One Company Just Did 3.6 Million Experiments in 25 Shots
How AI collapses the pharmaceutical formulation design space from years of empirical screening to weeks of intelligent, targeted experimentation - achieving in 25 experiments what traditionally required millions.
The Combinatorial Challenge
A typical solid oral dosage form has a design space exceeding 3.6 million possible experimental combinations. Three excipients at five concentration levels each, combined with five process parameters at three settings each, yield 3,645,000 permutations. That number has defined the pace of formulation development for decades.
This is a conservative estimate. Most formulations employ more excipients in varying ratios with additional process parameters; the real number often reaches tens of millions. The FDA's Inactive Ingredient Database contains approximately 10,000 entries, yet only about 30 excipients appear in marketed antibody formulations since 2015. The reason is straightforward: screening larger excipient combinations takes more time than drug development timelines allow.
Design of Experiments (DoE) methodologies can reduce the experimental burden to hundreds or thousands of runs. Machine learning goes further. Recent studies show comparable or superior results with as few as 25 experiments, a reduction of several orders of magnitude with direct consequences for R&D cost and speed.
Combinatorial Scale
Exponential growth makes exhaustive screening impractical at any realistic budget. A colon cancer project studying 11 FDA-approved drugs at 10 dosage levels each faces a design space of 100 billion potential combinations.
The Limitations of Traditional Design of Experiments
Design of Experiments has been the pharmaceutical industry's standard approach for managing combinatorial complexity. Full factorial designs, fractional factorials, and Response Surface Methodology all aim to maximize information per experiment. They are the basis of Quality by Design (QbD) initiatives mandated by the FDA, EMA, and PMDA.
DoE has hard limits, though. A full factorial design with seven factors at two to three levels generates thousands of experimental runs, well beyond any realistic timeline. Even fractional factorials struggle above five or six factors. As one industry analysis put it: "Even with fewer factors, 30-50 runs can be resource draining as well as pointless in the sense of gathering useful information about which factors are important."
Pharma was also late to adopt DoE relative to aerospace and automotive manufacturing. The blockbuster era rewarded One Factor At a Time (OFAT) studies over systematic experimental design. Many organizations still carry that legacy in their formulation workflows.
Traditional DoE methods are limited by difficulties in managing high-order complexities and a propensity to become trapped in local optima.
Source: BioPharm International, Characterizing the Formulation Design Space
Traditional Screening Requirements
| Design Type | Factors | Typical Runs |
|---|---|---|
| Full Factorial (2 levels) | 6 | 64 |
| Full Factorial (3 levels) | 4 | 81 |
| Fractional Factorial | 5-7 | 8-16 |
| Central Composite | 4 | 25-30 |
| Box-Behnken | 4 | 25-30 |
| Bayesian Optimization (AI) | 8 | 25 |
Machine Learning in Formulation
Machine learning moves formulation development from statistical optimization to predictive modeling. Deep Neural Networks now exceed 80% accuracy for pharmaceutical formulation prediction, outperforming classical ML models. Gradient boosting methods (XGBoost, LightGBM, CatBoost) produce the most accurate predictions for drug release profiles in comparative studies, with LightGBM consistently ranking highest.
Random Forest algorithms generalize well through ensemble learning; Support Vector Machines handle solubility and dissolution rate prediction effectively on smaller datasets. These methods each fill a role, but the largest efficiency gains have come from Bayesian optimization with active learning.
- Deep Neural Networks: Above 80% accuracy for pharmaceutical formulation prediction, outperforming classical ML models on non-linear relationships between formulation variables.
- Gradient Boosting (XGBoost, LightGBM, CatBoost): Tree-based models producing the most accurate drug release profile predictions. LightGBM ranks highest in pharmaceutical prediction benchmarks.
- Random Forest: Ensemble of decision trees with strong generalization through sample and attribute disturbance. Reliable across a range of pharmaceutical prediction tasks.
Bayesian Optimization in Practice
Bayesian optimization uses active learning to iteratively improve predictions, balancing exploration of uncertain regions with exploitation of promising areas. For pharmaceutical formulation, it consistently delivers the largest reduction in required experiments.
Approximates system behavior and quantifies uncertainty in predictions
Intelligently proposes next experiments based on expected improvement
Achieves global optimization with minimal experimental budget
Applies prior knowledge within and across similar systems
25 Experiments, 8 Variables
A study published in Molecular Pharmaceutics put these methods to a direct test. Researchers optimized thermal stability for tandem single-chain Fv variants across eight independent variables: pH, sodium chloride, l-arginine, l-lysine, l-proline, trehalose, mannitol, and Tween 20.
