From Trial and Error to Predictive Design: How AI Is Transforming Formulation Development

Introduction

The drug development industry has long accepted high attrition and extended timelines as an unavoidable cost of doing business. Although formulation science is crucial in ensuring exposure to deliver efficacy and in achieving patient acceptable dosing regimes, it remains a frequently underappreciated contributor to successful drug development. Candidate molecules fail not only because of fundamental biological shortcomings, but often because the formulation surrounding them was never quite right or wasn’t optimised early enough to make a difference.

Artificial intelligence is beginning to shift formulation development from an empirical discipline to a predictive one. Machine learning models integrated with formulation development, manufacturing and clinical research units are moving decision-making upstream, compressing timelines and enabling programmes to reach the clinic faster. The evidence for this shift is no longer confined to conference presentations and white papers: In May 2026, an AI-designed novel oral solid dose formulation entered Phase I clinical testing at Quotient Sciences, following approval by the Medicines and Healthcare products Regulatory Agency (MHRA). That milestone reflects both the maturity of the underlying technology and the readiness of the regulatory environment to accommodate it.

The Formulation Bottleneck

The gap between a promising molecule and a viable drug product that delivers is where programmes quietly stall. Formulation development sits at the intersection of chemistry, materials science, and manufacturing engineering, requiring a balance between bioavailability, stability, processability, and patient acceptability. That complexity does not lend itself easily to generalisation, and yet for decades the dominant approach has been precisely that: applying prior “formulation platform” experience from different drugs to new candidates and iterating experimentally until something works.

Andy Lewis, Chief Scientific Officer at Quotient Sciences, describes the structural problem: “Traditional trial-and-error formulation development is slow, material-intensive, and expensive. It is also frequently informed by prior experience with molecules with dissimilar physicochemical and biopharmaceutic properties, introducing assumptions that may not hold.”

Additionally, each formulation iteration consumes drug substance that is often scarce at early development stages. This can extend the timeline to advance the drug to first-in-human and/or proof of concept studies, while providing only limited mechanistic insight into why a particular formulation performs as it does.

The downstream consequences are significant: Suboptimal formulations entering clinical trials can mask a molecule’s true therapeutic potential, generating equivocal efficacy data that lead to incorrect programme decisions. In some cases, promising candidates are discontinued not because they lack biological activity, but because the formulation failed to deliver the drug to the right place, at the right rate, in sufficient quantity. The cost of that outcome, measured in time, capital, and patient impact, is substantial.

What AI Changes

The machine learning models with greatest promise for AI-guided formulation development are not general-purpose AI tools repurposed for pharmaceutical science. They are purpose-built models directed by humans, designed to learn the relationships between formulation variables, processing conditions, and measurable outputs, whether that is dissolution profile, physical stability, or in vivo performance.

The practical effect is a compression of the experimental cycle. Where traditional development might require multiple design-of-experiment iterations to identify a viable formulation space, an AI-guided approach uses the model to predict which combinations of excipients and process parameters are likely to deliver target performance, prioritising experiments that generate the most informative data rather than those that simply follow convention. This reduces the total number of iterations required, which in turn reduces drug substance consumption and shortens the time to a development decision. Across programmes, this approach has the potential to deliver estimated time savings of 30 to 50 per cent from formulation development through to clinical testing.

Critically, the model outputs are not treated as definitive answers. The architecture is human-in-the-loop: scientists interrogate model predictions, apply domain knowledge to assess plausibility, and design experimental confirmation studies accordingly. This is not AI replacing scientific judgement. It is AI making that judgement more efficient and better-informed. The models surface patterns across large and complex datasets that human experts may not readily detect, and they do so consistently, without the variability introduced by changing project teams or institutional knowledge gaps.

Digital twin outputs and in silico simulation extend this capability further. By generating a virtual representation of a formulation’s expected behaviour under defined conditions, scientists can test scenarios computationally before committing to physical experiments. This is particularly valuable in early development, where the cost of a failed experiment is highest and the available drug substance is most constrained. Identifying non-viable formulation strategies in silico, rather than at the bench, changes the economics of early-phase development in a meaningful way.

Putting It Into Practice

The value of AI-guided formulation development depends substantially on how it is integrated into a broader development workflow. A predictive model that exists in isolation from manufacturing and clinical capabilities can improve experimental efficiency but cannot, by itself, accelerate a development programme to the clinic. The connection to what happens next, including prototype manufacture, clinical supply preparation, and first-in-human dosing, is where the speed and confidence gains are fully realised.

For nearly two decades, Quotient Sciences have integrated more informed data and decision making processes as part of its integrated Translational Pharmaceutics® programs, combining formulation development, manufacturing, and adaptive Phase I clinical studies within a single operational unit. Recently, the company has embedded AI-guided formulation development within its Translational Pharmaceutics® platform, allowing a formulation selected based on AI model predictions to be manufactured and administered to human subjects within the same programme. Clinical data feeds back into the model, informing subsequent formulation decisions in near real time and creating a closed loop between computational prediction and clinical outcome.

This model distinguishes AI-guided development from predictive modelling in isolation and for clients, changes how early development decisions are made. Rather than committing to a single formulation candidate and accepting the risk that it may not perform in the clinic, programmes can be designed to test formulation variables alongside pharmacokinetic endpoints in healthy or patient participants. The AI model aids the optimisation process ahead of clinical testing; clinical data validates and refines the model’s predictions. The outcome is a more iterative, data-driven process that is nonetheless faster than conventional sequential development, with better-characterised drug products reaching the clinic sooner and with greater confidence.

Clinical Validation

In May 2026, an AI-designed novel oral solid dose formulation entered Phase I clinical testing at Quotient Sciences following MHRA approval. The formulation was developed in collaboration with Intrepid Labs, a company focused on developing proprietary AI tools for drug product development. This milestone represents, to our knowledge, the first instance of a formulation designed through an AI-guided approach progressing into clinical evaluation.

Regulatory acceptance of an AI-designed formulation in a clinical setting reflects a growing maturity in the evidence base supporting these methods and an increasing willingness on the part of regulators to engage with AI-informed development packages when presented with appropriate scientific rigour and documentation. It establishes a practical precedent: that AI-guided formulation development is not a theoretical proposition but a methodology capable of delivering regulatory-ready drug products and delivering them faster.

What This Means for the Industry

The shift underway in formulation science is not primarily a technological story; it is a strategic one. Pharmaceutical companies that treat AI as a formulation tool from the earliest stages of development, rather than as a late-stage rescue strategy or a supplementary analytical method, will accumulate advantages that compound over time. Faster progression to the clinic, reduced drug substance consumption, fewer experimental iterations, and better-characterised formulations are individually meaningful. In combination, they alter the economics of early-phase development in ways that affect portfolio decisions as well as individual programme outcomes.

The integration of AI with clinical infrastructure is what makes this commercially meaningful. Predictive models are valuable; predictive models that connect formulation composition to in-vitro performance, with the ability to incorporate clinical feedback in near real time, are transformative. The May 2026 milestone demonstrates that both conditions can be met: an AI-designed formulation, developed through a rigorous human-in-the-loop process, progressing to Phase I at the site where it was developed, on an accelerated timeline that would not have been achievable through conventional methods.

Organisations that engage with AI-guided formulation development early, and pair it with the clinical and manufacturing infrastructure needed to act on its outputs, will move faster and fail less. The discipline now is to apply that capability consistently and from the outset, across programmes where the opportunity to change outcomes is greatest.