Application of Predictive Modeling in the Development of Amorphous Solid Dispersions

Formulators challenged by promising investigational drug candidates with high bioactivity but limited solubility increasingly find amorphous solid dispersions (ASDs) an attractive solution for enhancing bioavailability. The FDA approved nearly 50 drug products formulated as ASDs between 2012 and 2023.1 However, identifying the optimum method of ASD preparation (e.g., spray-drying [SD] or hot-melt extrusion [HME]), and choosing the polymeric carrier, surfactant, and drug loading level can be challenging.

Traditionally, extensive physical experimentation using a trial-and-error approach has been required, which is both costly and time-consuming. As mechanistic understanding of ASD stability and solubility behavior and computational power have expanded, predictive modeling and simulation using quantum mechanics (QM), molecular mechanic (MM), molecular dynamic (MD), and physiologically-based pharmacokinetic (PBPK) approaches have helped accelerate formulation development by providing information on thermodynamic properties and potential drug-excipient interactions and miscibility, among other attributes that impact ASD performance during storage and in vivo.2-6

Predictive Modeling in ASD Formulation Development

For many poorly soluble molecules, dozens of potential ASD formulations are possible. Ongoing research is leading to the development of new ASD forms. Similarly, the choices of polymer carriers, surfactants, and other excipients suitable for generating optimal ASDs are continually expanding. Physical screening of possible formulations consumes expensive API in addition to consuming time and human resources.

Predictive tools help narrow the field from dozens of formulations to a targeted set of 2 to 6, reducing API consumption, effort, and time in the experimentation process, according to Josh Marsh, BAE and PBBM lead scientist with Lonza. “By reducing experimental burden early in development, they support more efficient process development and scale-up activities,” he observes.

Wider Use of Predictive Modeling and Simulation in ASD Formulation

Several trends have contributed to greater use of predictive modeling and simulation approaches in ASD formulation development. Advances in workflow automation and scalability have made broader simulation-based screening practical, including evaluation of excipients in the context of the intended manufacturing route, according to Shiva Sekharan, global portfolio leader, formulations and CSP software with Schrödinger. “Customized multistage MD workflows and advances in machine learning help better leverage historical experimental data, allowing prioritization of formulations and processing conditions for testing,” he explains.

Advances in computational resources are also reducing barriers, says Craig Sather, associate director CMC and principal investigator, bioavailability enhancement at Lonza. “The computing power available today enables faster iteration and larger-scale simulation studies,” he notes.

Using Predictive Modeling/Simulation to Identify the Best Enabling ASD Technology

Extensive knowledge of various molecule attributes is essential to selecting the optimum ASD technology. “Successful ASD development begins with a thorough understanding of the molecule's characteristics and behavior,” states Aimee Spenceley, pre-form team leader, with Codis.

Selecting among spray drying, hot-melt extrusion, or alternatives such as co-precipitation and freeze-drying, depends on both API properties and excipient choice, Sekharan adds.

Solubility limitations, thermal properties, crystallization tendency, chemical stability, and interactions with potential excipients all guide selection of the most appropriate enabling technology and processing conditions, whether SD, HME, or alternative approaches, Spenceley comments. “The objective is to develop a formulation that delivers enhanced bioavailability while maintaining a suitable stability profile and sufficient downstream processability to create a commercially viable drug product,” she concludes.

SD requires the API and polymer to be compatible with a common, process-appropriate solvent system, whereas HME requires a thermally stable, thermoplastic polymer with suitable melt viscosity at a temperature the API can tolerate, says Sekharan. “Some polymers can be used with either route, so the choice ultimately depends on drug–polymer miscibility, target drug loading, thermal stability, solvent constraints, processability, scale, downstream processing, and the intended dosage form,” he says.

Because every molecule presents a unique set of challenges, Marsh emphasizes the importance of starting with understanding the specific problem statement and physicochemical properties of the compound and letting the data lead to the right formulation.

