AI platform identifies synergistic drug combination for resistant prostate cancer

A UCLA research team has developed an artificial intelligence platform that combines large language models with three-dimensional tumor testing to identify drug combinations that may counteract treatment resistance in prostate cancer. In a peer-reviewed laboratory study, the system ranked thousands of natural products as possible partners for enzalutamide, an established treatment for advanced prostate cancer. Of 150 candidates selected for testing, 23 showed evidence of synergy with enzalutamide, while berberine produced consistent enhancement across four prostate cancer models with different resistance mechanisms. The findings identify a combination for further preclinical investigation; they do not establish that berberine or any other compound is effective in patients.

The platform, called MAESTRO—Mechanistic Augmentation and Ensemble Selection for Therapeutic Ranking Optimization—was described in Advanced Science. It was designed to address two limitations of conventional combination-drug discovery. Many methods depend on extensive experimental training data, while others provide little explanation for why particular compounds might work together. MAESTRO instead uses a large language model to integrate information about drug actions, cancer biology and treatment-resistance mechanisms. It then evaluates possible combinations, ranks candidate compounds and generates biological explanations that can be examined in laboratory experiments. The system is intended to support, rather than replace, experimental validation and mechanistic analysis.

The researchers focused on prostate cancer because enzalutamide blocks androgen receptor signaling, which many prostate cancer cells depend on for growth. Some cancer cells eventually adapt by activating alternative pathways that let them survive despite the treatment. Combining enzalutamide with compounds that target those escape routes could therefore be a strategy for overcoming resistance. The challenge is the size of the search space: thousands of possible compounds must be considered in the context of multiple, interconnected biological pathways. Screening every possible pairing directly is laborious, and examining isolated pathways can miss compounds whose effects emerge from interactions among several systems.

To construct its analysis, the team supplied the model with scientific information relevant to prostate cancer, including resistance mechanisms and known drug actions. MAESTRO evaluated how these mechanisms might interact with androgen receptor blockade and assigned candidate compounds a potential role alongside enzalutamide. Rather than treating a single model output as definitive, the researchers asked the system to evaluate every candidate more than 100 times. They then assessed both the height of each compound’s ranking and the consistency of those rankings across repeated evaluations. Compounds that consistently occupied high positions were regarded as the strongest candidates for biological testing. This ensemble approach was intended to make the final selections more robust than predictions based on a single run.

MAESTRO was used to assess 3,592 natural products as potential partners for enzalutamide. From that set, the platform selected 150 high-confidence candidates for testing in three-dimensional prostate cancer spheroids. These tumor-like cell clusters can reproduce some features of cancer biology more effectively than cancer cells grown in conventional flat layers. The researchers also supplemented the models with oxygen to reduce cell death caused by oxygen deprivation, improving their ability to measure how combinations affected cancer cells under more suitable culture conditions. The screening was therefore not a purely computational exercise: computational rankings were followed by experimental evaluation in an experimental system designed to retain more of the cellular organization of a tumor.

Among the 150 compounds tested, 23 produced an effect greater than that expected from the individual compounds, a result characterized as synergy. Berberine, a naturally occurring compound found in several plants, emerged as one of the most notable candidates. The berberine-enzalutamide combination showed consistent synergy across four prostate cancer models that had different mechanisms of treatment resistance. That consistency was important because the models did not all resist enzalutamide in the same way. The result suggested that adding berberine enhanced enzalutamide activity across distinct resistance settings, although the observation remains confined to laboratory models and requires further evaluation.

The team also checked whether the experimental response agreed with the mechanism proposed by MAESTRO. Adding berberine to enzalutamide increased activity in the AMPK pathway, which participates in the regulation of cellular energy and stress. At the same time, the combination suppressed mTORC1 signaling, a pathway associated with cell growth and survival. The combination was associated with increased cellular stress and with reduced DNA replication and cell-cycle progression. These observations provided a biological account consistent with the platform’s ranking, linking the observed synergy to coordinated changes in pathways that help cells manage stress, energy use, growth and division. The study did not determine whether this mechanism operates identically in human tumors or in patients receiving treatment.

The experimental result also illustrates why the platform’s use of broad biological knowledge may be useful in combination screening. Berberine was not an obvious candidate for pairing with enzalutamide because it affects several pathways, making its behavior with another drug difficult to infer from a single target. A method centered on one pathway at a time could overlook such a compound. MAESTRO was instead able to place it among consistently high-ranking candidates, after which the spheroid experiments showed enhancement across all four resistance models. The work does not prove that the approach will consistently identify clinically useful combinations, but it demonstrates a way in which an AI-generated, mechanistically grounded hypothesis can be narrowed and tested. The researchers have proposed expanding the system with additional forms of biological data and applying it to other combinations, resistance mechanisms and cancer types.

Several boundaries remain before the findings could have clinical implications. The experiments involved cultured prostate cancer models rather than patient-derived tumor models, animal studies or people with prostate cancer. Although the researchers plan to evaluate the berberine-enzalutamide pairing further, those next steps have not yet been reported in the study. MAESTRO’s selection process also relied on scientific information and experimental data particular to prostate cancer biology, and its performance in other cancers or against other drug pairs remains to be established. The supported conclusion is therefore that an AI platform integrated with three-dimensional testing can identify and prioritize a candidate combination for additional preclinical research. The results point to AMPK-mTORC1 signaling as a possible mechanism to examine in future work, while the ultimate effects on safety, dosing and patient outcomes require separate clinical evidence.

Subject of Research: AI-guided discovery of drug combinations to overcome prostate cancer treatment resistance

Article Title: New AI platform helps uncover promising cancer drug combinations

Article References: New AI platform helps uncover promising cancer drug combinations. Original publication (Bibliographic details unavailable.)

Image Credits: AI Generated

DOI: Not provided

Keywords: artificial intelligence, prostate cancer, drug combinations, treatment resistance, large language models, berberine, enzalutamide, cancer spheroids, drug discovery, platform, helps, uncover