Using AI and Molecular Dynamics to Overcome the Poor Solubility Challenge

For study, presented in poster form by Dineli Ranathunga, PhD, scientist III, Research and Development at Thermo Fisher Scientific, at AAPS PharmSci 360, held Nov 9-12, 2025 in San Antonio, Texas, investigators sought to use in-silico modeling to identify the best formulation approaches for use with poorly soluble drugs in order to support early decision making. Dr. Ranathunga, Thurman Falk, scientist III, Early Development at Thermo Fisher Scientific; Tom Reynolds, PhD, manager, Research and Development at Thermo Fisher Scientific; Nairuti Milan Mehta, scientist II, Research and Development at Thermo Fisher Scientific; Thomas Yonker, VP of Operations at Cerevance; and Sanjay Konagurthu, PhD, senior director, Science and Innovation at Thermo Fisher Scientific conducted the study based on the premise that the pharmaceutical industry faces persistent hurdles in bringing new treatments to market, particularly when dealing with compounds that are difficult to dissolve.

Currently, between 70% and 90% of drug candidates under development are categorized as poorly soluble (1). This widespread issue necessitates the early integration of appropriate formulation technologies into the development pipeline to prevent expensive delays later in the process, according to Dr, Ranathunga and team. Their work demonstrates how integrated predictive modeling can be leveraged to guide these critical, early-stage decisions for poorly soluble drugs, they stated.

The approach utilizes computational tools to rapidly identify optimal solubilization methods. As a case study, the researchers evaluated the poorly soluble compound CVN424. The study began by assessing the chemical structure and physicochemical properties of the compound, using in-house artificial intelligence (AI) and machine learning (ML) models to predict properties when experimental measurements were unavailable. All computational work, including quantum mechanical calculations, molecular dynamics simulations, and statistical modeling, was performed using the Quadrant 2 program.

How can predictive modeling streamline formulation selection?

The initial step in this systematic approach involved applying the Developability Classification System to categorize CVN424 based on its inherent characteristics. A combined framework of AI, ML, and statistical algorithms was then employed to steer the formulation technology choice. This framework analyzed several critical drug characteristics, including lipophilicity (Log P), pKa, melting point, dose, solubility, precipitation kinetics, and thermal stability. The analysis generated recommendations for the most effective solubilization technologies designed to enhance both solubility and bioavailability.

Following the selection of initial technologies, further computational studies focused on identifying appropriate excipients. The selection of suitable excipients is vital, the study team noted, as it reduces the time and expense associated with the typical empirical screening of the vast number of commercially available excipients. High-level quantum mechanical calculations were used to perform a conformational distribution analysis of CVN424. Advanced quantitative structure-activity relationship models, incorporating AI and ML, generated molecular descriptors (such as electrostatic potentials and hydrogen-bond donors/acceptors). These descriptors, combined with molecular interaction energies derived from molecular dynamics simulations, helped analyze drug–excipient interactions. By comparing these analyses against an extensive excipient database, the study predicted suitable polymeric dispersion excipients. Furthermore, molecular dynamics simulations were conducted at varying drug loadings to evaluate how CVN424 dispersed within the excipient matrix and to calculate the maximum feasible drug loading for each promising excipient.

What formulation strategies were recommended and tested?

The results of the predictive modeling provided specific recommendations. For instance, the findings displayed the predicted success likelihood for various formulation technologies across three dose ranges (Table), recognizing that the dose amount significantly impacts the formulation strategy, particularly when exact dosing is still unknown. The study specifically recommended lead excipients and their calculated maximum drug loadings for use in amorphous solid dispersion formulations, such as those created via spray drying. Snapshots from the molecular dynamics simulations visually represented the dispersion behavior of CVN424 within the excipient matrix as a function of drug loading.

Based on these computational predictions, Dr. Ranathunga and team selected spray-dried formulations and nano-milled suspensions for further evaluation. Spray-dried intermediate formulations were manufactured using five polymers—HPMCAS-H, HPMCAS-M, HPMC E3LV, Eudragit L100, and Soluplus—chosen based on the insights from the modeling and initial screening. Nano-milled suspensions of CVN424 were also developed; these suspensions were subsequently isolated using a spray-drying process. The feasibility of the chosen formulations was assessed using such methods as physical characteristics, chemical and physical stability, and in-vitro performance, including biorelevant two-stage dissolution performance.

This work highlights the utility of an integrated predictive approach in accelerating the formulation process for compounds like CVN424. By identifying suitable excipient candidates and solubilization strategies through computational means, the study efficiently guided the selection of spray-dried and nano-milled formulations for later evaluation. Ultimately, this process minimizes reliance on empirical trial-and-error, demonstrating the clear potential of predictive modeling to accelerate the overall development timeline, according to the poster presentation.

The application of in-silico modeling in drug formulation is akin to using a sophisticated digital blueprint before laying a single brick in construction, the investigators stated. Instead of blindly testing hundreds of materials and combinations, the predictive technology narrows down the most structurally sound options, saving immense amounts of time and resources before physical experimentation even begins.

References

1. Ranathunga, D; Falk, T; Reynolds, T; Mehta NM; Yonker, T; Konagurthu, S. AI/ML Model Informed Early-Stage Drug Development. AAPS PharmSci 360 Poster Presentation. Nov 12, 2025.