Developing effective oral drug formulations for medicines that don’t dissolve well in the body is a major challenge in pharmaceutical research. A newly published study introduces Andromeda 2, an advanced AI-powered system that designs and tests drug formulations in an automated mini-lab. This system intelligently builds on past experimental results to create better self-emulsifying drug delivery systems (SEDDS), which can improve how drugs are absorbed when taken by mouth. By automating and optimizing the formulation process, Andromeda 2 could help speed up drug development and increase the chances of finding successful treatments.
Key Takeaways
- Andromeda 2 achieved a 50% success rate in producing high-performing drug formulations, significantly outperforming its predecessor Andromeda 1 (17%) and traditional design-of-experiments methods (2%).
- The system identified 12 formulations meeting all four key product goals, compared to 6 by Andromeda 1 and none by standard lab approaches.
- One optimized formulation showed a drug loading more than three times higher than previously reported benchmarks, indicating better potential efficacy.
- Access to structured experimental data improved performance by 34%, highlighting the value of integrating prior knowledge into AI-driven design.
The researchers behind Andromeda 2 developed an “agentic” system—meaning it acts autonomously to plan and execute experiments—by combining computational reasoning with hands-on lab automation. This system continuously analyzes structured in-house experimental data, which includes detailed past results from similar drug formulation tests. Using this evidence, it decides which new formulations to create and test next, iteratively refining its approach to maximize success.
In this study, Andromeda 2 was tested on formulations of paclitaxel, a cancer drug known for poor oral bioavailability due to its low solubility. The system operated within a miniaturized automated laboratory environment, which allowed it to rapidly produce and analyze small batches of formulations within a fixed budget. Its performance was benchmarked against Andromeda 1, a previous probabilistic optimization model used on dozens of live projects, and a traditional wet-lab design-of-experiments (DoE) campaign, which is a systematic but manual approach to testing formulations.
Key to Andromeda 2’s success is its ability to reason over structured experimental evidence—essentially, organized datasets of past lab results—and use that information to guide its decision-making. This contrasts with earlier models that relied more heavily on probabilistic guesses or random exploration. By integrating computational tools with actual lab experiments in a closed loop, Andromeda 2 can efficiently home in on formulations that meet multiple target product profile (TPP) objectives, such as drug loading capacity and release characteristics.
The study’s results showed that Andromeda 2 not only found more high-performing formulations but also achieved better median drug absorption metrics (measured as area under the curve, or AUC) than the comparison methods. One standout formulation reached an apparent drug loading of 19% by weight at the first measurement, substantially exceeding the 5.7% loading reported for a previously published paclitaxel formulation. This suggests the potential for improved oral delivery of challenging drugs using AI-guided formulation development.
Looking ahead, systems like Andromeda 2 could transform pharmaceutical research by automating complex experimental workflows and leveraging accumulated lab knowledge to accelerate discovery. While this study focused on paclitaxel, the approach could be extended to other poorly soluble drugs, potentially reducing time and costs in drug development. Future work may explore integrating even richer data sources and scaling up the automated lab capabilities to tackle a broader range of formulation challenges.
Based on research published on arXiv by Michael M. Craig, Riley J. Hickman, Yingshan Ma et al..