Digital and In Silico Transformation in Biologics Formulation

Authors: Tim Menzen, Ph.D., and Andrea Arsiccio, Ph.D., Coriolis Pharma

Accelerating Toward the Biotech Singularity

In The Singularity is Nearer (2024),1 futurist Ray Kurzweil predicted that, as computing power continues to become exponentially cheaper, human biology becomes better understood, and engineering becomes possible at ever-smaller scales, the digital transformation of the biopharmaceutical industry will unlock therapies capable of curing today’s untreatable diseases and extending healthy life spans. These trends are no longer theoretical. Computing costs have plummeted, and vast computational resources are now instantly accessible in the cloud, a prospect that was barely imaginable just two decades ago. At the same time, scientists can visualize and manipulate structures at the nanoscale and employ additive manufacturing technologies, such as 3-D printing, to construct intricate particles and systems with unprecedented precision.

Protein science offers perhaps the most striking example of this accelerating digital revolution. The 2024 Nobel Prize in Chemistry recognized David Baker of the University of Washington for pioneering computational protein design and Demis Hassabis and John Jumper of Google DeepMind for their breakthroughs in protein-structure prediction. Building on those achievements, Baker’s group used DeepMind’s AlphaFold 2 in 2025 to generate the first fully de novo monoclonal antibody — designed entirely in silico — with binding affinities only a few orders of magnitude below those achieved using the latest experimental methods.2,3 Given the exponential trajectory of these tools, antibodies with 10-, 100-, or even 1,000-fold stronger affinities are likely to be designed digitally in the near future, without requiring a single wet-lab experiment.

Where the Money Flows: AI’s Hold on Drug Discovery

The conviction that digital technologies can transform drug development is evident in the enormous flow of investment capital into companies building in silico platforms and computational drug design tools. The greatest share of that funding has concentrated in the domains of drug discovery and clinical development, particularly among those applying generative AI, machine learning (ML), and natural language processing (NLP) to automate and accelerate research.

A prominent example is Isomorphic Labs, founded by 2024 Nobel Laureate Demis Hassabis as a spinoff of Google DeepMind. Dedicated to “harnessing AI’s extraordinary power to reimagine and advance the drug design process from first principles,” the company announced in early 2024 a $45 million collaboration with Eli Lilly and another $37.5 million partnership with Novartis, together carrying up to $2.9 billion in potential milestone payments.4–6 Major pharmaceutical firms, including Bayer, Merck, and AstraZeneca, are pursuing similar partnerships and internal initiatives, all aiming to exploit generative AI and large-scale digital data to compress the discovery cycle.

This focus on discovery and clinical development is logical. Identifying the right candidates earlier reduces the likelihood of failure downstream, while optimizing clinical programs can drastically cut the time and cost to reach approval, especially as clinical trials remain the most expensive and time-intensive phase of drug development. From an investment perspective, these stages offer the fastest and largest potential return: more viable molecules entering the pipeline, a higher probability of producing blockbusters, and ultimately greater revenue through faster, leaner development.

Across the discovery and clinical continuum, AI and ML are now embedded in nearly every workflow. During discovery, algorithms conduct literature mining, target identification, virtual screening, and lead optimization. In clinical research, they assist in biomarker selection and patient stratification, adaptive protocol design and authoring, clinical-site selection, participant matching and recruitment, trial execution, biomarker and outcome analysis, automated report generation, regulatory submissions, and adverse-event monitoring. These applications already deliver measurable efficiencies and predictive insight, explaining why the digital revolution in pharma has, so far, concentrated here.

However, this investment landscape also exposes a blind spot: the relative neglect of the preclinical and formulation development stages, where digital tools could similarly accelerate development and reduce attrition. That gap is precisely where the next wave of in silico innovation must take hold.

Closing the Gap Between Discovery and Formulation

Despite the explosion of funding for AI-driven discovery and clinical development, investment in digital tools for the preclinical space remains modest. Yet this stage represents one of the greatest untapped opportunities for transformation.

Existing in silico platforms already simulate in vitro and in vivo conditions, model and predict pharmacokinetic and pharmacodynamic (PK/PD) behavior, and estimate toxicological profiles — capabilities that can dramatically improve decision-making long before a molecule enters the clinic. What remains underexploited is their potential to streamline formulation development, a critical determinant of whether a promising candidate ever becomes a viable drug product.