Traditional DoE approaches would have required 128 experiments for a fractional factorial design, 145 for central composite design, or 113 for Box-Behnken design. Bayesian optimization identified the optimal formulation in just 25 experiments, at least three-fold lower than classical DoE and several orders of magnitude smaller than experimental screening of the entire combinatorial space.
The method efficiently transfers historical information as prior knowledge, enabling simultaneous optimization of multiple biophysical properties. Follow-up work demonstrated concurrent optimization of melting temperature, diffusion interaction parameter, and stability against air-water interfaces, achieving highly optimized conditions in just 33 experiments for a monoclonal antibody formulation.
Industry Results
Exscientia's Centaur AI platform delivered DSP-1181, the first AI-designed drug candidate to enter clinical trials, in under 12 months versus the typical four-year timeline. Their EXS21546 immuno-oncology candidate reached clinical stage in 8 months. EXS4318, a PKC-theta inhibitor, was the 150th molecule synthesized and reached development stage in 11 months.
Active learning strategies select the most informative experiments at each iteration. In one small-molecule discovery program, an initial training set of 150 synthesized analogs produced first-iteration recommendations of 25 compounds, with 80% of predictions falling within acceptable error ranges. Total development required 4 cycles versus the estimated 8-10 for traditional approaches, saving approximately 12 months.
The BO algorithm can effectively converge to global optimization objectives, particularly beneficial when available data and experimental budget are limited.
Source: ACS Molecular Pharmaceutics, Design of Biopharmaceutical Formulations
LNP Optimization for mRNA
AI-driven generation evaluated nearly 20 million ionizable lipids, identifying candidates matching or outperforming MC3 (used in approved vaccines).
LightGBM: R² > 0.87
Phase I Success Rates
AI-developed drugs show higher Phase I success rates than compounds developed through traditional discovery methods.
80-90% vs. 40% Traditional
Cost and Timeline Impact
The median cost to bring a drug to market is $985 million; the average reaches $1.3 billion when factoring in failures. Formulation development accounts for roughly 30% of total R&D costs, with new chemical entity formulation running EUR 60-190 million. Trial delays cost between $600,000 and $8 million per day.
AI-driven approaches are compressing these numbers. Deloitte estimates drug development costs reduced by up to 70%. Insilico Medicine completed development at one-tenth traditional cost. Bristol Myers Squibb reports clinical trial timelines shortened by nearly 2 years. Some implementations show 80% faster trial completion.
The market reflects these results. AI in pharmaceutical markets reached $4.35 billion in 2025, projected to grow to $25.73 billion by 2030 at a 42.68% CAGR. McKinsey estimates generative AI could save pharma $60-110 billion annually. Ninety-five percent of pharmaceutical companies already invest in AI capabilities.
Timeline Comparison
Regulatory Recognition
"AI models such as machine learning, generated using process development data, could be leveraged to more quickly identify optimal processing parameters or scale-up processes, reducing development time and waste."
Source: FDA Discussion Paper on Artificial Intelligence in CDER
API Conservation and Material Efficiency
AI-driven formulation development also cuts API consumption and material waste. Process Mass Intensity in pharmaceutical manufacturing often exceeds 100 kg of input materials per 1 kg of API produced. AI-powered retrosynthesis designs more efficient synthetic pathways, and digital twins allow virtual simulation in place of physical experimentation.
Continuous manufacturing enabled by AI optimization has shown capital expenditure reductions of up to 76%, with overall cost savings of 9% to 40%. The material-sparing angle is worth emphasizing separately: fewer experiments mean less API consumed, a higher probability of hitting target product profiles, and the ability to evaluate formulations that would otherwise require prohibitive quantities of scarce compound.
For early-stage assets where API availability is measured in grams, this changes what is feasible. Formulations that would have demanded kilograms of API for traditional DoE screening can now be optimized with an order of magnitude less material.
Where This Leaves Formulation Teams
The 3.6 million experiment problem is solvable. Bayesian optimization and active learning reduce the design space to a manageable number of experiments, cutting timelines from years to weeks and costs by up to 70%.
By 2030, the AI pharmaceutical market is projected to reach $25.73 billion, with 200+ AI-enabled drug approvals expected. Phase I success rates of 80-90% are becoming standard for AI-developed compounds. The pipeline has grown from 3 AI-assisted drugs in clinical stages in 2016 to over 3,000 currently in development.
The industry is shifting from empirical screening to predictive formulation design. The economics, the regulatory signals, and the clinical data all point in the same direction. Formulation teams that adopt these methods will run fewer experiments, use less API, and reach IND faster.