and disease; biomarkers and drugs; and more. 

To get to that point, however, the data science community must consider not only the specific problem they are solving in the here and now, but also the context in which they are doing it.

Progress will depend on each set of stakeholders playing their part, both individually and in collaboration with one another. Data scientists, for instance, must develop actionable models with causality at their heart. Biologists must determine how to effectively integrate data science into their workflows. Heads of R&D, meanwhile, must consider how best to affect the cultural shift now needed, to promote more seamless integration and symbiosis between the two sciences.

The rewards will be worth it. Ultimately, this is an opportunity to transform the success of drug discovery programs and avoid the high costs associated with unnecessary experimentation and/or misdirected decision-making.

From DDW Volume 26 – Issue 1, Winter 2024/25 Read the digital issue

References:

  1. https://www.nature.com/articles/s41591-024-02857-3
  2. https://www.nature.com/articles/s41467-022-29292-7

About the author:

Ben SiddersDr Ben Sidders, Chief Scientific Officer at Biorelate, has been working at the forefront of pharma data science for the last two decades. Formerly Executive Director and Head of Early Data Science within Oncology R&D at AstraZeneca, responsible for advancing the company’s understanding of disease through computational oncology, Sidders also previously spent eight years at Pfizer, and has extensive experience of many aspects of drug discovery for major pharma.