▎WuXi AppTec content team editor
Recently, ChatGPT has been popular all over the Internet. Yesterday, Microsoft said it would integrate ChatGPT into search engine Bing and web browsers, and Google today demonstrated its artificial intelligence dialogue system called Bard. These systems can provide comprehensive and integrated answers based on complex questions provided by users, from detailed travel plans to analyzing a company's operational strategy. In the field of biomedicine, the application prospect of ChatGPT has also received widespread attention. Today, WuXi AppTec's content team will look forward to the application of this emerging AI model in the bioindustry based on publicly available information.
Image source: 123RF
The right-hand man in scientific exploration
Nowadays, scientific research is developing rapidly, hundreds of scientific papers are published every day, and how to keep up with the pace of scientific research is a challenge that researchers need to face. Based on ChatGPT's AI system, Microsoft has developed an AI system called BioGPT, which has been trained on more than 15 million abstracts on the scientific literature website PubMed to quickly provide relevant answers based on users' questions. In PubMedQA detection, this AI model achieves 81.0% accuracy.
Image source: Reference[8]
In introducing the browser that integrates ChatGPT, Microsoft said that the system can open a new window when reading lengthy earnings reports, allowing users to distill the main points of the article by asking questions and comparing them with other earnings reports. Applied to scientific literature, this system is expected to change the way we query and read papers in the future. The artificial intelligence system can not only help us find literature, but also "focus on it with one click" and compare it with other literature, greatly improving the speed of obtaining information from scientific literature.
Uncover scientific insights
The large-scale language model behind ChatGPT uses the analysis of massive amounts of human language data to learn the grammar and other features of human language. This learning method can also be used to interpret genomic DNA sequences. Technology company Nvidia pointed out at this year's JP Morgan Healthcare Conference that with the acceleration of the speed and cost of next-generation genome sequencing, our ability to sequence genomic DNA has now surpassed the ability to analyze DNA sequences and gain insights from them. Faster and more efficient processing of massive genome sequence information is inseparable from artificial intelligence. By analyzing DNA sequences in the same way as human language, large language models can speed up genome splicing, the discovery of genetic mutations, and present findings to researchers in the form of human dialogue.
For example, a gene sequencing analysis system integrated with ChatGPT may process a patient's genome sequencing data to give a summary that a mutation in the patient's X gene may cause a rare genetic disorder Y, supporting clinicians to make faster decisions.
▲ Large language models and generative artificial intelligence are essential for genomics (Image source: Nvidia official website)
Powering scientific breakthroughs
Artificial intelligence systems based on large language models have been used to learn the relationship between amino acid order and protein structure and function in proteins, helping to manually design new proteins. A recent article published in Nature Biotechnology
thesis
Using the ProGen system based on a large language model, the researchers designed a new lysozyme with similar activity to natural lysozyme. They say the new technology could be more powerful than the Nobel Prize-winning directed evolution protein design technique, injecting new life into the field of protein engineering.
AI systems such as ProGen can design new proteins with specific functions from scratch (Image source: Reference [13])
Improve the efficiency of scientific paper and medical report writing
Articles published recently in Nature and The Lancet Digital Health point out that an important future application of ChatGPT is to free scientists and doctors from some repetitive tasks and better focus on scientific research and treating patients. For example, many researchers are already using ChatGPT to help write the background material portion of a scientific paper, or to aid in editing a paper. In a hospital environment, ChatGPT has the potential to replace doctors writing reports with standard formats, such as discharge summaries.
What challenges need to be overcome?
Although ChatGPT has broad application prospects in the field of biomedicine, the industry also pointed out some hidden dangers in this system. For example, one of the shortcomings of current large-scale language systems is that the authenticity of the information provided needs to be improved. Since ChatGPT provides answers based on learning from existing linguistic data, its responses are also affected by untrue, biased, or outdated knowledge in the database. This means that for highly specialized topics, large language systems are likely to provide incorrect answers if they are not trained on enough specialized data. Researchers with sufficient expertise can still find and correct these problems, but users without expertise can easily be misled.
In addition, the language data of training ChatGPT also contains human historical biases, including race, gender, culture, age discrimination and other adverse factors. Since these historical biases are widespread in language databases and difficult to manually culle, how to prevent ChatGPT from outputting harmful speech based on this data is another challenge that needs to be addressed.
Some researchers point out that it is critical to establish norms and regulations for the use of ChatGPT to ensure that the technology is used properly, transparently and fairly. For example, several academic journals, such as Nature, have issued statements pointing out the need to explicitly point out the use of large language models such as ChatGPT when submitting academic papers for publication.
Dr. Eric Topol, a well-known scholar at the Scripps Research Institute, looked ahead to the future of AI applications and said that AI systems, including large language models, are expected not only to help diagnose cancer, but also to enhance the understanding of diseases by linking features in scanned images of the body with words in academic literature. He also stressed that these efforts should be carried out under expert supervision.
Generative AI like ChatGPT is advancing rapidly, and how researchers choose to use them will determine our future. "2023 is just the beginning!" Dr. Topol said.