In the high-stakes world of pharmaceutical development, formulation scientists are often caught in a relentless cycle of "try, test, and make again." The process of optimizing drug release—particularly for complex long-acting injectables and implants—is traditionally slow, iterative, and resource-heavy. A single release study can stretch from 30 to 90 days, and every reformulation requirement restarts the clock, leading to significant delays in clinical filings and soaring costs.
However, a paradigm shift is underway. Shawn Zhang, Ph.D., co-founder and CEO of DigiM Solution, is pioneering a method to bypass this traditional trial-and-error cycle. By leveraging high-resolution 3D imaging and advanced machine learning, DigiM is turning the "hidden" internal architecture of drug doses into predictive maps of drug release, effectively allowing researchers to see inside a formulation before it ever reaches a dissolution bath.
The Challenge: Why Conventional Testing Fails
The core frustration for formulators is that two drug products, identical to the naked eye and even under standard laboratory analysis, can behave completely differently once administered.
"From the outside, two products can look the same, but once you open them up, you see the primary particles are deposited differently because of different process conditions," Zhang explains. Conventional techniques like laser diffraction or standard scanning electron microscopy (SEM) are often insufficient because they provide only surface-level data or bulk measurements that mask critical internal attributes.
These internal variables—the spatial distribution of the active pharmaceutical ingredient (API), the configuration of the pore network, and the arrangement of excipients—are the "hidden variables" that dictate a drug’s performance. As Zhang notes, conventional testing might show identical particle sizes, yet the actual release profiles will diverge significantly based on invisible process changes, such as variations in stir rates or solvent viscosity during manufacturing.
Chronology of a Digital Transformation
DigiM’s journey began in 2014, long before the current generative AI boom, with a pragmatic focus on digital infrastructure and high-fidelity data. The company has spent the last decade building a robust network of over 50 specialized labs and synchrotron facilities. This infrastructure allows them to manage representative sample preparation, rigorous quality controls, and data reproducibility—the trifecta of requirements for moving imaging from "pretty pictures" to regulatory-grade data.
The Evolution of the Workflow:
- 2014–2018: Establishing the Infrastructure. DigiM focused on standardizing 3D imaging protocols, utilizing FIB-SEM (Focused Ion Beam Scanning Electron Microscopy) to mill away layers of a dose and create high-resolution 3D maps.
- 2019–2021: Integrating AI. The company moved beyond raw imagery by developing digiM I2S, software that uses supervised machine learning and convolutional neural networks to perform "semantic segmentation." This allows software to assign specific numerical labels to voxels in a 3D image, distinguishing between API, excipients, and pore space.
- 2022–2024: Generative Advancements. Recognizing that prediction was the next hurdle, the company pivoted to generative AI. This allowed for the creation of virtual microstructures that could be manipulated digitally to test different "what-if" scenarios, such as changing the concentration or spatial distribution of an API.
- 2024–Present: Validation and Industry Adoption. Landmark studies, including a 2024 Nature Communications paper co-authored with industry giants Genentech and Merck, validated these predictive models against physical reality, proving that digital simulations could accurately predict percolation thresholds in tablets.
Supporting Data: Turning Pixels into Predictors
The strength of DigiM’s approach lies in its ability to convert grayscale intensity—the raw data from a microscope—into quantitative insights. In their workflow, a domain expert labels a small region of an image, and the machine learning model applies this logic across the entire sample set.
This process results in what DigiM calls "instance segmentation," which isolates individual objects, such as a single pore or a specific particle, for detailed analysis. When this data is fed into their simulation platform, ivisLab, the system models how the physical structure evolves over time.
Crucially, the simulation accounts for mass-transport properties, including diffusivity and the evolution of the API network. Unlike traditional dissolution modeling, which requires fitting a model to existing, previously measured dissolution curves, the DigiM approach uses image-derived microstructure as a de novo initial condition. This means researchers can obtain a directional release prediction before a single batch is manufactured, saving the industry from wasting milligrams of material—which, in early-stage development, are often more expensive than gold.
Official Responses and Regulatory Traction
The pharmaceutical industry does not shift its standards easily, but DigiM’s work has caught the attention of the FDA. The agency’s interest in microstructure characterization has intensified as part of its ongoing mission to modernize bioequivalence pathways for complex generics.
In recent years, the FDA has released draft product-specific guidances that echo the methodologies championed by DigiM. For example:
- November 2025: A draft guidance for generic minocycline hydrochloride dental extended-release powder suggested an in vitro alternative to comparative clinical studies, explicitly calling for the characterization of microsphere porosity and spatial distribution using orthogonal methods.
- May 2026: A draft guidance for generic dexamethasone ophthalmic inserts listed microstructure imaging as a mandatory comparative study, citing research that utilized the exact imaging workflows DigiM has helped refine.
While these guidances remain nonbinding, they signal a major regulatory pivot. The FDA is increasingly moving toward a "totality of scientific evidence" framework, where high-resolution structural data provides a more reliable and reproducible metric for product performance than traditional, bulk-scale testing.
Implications for the Future of Medicine
The implications of this shift are profound. By utilizing generative AI to "digitally increase the amount of chocolate chips" (the API) in a formulation, scientists can identify the exact percolation threshold—the point at which the API forms a continuous, release-contributing network—without ever entering the wet lab.
This capability has three primary impacts on the pharmaceutical landscape:
- Cost Reduction: By limiting the number of physical manufacturing campaigns, firms can drastically reduce the cost of R&D.
- Accelerated Timelines: Moving from a 90-day physical test to a digital simulation allows for rapid iteration and "fail-fast" development cycles.
- Enhanced Safety: A better understanding of the internal architecture of long-acting injectables allows for more precise control over release profiles, reducing the risk of dangerous "burst" effects where too much drug is released at once.
As Shawn Zhang prepares to present his keynote at the PharmSci 360 conference in New Orleans, the message is clear: the era of "black box" drug formulation is ending. The future of medicine is transparent, digital, and built from the inside out. By bridging the gap between raw microscopy and predictive AI, DigiM is providing the tools to turn complex, microscopic structures into the next generation of safe, effective, and rapidly developed therapeutics.
The industry is now faced with a choice: continue to rely on the traditional, time-consuming cycles of the past, or embrace a digital-first strategy where the microstructure is not a mystery, but a blueprint. As the regulatory environment catches up to this innovation, the companies that adopt these predictive technologies will likely lead the next wave of pharmaceutical success.
About the Author