Data gathered using micro-scale screening approaches that allow rapid assessment of polymer compatibility, process suitability, and formulation performance while minimizing API consumption can be used to model and extrapolate formulation and process behavior, establishing operating windows that support scalability and technology transfer, Spenceley observes. “This strategy enables early identification of the most promising development pathway, reducing technical risk and accelerating progression towards a robust commercial process,” she says.

Physical Experimentation Informs Predictive ASD Modeling

ASD development is best supported by an integrated modeling and experimental strategy. “Structure-property assessments, thermodynamic and physical-stability models, spray-drying process models, dissolution models, and physiologically based biopharmaceutics models can be used together to assess API-polymer compatibility, drug-loading limits, crystallization risk, manufacturability, and the translation of in vitro performance to in vivo exposure,” Sather explains. He adds that these models are most useful when they generate hypotheses that can be tested using targeted, scale-relevant experiments.

Sather highlights scale as being particularly important, as changes in droplet size, drying kinetics, secondary solvent removal, and thermal history can affect ASD phase behavior even when conventional bulk characterization and dissolution results appear comparable. This issue came to light in a recent development program in which surface-sensitive analysis identified crystallization in solvent-wet ASD produced at larger SD scale that was not detected initially by bulk solid-state methods. “This finding led to a more integrated evaluation of formulation composition, process thermodynamics, drying kinetics, and wet-product handling,” he notes.

Specific experimental methods used at Codis to screen excipients, including thermal miscibility studies, film-casting assessments, and liquid-liquid phase separation investigations, provide an early understanding of ASD prototype formulation behavior and help define compositions that can deliver enhanced solubility while maintaining physical stability, according to Spenceley. Micro-dissolution testing in biorelevant media is used to benchmark solubility and supersaturation behavior and provide an indication of likely in vivo performance, while micro-flux studies can be used to assess permeation enhancement potential. “These tools generate data that can be modeled and extrapolated to support formulation optimization, reduce development risk, and improve confidence in progression towards scalable ASD products,” she contends.

Modeling/Simulation Techniques for Formulation Development

Several techniques are being used for predictive modeling and simulation applied to ASD formulation. Two models widely used to predict the physical stability of ASDs, according to Marsh, include perturbed chain-statistical associating fluid theory (PC-SAFT) and conductor-like screening model for realistic solvents (COSMO-RS), which are used to predict favorable or unfavorable interactions between APIs and common polymers used in ASD formulations. Marsh notes, though, that these models are limited because their applicability for assessment of long-term stability of ASDs is not universal.

PC-SAFT is a molecular thermodynamics model, while CSOMOS-RS is a quantum chemistry-based equilibrium thermodynamics method. They are used to predict attributes such as density, thermal expansion coefficient, glass transition temperature, and heat capacity and to calculate phase equilibria and predict chemical reactions within ASDs.2

Some accelerated stability modeling techniques are also used, including the accelerated stability assessment program (ASAP) and advanced kinetic modeling (AKM).2 These models are particularly useful for ASD formulations that tend to crystallize or undergo chemical degradation.

The integration of thermodynamic modeling, multiscale molecular simulation, and targeted experiments has also been shown to be effective, according to Sekharan. In 1 study, PC-SAFT indicated hydration-induced liquid–liquid phase separation, while coarse-grained simulations showed increasing drug-rich aggregation with drug loading, and microscopy confirmed that, at 40 weight%, a drug-rich interfacial layer impeded further dissolution.7 “This study demonstrates how modeling can identify formulation-specific release mechanisms and guide experimental design rather than simply rank candidates,” observes Sekharan.

What Predictive Modeling and Simulation Tools are Useful in ASD Formulation Development?

Numerous tools leveraging various techniques provide ASD formulators the ability to model and simulate important aspects of ASD formulation performance. Accelerated stability modeling tools, such as FreeThink ASAP, are increasingly used to predict long-term stability risk and support formulation selection, according to Spenceley.

