The peer-reviewed study, made available for early access in Nature Communications in 2026, presents an expert-mimicking AI framework called E-MAF to accelerate the development of polymeric controlled-release systems. Combining a process-aware surrogate model with an optimized genetic algorithm, E-MAF emulates expert reasoning that progresses from coarse evaluation to detailed design, producing laboratory-feasible formulations tailored to targeted release profiles.
The researchers reported that the system reduced the design cycle, which normally takes 6-12 months, to approximately 1 hour. Studies involving 7 active pharmaceutical ingredients found that in vitro release, in vivo pharmacokinetics, and therapeutic effects matched or surpassed commercial benchmarks. E-MAF also delivered profiles that are difficult to achieve using conventional methods; while risperidone burst release within 1 h was maintained at ≤2%, near-zero-order constant-rate release was achieved in vivo for approximately 28 days. The findings indicate that the approach has the potential to accelerate the translation of controlled-release systems and drug delivery technologies into clinical practice and industrial production.
The study’s equally contributing authors were identified as Ying Qin, Rushuang Zhou, Yali Ming, and Bohao Li; Yuedong Yang, Yining Dong, or Yufei Xia were responsible for correspondence. Teams affiliated with City University of Hong Kong, Chinese Academy of Sciences, University of Chinese Academy of Sciences, Sun Yat-sen University, and Huili Pharmaceutical Co., Ltd. participated in the research. Huili employees Y.H. and H.W. reportedly provided the company’s proprietary formulation and process data. A patent application was filed for E-MAF and its applications, with Y.Q., R.Z., and Y.D. listed as inventors. The other authors declared no conflicts of interest.
During preparation of the article, ChatGPT and DeepSeek were used solely for language improvement under the authors’ supervision. The study was supported by various Chinese research programs and funds. The article was published under the CC BY-NC-ND 4.0 license, and its DOI was given as https://doi.org/10.1038/s41467-026-77619-5.
Why it matters
By shortening the workflow between candidate screening and detailed design in controlled-release formulations, this approach allows drug development teams to determine earlier which options should receive laboratory resources. The comparison of in vitro, in vivo, and therapeutic results with commercial benchmarks means the findings go beyond a claim of computational performance alone; however, the evaluation was limited to seven active pharmaceutical ingredients. It therefore remains unclear whether the same success can be replicated across different compounds, manufacturing conditions, and clinical applications. While the company’s provision of proprietary formulation and process data and the patent application demonstrate the industry connection, they also keep questions open about the extent to which the method can be reproduced by independent teams and made accessible.