Some Assembly Required: Building Pharma’s Digital Future

As the pharmaceutical industry faces a fresh round of digital transformation, the promise of technologies like predictive artificial intelligence (AI) is no longer just an exciting hypothesis. From drug discovery to clinical trials and patient engagement, the integration of digital tools could shrink development timelines and fundamentally reshape how life-saving treatments are designed and brought to market.¹^,²

As with any revolution, however, success depends both on what new tools can do and how responsibly and effectively they are implemented. For AI in particular, its value is inseparable from the quality of the data it draws from. Fragmented historical records, inconsistent standards, and siloed systems continue to hinder progress, while pressing ethical, technical, and operational questions persist surrounding data stewardship, algorithmic transparency, and patient privacy.³^,⁴

In this article, we will explore some of the foundational steps the industry can take to harness the full potential of predictive AI. Plus, I explain how modernizing legacy data into robust information ecosystems can help build a future where AI innovation serves not just efficiency, but real improvements in patient care.

Making the past work for the present 

Much of the data researchers still rely on today was gathered from studies conducted years, if not decades, ago. Just as early adopters of digital technology had to contend with the limitations of physical records, those hoping to harness the power of AI are finding that most historical research was not organized with future  modeling in mind. The data in question can be scattered across incompatible sources like spreadsheets, equipment memory files and presentations, or be stored in dense, unstructured libraries that make easy access virtually impossible. 

Grappling with the issues inherent to historical research can be daunting, but with time, patience and the right processes, they can be overcome. A good place to start is by collecting any available data from all available sources, regardless of its perceived quality. Once gathered, all this data must be screened for completeness, with any missing values filled, and incomplete or inaccurate points discarded, to guard against future inaccuracies. There is, however, always a significant risk of obtaining a very partial dataset from historical data. In one example, researchers at Roquette began with a collection of raw data drawn from 20,000 separate trials. Upon inspection, they determined that just 5,400 featured an associated study report suitable for mining, and of this group, only 600 were fully complete. This demonstrates just how important proper data checks are to the construction of accurate digital tools, because from the original cohort of 20,000 trials, only 3% proved viable. It also shows just how important it is to capture data when it is produced, rather than after the fact, to guarantee completeness. 

Stepping up study designs for the digital age

It’s not only historical study data that deserves drug developers’ attention, but also the thousands of points of experimental information produced every day in R&D and pharma laboratories around the world. The objective here is to build robust data management practices into the process from the start, ensuring data is as accurate and organized as possible to fully unlock its value later down the line.  

Establishing good data management practices is—in theory—simple. First, the experiments should be set up using a dedicated tool, such as an Electronic Laboratory Notebook (ELN), rather than relying solely on in-built equipment software. This record should contextualize the project, summarize important experimental conditions, like the formulas or equipment used, and generate a unique study identification (ID) code to ensure the traceability of the data in any future applications. 

The next step is for technicians to perform the research, making certain to measure all critical attributes and parameters to establish good data coverage. Running in parallel with the experimental step is the equally vital process of data recording, during which technicians should capture information from connected equipment and implement simple, yet comprehensive manual recording measures for any nonintegrated apparatus. Though it may seem counterintuitive, it is essential to record failed assays to define the boundaries between effective and ineffective formulas. After all, AI models cannot predict what they have not had the opportunity to learn, making it doubly important to provide them with records of what to avoid. 

All in the preparation

Generation and recording done—extraction, standardization and centralization still to come. Data preparation steps like these are crucial to the success of any future AI model, but they also require a set of skills not typically found in pharmaceutical science organizations. As such, technicians may require additional training or the support of external partners to fully reconcile differences in data architecture between equipment systems and automate the centralization of information as it is created.

Once they have a comprehensive bank of accurate, homogenized data at their disposal, drug developers can finally move on to its use. Even if the goal is to move straight to the creation of advanced AI models, there are many benefits to be found in visualizing gathered data into readily accessible graphs and study reports. Easily digestible resources like these can help technicians share knowledge and avoid duplicating experiments—saving time, effort and valuable resources (Figure 1). 

