A team of American scientists has created CancerGPT, an artificial intelligence (AI) model utilizing large pre-trained language models (LLMs) to estimate the potential impact of various drug combinations on the uncommon human tissues present in cancer patients. This innovative technique holds great promise for advancements in medical research, especially in situations where structured data and sample size are restricted, offering a significant breakthrough in the field.
A collaborative effort between the University of Texas and the University of Massachusetts involved the utilization of LLMs to extract existing information from medical research texts, which was then employed in biological inference tasks. The study successfully showcased the model’s remarkable accuracy.
The research paper states that the experiments conducted on seven rare tissues across various cancer types indicated that the prediction model based on LLMs achieved notable accuracy, even with minimal or no samples available for analysis.
The application of AI LLMs in medical research has garnered significant attention in 2023. An example of this is Ankh, an LLM developed by a team of experts from Munich and Columbia Universities in partnership with Protinea, which possesses the ability to comprehend protein communication, as reported by Decrypt. Additionally, another group of researchers utilized AI technology to identify three potential senolytic drug candidates, capable of targeting and eliminating “zombie cells” to potentially slow down aging and alleviate age-related illnesses.
CancerGPT, an LLM consisting of around 124 million parameters, is similar in size to the larger fine-tuned GPT-3 model that has approximately 175 million parameters. In the research study, a type of LLM called zero-shot GPT-3 was utilized to generate coherent responses. The accuracy of the LLM’s responses was assessed by comparing them with existing scientific literature, revealing that the LLM presented predominantly accurate arguments.
The researchers acknowledged that the accuracy of the LLM’s arguments cannot be confirmed on all occasions and may be vulnerable to hallucinations.
The researchers argue that despite having limited structured data for certain cancer types, there is still valuable information available in scientific literature. They harnessed the capabilities of pre-trained language models to utilize these existing resources, gaining “generalizability” in the process. This enhanced their ability to make predictions for future reactions and improve overall prediction accuracy.
Generalizability refers to a model’s capacity to utilize the knowledge acquired from training data in order to make predictions for new and unseen data. This is a key aspect that distinguishes AI from traditional deterministic computer programs.
The researchers suggest that future studies should further explore the approach and focus on developing an ensemble method that combines existing structured features with newly discovered prior knowledge found in LLMs.
While there may be some potential challenges, the findings of the study underscore the significance of AI technology in contemporary biology. AI has the ability to enhance personalization, improve efficiency, and increase the likelihood of successful outcomes, making it a transformative tool in the field.
#AI #ArtificialIntelligence #CancerGPT