Marsh highlights Simulations Plus’s ADMET Predictor as a tool for predicting physicochemical properties and informing experimental design before moving to in vitro screening. ADMET Predictor, together with Simulations Plus’s GastroPlus, can also support the prediction of in vivo performance.

Lonza also constructs physiologically based biopharmaceutics models (PBBMs) that combine in vitro data with in silico predicted properties to determine which formulations perform best in various simulated animal models and physiological conditions. “These models can be used to evaluate food effects, proton pump inhibitor effects, and performance across patient populations, helping us better understand formulation risks and identify strategies to mitigate them before clinical development,” Marsh notes. This approach is critical, he adds, because relying on in vitro testing alone can overlook the interplay of formulation and physiology and rule out potentially viable formulation options.

Schrödinger believes a multi-scale approach for predictive modeling of formulations is essential because, notes Sekharan, ASD performance is inherently a multiscale problem. He explains that QM and atomistic MD characterize API-polymer energetics, specific interactions, and compatibility, while coarse-grained MD extends the analysis to phase separation, aggregation, and dissolution mechanisms over larger lengths and time scales.

FEP+, a proprietary physics-based free energy perturbation technology for computationally predicting protein-ligand binding, can be used to calculate relative free energy differences associated with molecular transfer between solvent environments, providing an additional quantitative approach to questions such as solvation and partitioning. Glass-transition and penetrant loading calculations assess molecular mobility and water uptake.

In addition, these capabilities can be integrated with established thermodynamic methods such as PC-SAFT and Flory–Huggins, as well as dosage-form, biopharmaceutic, and PK models to connect molecular interactions with bulk phase behavior, tablet performance, and in vivo exposure and guide experimental follow-up, Sekharan observes. He adds that formulation ML further supports screening by mapping ingredient structure and composition to target properties and prioritizing candidates for testing.

The Role of AI and ML in ASD Formulation Development

There is growing interest across the industry in how artificial intelligence (AI) and machine learning (ML) approaches can be applied to drug development and formulation science, including for ASDs, particularly as these approaches apply to chemical structure and formulation data rather than traditional language-based datasets.

ML algorithms such as transfer learning, one-shot, zero-shot, and Bayesian-based optimization, for instance, help formulators uncover nonlinear relationships between various formulation parameters and target quality attributes that human researchers cannot detect.2 In 1 study, researchers trained multiple ML models over 3-6 months using stability data for over 600 ASDs to accurately predict ASD stability and other properties.8 PharmSD, meanwhile, is freely available web-based AI-driven computational platform for prediction of the physical stability, dissolution type, and dissolution rate of ASDs.9

Mechanistic models, Sather says, remain particularly valuable because they are grounded in established physical and thermodynamic principles and can be applied even when large training datasets are unavailable. Best results with ML models are obtained, believes Sather, when they are paired with historical ASD expertise, high-quality experimental data, and mechanistic understanding. “The value of these models ultimately depends on the quality of underlying data and scientific rigor used to develop and validate them. Successful implementation requires expertise in both data science and formulation development to ensure model outputs are relevant, reliable, and actionable,” he observes.

Lonza, for instance, has demonstrated the use of advanced ML algorithms for particle size prediction for SD dispersions to support manufacturability assessment across scales. “Our success has shown that AI and ML can expand the existing experimental and predictive modeling toolkit,” he comments.“Their role, however,” he continues, “is not to replace experimentation or scientific judgment, but to focus development on the highest-value experiments, reduce API consumption, and improve confidence in performance, manufacturability, and scalability across the product lifecycle.”

Future Use of Predictive Modeling/Simulation for ASD Formulation Development

As understanding of ASD kinetics and behavior mechanisms grows, modeling and simulation techniques advance further, and computational speed increases, predictions of general ASD formulation properties, long-term stability, and in vivo performance will improve and become a more valued and notable component of ASD formulation development.3

Marsh expects the focus to be on improving prediction of long-term ASD formulation stability. “That will require generation of a significant amount of high-quality data and thus significant investment, but the resulting models may support higher active loading in ASD intermediates and thus optimized active loading in final solid dosage forms, reducing pill burden and improving the patient experience,” he says.

