Mapping and studying human genes is a Sisyphean task, with around 20,000 genes to analyze! This is where artificial intelligence is making a difference, particularly with OpenAI’s new model, o1, which is transforming genetic research.
Dr. Catherine Brownstein, a geneticist at Boston Children’s Hospital, works on rare and complex medical cases where patients have symptoms that no one has seen before. These patients described as ‘medical refugees,’ they often have a long and fruitless search for answers.
In the past, Dr. Brownstein would spend hours researching, poring over multiple studies to connect genetic mutations with physical symptoms (known as phenotypes). With OpenAI o1, this process has become much faster. Instead of manually sifting through articles, she can ask the AI questions like “What is the role of citrate in the bladder?” The AI quickly provides explanations, including potential connections to bladder health.
For example, when looking at a genetic mutation, it does not just give one answer, it shows different options, like whether the mutation increases or decreases activity. This flexibility is important because small details can make a big difference in diagnosing or treating a condition.
Dr. Brownstein emphasizes how much time this tool saves. The AI focuses on meaningful insights, making it easier to address the toughest cases. For her and many others in genetics, AI is not just a helpful tool, it is becoming a vital part of solving medical mysteries.
Why is this important?
Genetic research often relies on a process called cell sequencing, which allows scientists to analyze the DNA or RNA within individual cells. Especially for understanding diseases and uncovering how the human body works. However, the amount of data generated by cell sequencing can be overwhelming, making it difficult to analyze efficiently.
With OpenAI o1 data can be processed much faster and with fewer errors than humans. By automating time-consuming tasks, scientists can focus on making discoveries instead of getting bogged down in data management. This can be valuable for advancing fields like cancer research, personalized medicine, and rare genetic disorders.
Evaluating OpenAI o1’s reasoning capabilities in healthcare
OpenAI o1’s ability to excel in challenging scenarios lies in its advanced reasoning process, often referred to as a chain of thought. To solve difficult questions o1 uses a step-by-step reasoning process, similar to how a human carefully thinks through a problem before getting an answer.
Through reinforcement learning, o1 has learned to:
- Recognize and correct its mistakes
- Break down complex tasks into simpler steps
- Pivot to alternative strategies when the current approach is not working
This iterative process dramatically enhances the AI’s reasoning skills and accuracy, particularly in fields like healthcare.
Real world example: diagnosing a genetic disorder
To evaluate the difference between OpenAI o1 and its predecessor, GPT-4o, the models were tested with a complex medical case. They were asked to analyze a report containing a detailed list of phenotypes (observable traits) and excluded conditions, then propose a likely diagnosis.
GPT-4o’s response
GPT-4o identified Cornelia de Lange Syndrome (CdLS) as the most likely diagnosis, reasoning that several core features, intellectual disability, developmental delays, short stature, and distinctive facial characteristics aligned with the syndrome. However, it failed to account for some nuanced features, such as macrodontia (large teeth), which is not a hallmark of CdLS but is critical for differentiating conditions.
OpenAI o1-preview’s response
OpenAI o1-preview arrived at a more precise diagnosis: KBG Syndrome. By employing a chain of thought, the model systematically evaluated the phenotypes, including macrodontia, triangular face, thick eyebrows, skeletal anomalies, and developmental delays, which are hallmark features of KBG Syndrome. Additionally, o1-preview considered excluded phenotypes (e.g., microcephaly, cardiac anomalies) to narrow the diagnosis further.
Key insights from o1’s diagnosis:
- Inclusion of critical features: o1’s reasoning incorporated macrodontia, a hallmark of KBG Syndrome that GPT-4o overlooked.
- Exclusion of mismatched traits: the model ruled out syndromes with unlisted phenotypes, such as heart defects or unibrow (synophrys), which helped refine the diagnosis.
- Actionable recommendations: o1 not only provided a diagnosis but also recommended genetic testing for the ANKRD11 gene (linked to KBG Syndrome) and multidisciplinary care for the patient.
What makes OpenAI o1 unique?
OpenAI o1 has been specifically trained to reason through complex problems instead of simply providing quick answers. For example, in tests of mathematical problem-solving, earlier models like GPT-4 had a success rate of 13%, while OpenAI o1 achieved an impressive 83%. This advanced reasoning capability translates directly to healthcare, enabling researchers to uncover complex genetic patterns and relationships more effectively.
It can assist researchers in studying protein folding, a critical step in developing treatments for diseases like Alzheimer’s. Additionally, o1 can help to streamline drug development and testing.
The o1 can also be used to build smarter diagnostic tools or automate repetitive research tasks, saving time and resources. It’s a versatile tool for many healthcare applications.
o1 greatly improves over GPT-4o on challenging reasoning benchmarks. Solid bars show pass@1 accuracy and the shaded region shows the performance of majority vote (consensus) with 64 samples.
OpenAI has ensured o1 is safe and reliable by setting strict guidelines. The AI is designed to resist misuse, such as giving harmful advice, and has been extensively tested to meet ethical standards, making it a dependable tool for medical research.
Currently, o1-preview and o1-mini are available, while the full o1 model has not been released yet. The key difference lies in their capabilities: o1-preview excels at advanced reasoning for complex tasks, o1-mini is optimized for faster and simpler operations, and o1 is expected to integrate these strengths into a comprehensive, full-capacity model.
