Groups working in genomics, bioinformatics, and materials science started experimenting with similar approaches. Once-skeptical teams began to see the value of embedding AI pipelines into older frameworks rather than discarding them
The modern scientist has no shortage of data. Nearly 80 percent of all scientific information now lives in digital form, and in 2022 alone, more than 2.3 million journal articles were published worldwide. But while the data flood grows, the tools to make sense of it haven’t kept up. Many labs still rely on workflow systems built years ago, long before machine learning became central to research.
That gap matters. In diagnostics, drug development, and personalized medicine, speed is everything. Data is not just the byproduct of research-it is the foundation. Yet without systems capable of handling AI-driven workloads, even the most promising insights risk getting lost in translation.
It was inside one of Southern California’s premier research centers that computer scientist Aditi Jain found herself confronting this problem. Her work revolved around an open-source scientific workflow management platform software widely adopted across the global research community for large-scale studies. The challenge was deceptively simple: could this trusted but traditional platform be adapted to meet the demands of modern AI? What followed was less about reinvention and more about careful adaptation.
The system she inherited was reliable, trusted, and designed for large-scale batch computations. But machine learning requires something different iterative testing, flexible resources, and reproducibility. The question was whether an old platform could meet those new demands.
Her first test case came from the medical field: lung image segmentation.
