AI Won't Save Subcutaneous Drug Development. Better Data Will.
Artificial intelligence has become the pharmaceutical industry's favorite answer to almost every R&D challenge. Faster decisions. Better predictions. Shorter development timelines. But when it comes to developing subcutaneous biologics, AI isn't the breakthrough. The data are.
As more companies convert successful IV biologics into subcutaneous (SC) formulations, or develop new therapies for SC delivery to start with, the industry faces a reality that is often underestimated: changing the route of administration changes everything.
An IV therapy enters the bloodstream immediately. A subcutaneous injection embarks on a far more complicated journey. Before reaching systemic circulation, a biologic must navigate the extracellular matrix, diffuse through tissue, interact with its environment, and ultimately enter the lymphatic or vascular system. Concentration, viscosity, injection volume, release kinetics, precipitation, and tissue interactions all influence whether enough drug reaches circulation to achieve the desired therapeutic effect.
Yet many of the critical decisions that determine success are still made with limited insight into how formulations will actually behave beneath the skin. That uncertainty is very expensive.
Development teams can spend years optimizing formulations, scaling manufacturing, and advancing programs before human bioavailability data reveals that a formulation simply isn't performing as expected. By then, millions of dollars and valuable time have already been invested.
For decades, the industry has largely accepted this uncertainty as the cost of doing business. Animal models provide useful information, but they rarely capture the full complexity of human SC absorption. Clinical studies remain essential, but they come late in the development process, after many of the most consequential formulation decisions have already been made.
This is exactly where predictive science should change the equation. Unfortunately, many conversations today begin and end with AI. That's backwards.
Machine learning models are only as good as the data they're trained on. If experimental data fail to capture the biological mechanisms that drive SC absorption, no algorithm, no matter how sophisticated, can reliably predict clinical performance.
The industry's competitive advantage won't come from having the most advanced AI. It will come from generating the most biologically relevant data.
That means investing in physiologically relevant in vitro models that reproduce key aspects of the subcutaneous environment. These systems allow scientists to measure release kinetics, diffusion, precipitation, and other mechanistic behaviors long before a candidate enters the clinic. Those insights not only help researchers understand why one formulation outperforms another, they also create the high-quality datasets needed to train predictive models with real clinical relevance.
This is where the future becomes exciting. Mechanistic experimentation and AI aren't competing approaches. They're complementary ones. Mechanistic science explains what is happening. AI identifies patterns across thousands of variables that humans cannot easily detect. Together, they create a far more powerful framework for predicting formulation performance than either approach can deliver alone.
The goal isn't to eliminate uncertainty. Biology is too complex for that. The goal is to eliminate avoidable surprises.
Every program that identifies a weak formulation before expensive preclinical or clinical work begins saves time, conserves resources, and allows teams to focus on candidates with the greatest chance of success.
As subcutaneous biologics continue to reshape therapeutic development, companies that embrace predictive formulation science will gain more than operational efficiencies. They'll make smarter decisions earlier, reduce development risk, and accelerate innovation where it matters most. The question facing the industry is no longer whether AI belongs in formulation development. The real question is this:
Are you generating the kind of data that AI can actually learn from, or are you simply asking better algorithms to make sense of inadequate science?