The preclinical domain is where efficiency gains can deliver exponential value: identifying liabilities early, learning from iterative digital–experimental cycles, and de-risking programs before they consume significant clinical resources. Expanding the use of in silico modeling here could not only accelerate development timelines but also reduce attrition and waste across the entire pipeline.

From Art to Algorithm: Digitalizing Formulation Science

Formulation development remains one of the most underappreciated and underestimated phases of drug development, despite being indispensable to the development of successful biologic products. The most potent therapeutic candidates emerging from discovery cannot advance to market if they are not stable, manufacturable, and of consistently high quality.

Each biologic drug substance must be formulated within an optimal matrix containing the right combination of excipients to preserve its integrity during fill–finish, storage, shipment, and administration. Factors such as molecular structure, concentration, pH, temperature, and exposure to oxygen or light can all influence degradation, aggregation, or potency loss. Freeze–thaw cycles and other process stresses further challenge stability. Determining the most robust formulation, therefore, requires mapping a multidimensional space of variables — each potentially interacting with the others — and experimentally testing all combinations is neither practical nor economical.

Even with accelerated stability studies, physical testing of every variable consumes vast time and resources. This complexity underscores the potential of AI and ML to simulate and predict formulation behavior, reducing dependence on purely empirical experimentation.

The concept of the digital twin provides a useful analogue for where formulation development is headed. A self-learning digital twin captures process understanding in a dynamic model that continuously integrates new data. With each iteration, its predictions improve, enabling more targeted experimentation and reducing the number of physical studies required.7

At Coriolis Pharma, we are adapting this philosophy to formulation development. The challenge lies in the fact that, unlike cell culture, which involves physics that are well characterized and readily modeled, the underlying mechanisms governing protein stability, excipient interactions, interaction with numerous contact-materials, including production equipment and primary packaging material, and long-term formulation behavior remain only partially understood. Consequently, formulation development is still as much an art as a science.

To bridge this gap, Coriolis’s vision is to employ a multiscale modeling approach, building and combining models at different levels of resolution. Atomic-scale simulations, though accurate, are not yet computationally feasible for large ensembles of protein molecules. Meso-scale models treat systems as continua, while process-level models focus on performance outcomes rather than molecular detail. By linking these hierarchies — feeding the outputs of lower-scale models into higher-scale ones — it becomes possible to construct an increasingly comprehensive representation of how formulation variables influence both stability and downstream pharmaceutical processes.

While the ultimate goal is a fully integrated, predictive digital twin for formulation, meaningful progress is already underway through machine-learning–driven optimization. As demonstrated by Narayanan and colleagues,8 iterative models can learn from prior experiments to design subsequent formulations more efficiently. The group’s early model in 2021 optimized two variables simultaneously; by 2025, their refined Bayesian framework could handle three interacting factors — a major advance toward practical, multidimensional prediction.9 The pace of progress suggests that exponential improvement is near, and that the digitalization of formulation development will soon redefine how biologics are stabilized, screened, and scaled.

Learning Loops for Smarter Formulation Science

To fully realize the promise of digital formulation development, the industry must deepen its understanding of the physics underlying biologic formulations: how proteins interact with their environments, how excipients stabilize or destabilize them, and how those mechanisms evolve over time. Only by coupling that knowledge with ML can predictive, physics-informed models truly accelerate formulation design.

A powerful model for this evolution lies outside the life sciences: the agile systems engineering approach exemplified by SpaceX. In 2006, the company’s first Falcon 1 launch ended in failure.10 Less than two decades later, SpaceX routinely launches and lands reusable rockets at scales once deemed impossible.11 The key to that transformation was the company’s disciplined embrace of the design–build–test–learn (DBTL) cycle: build rapidly, test aggressively, analyze the data, and apply the lessons immediately to the next iteration. Each failure became a feedback loop for improvement rather than a setback.

Coriolis Pharma sees clear parallels between this philosophy and the challenges of biopharmaceutical development. Drug formulation, like rocket design, involves countless variables, complex physics, and enormous cost when failures occur late in the process. Applying a DBTL mindset enables development teams to “crash early” in controlled, low-risk environments, iterating through predictive modeling, small-scale experimentation, and digital simulation before moving into linear clinical phases.