Sather anticipates better integration of modeling tools and data streams to allow teams to evaluate complex formulation challenges more efficiently and build on prior knowledge. “Ultimately, predictive modeling is most powerful when combined with experimental data, scientific expertise, and a clear understanding of the development pathway. By integrating these capabilities early on, drug developers can make more informed formulation decisions, reduce development risk, and build confidence in formulation performance, manufacturability, and scalability through to commercialization,” Sather believes.

Use with Caution

Regardless of future advances in predictive modeling/simulation for ASD formulations, Marsh emphasizes that human expertise remains the gold standard when it comes to interpreting in vitro data, understanding model inputs and outputs, and informing formulation and development strategies for drug developers. Access to sufficient quantities of high-quality experimental data is also essential and a major challenge, according to Sather.

It is important to remember that models and simulations will always provide answers; informed interpretation is necessary to ensure those outputs are meaningful.3 Models/simulations used to aid ASD formulation development should be selected and applied with careful thought and consideration. In addition, more research is needed to overcome current limitations and improve their performance.

References

1. Predicting physical stability of amorphous solid dispersions (ASD). Crystal Pharmatech www.crystalpharmatech.com/predicting-physical-stability-of-amorphous-solid-dispersions-asd.html
2. Bhatta HP, Han H-K, Maharjan R, and Jeong SH. Recent techniques to improve amorphous dispersion performance with quality design, physicochemical monitoring, molecular simulation, and machine learning. Pharmaceutics. 2025 Sep 24;17(10):1249. doi: 10.3390/pharmaceutics17101249
3. DeBoyace K and Wildfong PLD. The application of modeling and prediction to the formation and stability of amorphous solid dispersions. J. Pharm. Sci. 2018 107 (1);57-74. doi: 10.1016/j.xphs.2017.03.029
4. Walden DM, Bundey Y, Jagarapu A, Antontsev V, Chakravarty K, Varshney J. Molecular simulation and statistical learning methods toward predicting drug-polymer amorphous solid dispersion miscibility, stability, and formulation design. Molecules. 2021 Jan 1;26(1):182. doi: 10.3390/molecules26010182.
5. Higginbotham T,Meier K,Ramírez J, and Garaiza A. Predicting drug-polymer compatibility in amorphous solid dispersions by MD simulation: on the trap of solvation free energies. Mol. Pharmaceutics (2025) 22 (2): 760–770. doi: 10.1021/acs.molpharmaceut.4c00810
6. Gao H, Wang W, Dong J, Ye Z, and Ouyang D. An integrated computational methodology with data-driven machine learning, molecular modeling and PBPK modeling to accelerate solid dispersion formulation design. Eur. J. Pharm. and Biopharm. 2021 Vol 158; 336-346. doi:10.1016/j.ejpb.2020.12.001
7. Walter S., et al. Predicting the release mechanism of amorphous solid dispersions: a combination of thermodynamic modeling and in silico molecular simulation. Pharmaceutics 2024, 16, 1292. doi: 10.3390/pharmaceutics16101292.
8. Han R, Xiong H, Ye Z, Yang Y, Huang T, Jing Q, Lu J, Pan H, Ren F, Ouyang D. Predicting physical stability of solid dispersions by machine learning techniques. J Control Release 2019 Vol 311-312; 16-25.
9. Dong J, Gao H, Ouyang D. PharmSD: A novel AI-based computational platform for solid dispersion formulation design. Intl J Pharm 2021 Vol. 604; 120705. doi: 10.1016/j.ijpharm.2021.120705

About the Author

Cynthia A. Challener, PhD, is a contributing editor to PharmTech.