The main event: AI modeling 

Now we come to the crux of this careful preparation—AI-enabled drug development tools, harnessing high quality historical data to make accurate predictions that have the power to change and save lives. Specialized AI models capable of analyzing vast quantities of data points almost instantaneously are already beginning to make their mark on drug development.⁵ The rapid and comprehensive assessment offered by these tools frees pharma technicians to concentrate on the more intricate, qualitative analysis currently only achievable with human expertise and experience. This combination of organic intuition with digital efficiency has the potential to fast-track the discovery of new active ingredients and excipients, as well as their development into fully approved drugs.⁶

A compelling illustration of this is Pfizer’s application of AI to the formulation of its COVID-19 treatment, PAXLOVID™ (nirmatrelvir tablets; ritonavir tablets). Here, AI tools were leveraged to identify the optimal crystal structure that would allow the drug to be developed as an oral medication, rather than a traditional vaccine administered via injection.⁷ The adaption of treatments to more convenient delivery formats is not the only patient-centric advancement offered by AI though. By shortening timelines for drug discovery and approval, these models can hasten the development of treatments for rare and chronic conditions, in turn allowing healthcare providers to arrest diseases sooner in their progression for improved prognoses and faster symptom relief.

The possibilities unlocked by AI are virtually limitless—as expansive as the available data and human imagination will allow. Focusing on drug formulation in specific, however, what could an artificially intelligent platform combining mechanistic models and machine learning for accurate predictions look like in practice? 

Practical theory: Digital formulation assistants 

Such a model could be designed in two directions: either as a “formulation optimizer” that predicts the properties and performance of the final drug based on provided excipient, API and process parameters, or as a “formulation assistant” that recommends excipient selection and process parameters, based on the stated end goal of the formulation and API characteristics. Grouped together as “digital formulation assistants” (DFAs), the impact of tools like this on drug design workflows could be transformative. 

Beyond their obvious efficiency-enhancing capabilities, DFAs offer an array of powerful advantages to drug developers. Their ability to speed up decision-making while improving accuracy and consistency significantly streamlines production scale-up, and simplifies regulatory approvals with comprehensive, traceable documentation already generated in tandem with a drug’s development. Ultimately, these benefits add up to optimized resource use, reduced expenditure on experimental trials, and a generally enhanced spirit of competitive innovation—all in service of better care for patients.

A note of caution: Integrating ethics from the ground up

The advancements offered by DFAs, and AI more generally, will have an undeniably positive impact on millions of patients—providing they are carefully designed and meticulously monitored. The use of AI in the context of healthcare is already a contentious topic, with several nations implementing regulations mandating that patient information be stored exclusively on servers within their own borders to prevent sensitive data being incorporated into global AI datasets.⁸

In all their data gathering, management and application activities, therefore, pharmaceutical companies must make security, accuracy and accountability core priorities. 

Constructing commitment

Whether it is old, incomplete or scattered, the swathes of scientific data at drug developers’ disposal are an underexploited goldmine. The technology to harness this treasure trove is already available, and yet many organizations remain reticent to take the first step. Ultimately, unlocking the full potential of pharma development data demands more than just technology; it requires a collective change-mindset. Every individual involved in the process must be bought into the project of data gathering and management, with no shortcuts or deviations. Without this shared commitment, gaps and errors will inevitably find their way into future models, compromising their integrity. The takeaway here is that digital evolution is a marathon that requires a strategic, step-by-step approach, all in service of an ambitious end goal. By embracing this challenge with a unified vision, pharmaceutical brands and the industry as a whole can finally turn its wealth of data into a powerful engine for progress, built for success right from the start.

References

1. Topol, E. (2019). Deep Medicine: How Artificial Intelligence Can Make Healthcare Human Again. Basic Books.
2. Esteva, A., Robicquet, A., Ramsundar, B., et al. (2019). A guide to deep learning in healthcare. Nature Medicine, 25(1), 24–29
3. Price, W. N., & Cohen, I. G. (2019). Privacy in the age of medical big data. Nature Medicine, 25(1), 37–43.
4. Rajkomar, A., Dean, J., & Kohane, I. (2018). Machine learning in medicine. New England Journal of Medicine, 380(14), 1347–1358.
5. https://www.mckinsey.com/industries/life-sciences/our-insights/generative-ai-in-the-pharmaceutical-industry-moving-from-hype-to-reality 
6. https://www.mckinsey.com/industries/life-sciences/our-insights/generative-ai-in-the-pharmaceutical-industry-moving-from-hype-to-reality
7. https://www.pfizer.com/sites/default/files/investors/financial_reports/annual_reports/2022/story/data-and-ai-are-helping-to-get-medicines-to-patients-faster/
8. https://www.mckinsey.com/industries/life-sciences/our-insights/generative-ai-in-the-pharmaceutical-industry-moving-from-hype-to-reality

Nicolas Descamps is Global Technology Liaison Manager for Pharma at Roquette Health & Pharma Solutions, a provider of excipients and biopharma ingredients that support formulation innovation, drug delivery, and patient-centric solutions worldwide.