Failure in clinical development is far too late and far too expensive. By failing fast and learning early, developers can focus only on candidates with a high probability of success in the clinic. For Coriolis, embracing this agile, data-driven systems approach is not just about faster formulation optimization but fundamentally changing how knowledge is built, shared, and reused across programs to ensure that each experiment strengthens the collective intelligence of the next.

Kinetic Modeling, Automation, and the New Pace of Stability Science

Even the most agile development philosophy must ultimately contend with one unyielding constraint: the laws of physics. In space travel, for example, a rocket launched on the most fuel-efficient trajectory to Mars still requires two years to reach its destination.12 SpaceX can improve its odds of success by sending multiple rockets in parallel, but the fundamental delay — two years before new flight data return to Earth — cannot be shortened.

Formulation science faces an analogous dilemma. Generating long-term stability data simply takes time: demonstrating a drug product’s integrity over its full shelf life can require 24 months or more. These data are essential for confirming product robustness and informing formulation design, yet waiting two years for results would stall innovation. Running several promising formulations in parallel can de-risk the process, but it does not eliminate the temporal bottleneck.

To accelerate learning without violating physical limits, formulators increasingly rely on advanced kinetic modeling (AKM). This approach uses Arrhenius-like kinetic equations to extrapolate long-term degradation trajectories from data collected at earlier time points, typically within the first one to three months of testing.13 By modeling degradation kinetics, scientists can make educated predictions about stability and product lifetime, allowing confident decisions well before conventional studies would conclude.

Automation is also becoming integral to this acceleration. Robotic systems can execute repetitive stability assays with precision and consistency, reducing human error while dramatically increasing throughput. Meanwhile, augmented and mixed-reality platforms, integrated into laboratory information management system (LIMS) environment, enable scientists to visualize experimental protocols directly in the workspace, streamlining execution and documentation. Together, these technologies help formulation teams move faster, learn earlier, and make better decisions without waiting for the calendar to turn.

The Data Question

Beyond the challenges of understanding the physics of protein interactions, selecting the right modeling frameworks, and integrating AI and ML lies a more fundamental issue: data. Predictive models are only as strong as the information that feeds them. High-quality, relevant data are essential to train learning systems capable of generating meaningful and reliable insights. The more data available — provided they are consistent, well-curated, and representative — the more accurately those models can capture the true physical behavior of biologic formulations.

Such data may come from internal experiments or from published literature, but its value depends entirely on quality and context. Sparse, noisy, or poorly standardized data sets can mislead algorithms and obscure real correlations. By contrast, robust datasets grounded in sound experimentation not only improve model performance but also deepen mechanistic understanding, creating a virtuous cycle between empirical observation and digital prediction.

At Coriolis Pharma, this philosophy guides every digital initiative. The company is systematically generating and curating high-quality experimental data while constructing mechanistic models that reflect the physical realities of protein stability and excipient interactions. By ensuring that model inputs are scientifically grounded and that outputs are validated against real-world observations, Coriolis is building a foundation of confidence that allows digital models to predict formulation behavior across a wide range of conditions with ever-increasing fidelity.

Addressing Scientific, Financial, and Cultural Challenges

While the early results are deeply encouraging and there is little doubt that digital platforms will transform preclinical and formulation development, the transition will not happen overnight. Several interrelated hurdles must be addressed before this transformation becomes routine: scientific, financial, and cultural.

Scientific Understanding
A great deal of research remains to be done to fully elucidate the physics of protein–excipient interactions and to translate that knowledge into predictive, digital models. The use of AI and ML in formulation development is still in its infancy, constrained by limited data sets and incomplete mechanistic insight, but progress is accelerating. Each advance in in silico simulation, big data analytics, and digital-twin technology brings the field closer to a point where biologic formulations can be optimized virtually before entering the lab. The direction is clear, and it is only a matter of time before these capabilities mature into practical, validated tools.

Financial Perspective
Digital transformation requires a mindset shift in how investment and return are measured. The biopharmaceutical industry’s focus on short-term deliverables can slow adoption of technologies that yield compounding value over longer horizons. Building robust data infrastructures, simulation environments, and integrated digital workflows demands upfront investment and patience. Companies are already realizing near-term efficiencies from AI and NLP tools for documentation and reporting, which are important but relatively simple applications. The real gains will come from more complex, high-value implementations that will take time to develop and scale.

