UCLA MAESTRO AI Ranks 3,592 Compounds for Cancer Drug

Picking which two drugs to combine is a numbers problem. With thousands of candidates, testing every pair in the lab is slow and expensive, and guessing from the literature misses the unexpected. Researchers at UCLA say they have built a system that uses large language models to shrink that list, and it pointed them to a compound found in plants.

The system is called MAESTRO, short for Mechanistic Augmentation and Ensemble Selection for Therapeutic Ranking Optimization. In a UCLA Samueli release, the team says MAESTRO ranked 3,592 natural products as possible partners for enzalutamide, a prostate cancer drug first discovered and developed at UCLA. Lab tests then picked out berberine, a naturally occurring plant compound, which worked with enzalutamide in four prostate cancer cell models. The work is published in Advanced Science. All of it is in lab-grown cells. It is not a treatment and has not been tested in animals or people.

How MAESTRO ranks drug pairs

MAESTRO does not ask a language model a single question. It feeds the model literature-supported information about each compound’s targets and effects, plus curated knowledge of how signalling pathways in prostate cancer cross over when the main hormone pathway is blocked. The model then ranks candidates for how likely they are to work alongside the base drug and explains its reasoning in pathway terms.

Language models give different answers each time, so the team turned that into a measure of confidence. According to the paper, the ranking “is repeated until a minimum of 100 ensemble runs is accumulated for each compound”. The average rank shows how promising a compound looks. The spread of ranks shows how sure the model is. Candidates with a high average and a low spread go forward to the lab.

The paper also tested whether the method depends on one model. Rankings from open-weight models, including GPT-OSS-120B and Qwen3.5-35B-A3B, correlated strongly with those from OpenAI’s GPT-4o. The authors argue open models could make the method easier to reproduce as commercial models change.

From 3,592 candidates to one combination

The team took 150 high-confidence natural products from MAESTRO’s ranking and tested them with enzalutamide in 3D clusters of prostate cancer cells, called spheroids. According to the paper, 23 showed synergy, a hit rate of 15.3%. An earlier screen of kinase inhibitors, a class of targeted drugs, found 17 hits among 95 tested in one resistant cell line.

The spheroids were grown with extra oxygen. Large spheroids normally die in the middle because oxygen cannot reach the centre, and that dead core can hide whether a drug pair is working. The paper says the added oxygen reduces that problem, but it also cautions that the condition is not meant to copy the oxygen levels inside a real tumour.

Berberine stood out because it worked with enzalutamide in all four cell models tested. Two of them are resistant to enzalutamide and two are sensitive. “Berberine was not an obvious candidate because it affects several biological pathways,” said first author James C.H. Lo, a former postdoctoral researcher in the lab. The team found the combination raised AMPK activity and suppressed mTORC1 signalling, both involved in how cells manage energy and growth, which is consistent with the reasoning MAESTRO gave.

“Developing drug combinations currently relies on either brute-force screening or researchers hypothesizing potential candidates based on existing literature,” said senior author Neil Lin, an associate professor of mechanical and aerospace engineering and bioengineering at UCLA. “MAESTRO could substantially accelerate this process.” Other senior authors include Robert Damoiseaux, Cho-Jui Hsieh, Tanya Stoyanova and Andrew Goldstein.

A hypothesis machine, not a cure finder

This is TSN’s analysis. The interesting part of MAESTRO is how it handles the usual weakness of language models. Instead of trusting a single answer, it asks more than 100 times and treats disagreement as a warning. That turns randomness into a confidence score scientists can use to decide what to test. Lin describes the models as tools that “generate hypotheses that can be tested experimentally”, and the paper’s own abstract calls the result “proof of concept”.

The approach sits alongside other recent attempts to use computing to choose drug pairs, such as the UCL digital twins suggesting extra drugs for breast cancer and Stanford’s immune digital twins flagging a drug pair for preterm labour. Like the AI screen that helped find the antibiotic abaucin, the computer narrows the search and the lab still has to prove each result.

The berberine problem

The paper is frank about the gap between a lab hit and a medicine. It says berberine taken by mouth reaches the bloodstream poorly, “with bioavailability below 1%”, so any future use would need a way to raise exposure. The cells in the study were also grown under conditions chosen for screening, not to mimic a tumour.

Drug companies are also testing language models in early discovery, with Anthropic reporting Claude work at six partners. MAESTRO’s value will depend on whether its picks keep holding up in harder tests. The team plans to move next to patient-derived tumour models and animal studies, and to apply MAESTRO to other cancers. Those results, not the ranking, will show whether berberine and enzalutamide deserve a place in a clinical trial.