Cultural Change
Perhaps the most difficult barrier is cultural. True digital transformation depends on widespread trust in data-driven systems. Scientists, regulators, and patients alike must be confident that digital platforms can ensure safety and quality at least as reliably as traditional processes. Drug products must be manufactured under consistently controlled conditions, not just to meet compliance standards but to safeguard patient health and confidence. In the end, whether quality is guaranteed by human oversight or an AI-driven control system should not matter; what matters is trust. Building that trust through transparency, validation, and shared understanding is the industry’s most important and enduring challenge.

Redefining Preclinical Science Through Digital Partnership

Coriolis Pharma is dedicated to helping drug developers de-risk formulation development and strengthen the bridge between discovery and clinical success. We recognize that preclinical development, particularly formulation, plays a decisive role in determining whether a promising molecule ultimately reaches patients. Digital technologies, including AI and ML, hold extraordinary potential to accelerate and de-risk product development far beyond their current, limited applications.

Even with today’s capabilities, Coriolis is moving key development insights earlier than ever before. Through our formulatability assessment services, we extend our perspective upstream — from the drug product level into the biophysical properties of candidate molecules — enabling early identification of liabilities that might compromise stability, manufacturability, or efficacy. Using in silico models and predictive tools informed by more than 15 years of empirical experience across diverse molecular classes, we minimize the number of required physical experiments, reduce cost and risk, and accelerate the overall development timeline.

Our current digital focus centers on predicting the stability of established biologic modalities, such as antibodies and other proteins. However, we are rapidly expanding these capabilities to address more complex molecular formats, including emerging therapeutic modalities. In parallel, we are developing predictive models to identify the optimal primary packaging systems for each specific drug substance and formulation, recognizing that container closure, material compatibility, and environmental interactions are integral components of overall product stability.

Coriolis’s progress is powered by collaboration. We partner closely with technology vendors advancing digital modeling platforms and with drug developers seeking deeper insight into their biologic candidates. Remarkably, we can evaluate the intrinsic stability of a molecule based solely on its sequence. Our analyses yield detailed reports on molecular liabilities and formulatability, providing developers with actionable guidance.

Looking ahead, our vision extends toward an era where drug product development is supported by advanced digital ecosystems and ultimately by digital twins of human beings. Such models could one day represent individuals across diverse genetic, physiological, and cultural spectra, allowing precise prediction of pharmacokinetic and pharmacodynamic behavior without the need for animal testing. This would enable virtual preclinical studies and even early-stage clinical trials, accelerating the development process while dramatically reducing ethical and logistical burdens. Achieving that vision will take time, but every digital advancement we make today brings it closer to reality.

References

1. Kurzweil, Ray. The Singularity Is Nearer: When We Merge with AI. Viking, 2024.
2. Nobel Prize in Chemistry 2024. The Royal Swedish Academy of Sciences. 9 Oct. 2024.
3. Bennet, Nathaniel R, et al. “Atomically accurate de novo design of antibodies with RFdiffusion.” bioRxiv. 28 Feb. 2025.
4. “Our Tech.” Isomorphic Labs. Accessed 7 Nov. 2024.
5. Isomorphic Labs Announces Strategic Multitarget Research Collaboration with Lilly. Isomorphic Labs. 7 Jan. 2024.
6. Isomorphic Labs Announces Strategic Multitarget Research Collaboration with Novartis. Isomorphic Labs. 7 Jan. 2024.
7. von Stosch, Moritz. “A Model-Based Upstream Process Optimization.” DataHow Symposium. Zurich, Switzerland. 15 Jun. 2023.
8. Narayanan, Harini et al. “Design of Biopharmaceutical Formulations Accelerated by Machine Learning.” Pharmaceutics. 18: 3843–3853 (2021).
9. Waibel, Isabel, et al. “Bayesian Optimizaiton for Efficient Multiobject Formulation development of Biologics.” Mol. Pharmaceutics. 22 (2025).
10. Bergin, Chris. “Falcon 1 suffers launch failure.” NASA Space Flight. 24 Mar. 2006.
11. “SpaceX” SpaceX. Accessed 7 Nov. 2025.
12. “Hohmann Transfer Orbit.” NASA. Accessed 7 Oct. 2025.
13. Huelsmeyer, M et al. “A universal tool for stability predictions of biotherapeutics, vaccines and in vitro diagnostic products.” Sci. Rep. 13(1):10077 (2023).