How AI Helps Solve Medical Mysteries at Boston Children’s Hospital | OpenAI Forum
Researchers from Boston Children's Hospital and OpenAI collaborated to apply AI (specifically, OpenAI o3 Deep Research) to tackle the "diagnostic odyssey" of rare diseases. By analyzing complex genomic and phenotypic data, the AI model helped identify 18 new diagnoses in 376 previously unsolved pediatric cases, showcasing its ability to accelerate literature review, generate hypotheses, and uncover obscure but critical information. This human-in-the-loop approach aims to make diagnosis faster, more accessible, and more precise, offering hope for personalized medicine and improved patient outcomes.
This conversation highlights a research collaboration between the Manton Center for Orphan Disease Research at Boston Children’s Hospital and Harvard, and OpenAI, focusing on leveraging AI to accelerate rare disease diagnosis. Historically, rare disease patients face a “diagnostic odyssey,” often taking six to seven years to receive a diagnosis, leaving many undiagnosed despite modern genetic testing.
The Challenge of Rare Disease Diagnosis
Stav Rones, a participant in the Manton Center’s research, shared his personal struggle, experiencing symptoms as a teenager but enduring years of searching for a correct diagnosis. Initially misdiagnosed, it took him three to four years to identify his condition as neurological and another five years to confirm its genetic basis. This journey is common, as emphasized by Dr. Alan Beggs, Director of the Manton Center, who describes it as a “diagnostic odyssey.” The human genome, with its three billion bases and 20,000 genes, presents an enormous challenge. Around 8,000 to 9,000 genes are known to be associated with disease, and each can have numerous damaging variations. While next-generation sequencing allows for comprehensive genomic data collection, the sheer volume—hundreds of thousands to millions of genetic data points—overwhelms human analysts.
Dr. Katherine Brownstein, Scientific Director of the Manton Center’s gene discovery core, detailed the bottleneck: geneticists manually sift through variants, cross-referencing databases, complex ambiguous symptoms, and new research. This labor-intensive process involves:
- Examining known genes.
- Identifying de novo mutations (not present in parents but found in the child) or cases where children inherit two variants from carrier parents.
- Analyzing large genomic changes like deletions or duplications.
- Consulting multiple databases to rule out common variants.
- Extensively reviewing scientific literature for each potential gene. This process is incredibly time-consuming and often leads to dead ends, making it a prime candidate for AI assistance.
The Impact of Diagnosis
Getting a diagnosis, as Stav explained, provides crucial clarity, enabling individuals to understand their condition, adapt their lifestyle, and avoid harmful activities. It also connects them to a community of others with the same disease, fostering support and shared understanding. Critically, a diagnosis opens doors to potential cures and solutions, allowing patients to engage with companies and teams developing targeted therapies. For Stav, it was vital for family planning, enabling informed decisions about genetic implications. Dr. Beggs highlighted that a diagnosis can lead to personalized medicine, exemplified by Stav connecting with a company developing a drug specifically for his condition. Katherine shared an emotional anecdote about diagnosing a 91-year-old, underscoring the deep relief and understanding a diagnosis brings, even after decades of uncertainty.
AI’s Role in Accelerating Diagnosis
Suya Shringerpur, a machine learning researcher at OpenAI, explained how OpenAI o3 Deep Research, an LLM, was integrated into the workflow. The goal was to accelerate the literature search, identify hypotheses, and present prioritized lists to human experts. The team collaborated closely with Alan and Katherine to define prompt engineering, desired output formats, and pitfalls to avoid, especially concerning complex genetic data. The AI model’s output provides evidence-driven lists, detailing supporting literature and variant information, rather than a black box nomination. Experts then review and validate these hypotheses.
Tackling Hard Cases
The AI model was initially tested on solved cases to refine its accuracy and learn to avoid common error patterns, such as requiring high sequencing depth for variant confidence. After achieving 80-90% accuracy, it was applied to 376 previously unsolved, difficult pediatric cases. These were “cold cases” accumulated over 10-15 years, where manual methods had failed. Alan explained that the LLM processes raw genetic data (up to 500,000 genetic changes with 20-30 data columns per change) and combines it with patient metadata, including numerical ontological codes describing clinical phenotypes (e.g., weakness, foot drop). The AI then intersects this data with its vast knowledge base to nominate a small, focused number of potential genetic changes, significantly reducing the search space for human analysts.
Breakthrough Discoveries
From the 376 unsolved cases, the AI workflow surfaced evidence that led to 18 new rare disease diagnoses. Katherine provided a remarkable example: an old case involving vitiligo, transposition of the great vessels, and pulmonary hypertension. The model nominated the S1PR1 gene. A comprehensive literature search by the AI uncovered a 26-year-old paper proposing S1PR1 as an etiological factor for vitiligo, along with other papers detailing its involvement in relevant pathways. This discovery led to connecting with the original researcher who cloned the gene, highlighting the AI’s ability to unearth obscure yet critical information that human researchers might miss due to time constraints and the sheer volume of scientific literature. Furthermore, in one instance, the model identified two distinct genetic issues in a patient where clinicians had previously focused on only one, demonstrating its potential to reveal complex comorbidities.
Future Implications and Hopes for AI in Medicine
Looking forward, the Manton Center is working with OpenAI to build a secure and accessible tool so that the technology is not limited to tertiary medical centers. The goal is to allow any clinician or researcher to periodically rerun undiagnosed cases, as genetic knowledge is constantly evolving.
Transforming Medicine
- Sash: Envisions models continuously re-analyzing cases as new knowledge emerges, proactively flagging relevant findings to researchers. This would allow human experts to focus on the truly hardest cases.
- Alan: Foresees AI dramatically speeding up and making diagnosis more efficient, rapidly pointing to diagnoses based on known information. It could also synthesize existing information to propose novel hypotheses, not just based on existing knowledge. He emphasizes that AI will not replace human diagnosticians but make their process faster and more efficient.
- Katherine: Hopes for a future where every undiagnosed case is efficiently run through such tools, leading to accurate diagnoses and guiding patients toward appropriate treatments or clinical trials. She believes the positive impact on families facing rare diseases is undeniable.
Overcoming Bottlenecks
Discussion also touched upon the importance of early whole-genome sequencing, which can now be done for less than $1,000 within days or weeks. While insurance reimbursement remains an issue, the medical community advocates for its broader adoption. The ability of AI to connect seemingly unrelated aspects of a patient’s health and reveal diagnostic patterns that clinicians might miss, was highlighted as a major potential area for future development.
This study utilized OpenAI o3 Deep Research, but newer models like GPT Rosland are continually advancing, promising even greater capabilities in life sciences research. The collaboration underscores the immense potential for AI to bring relief and better care to humanity by transforming rare disease diagnosis.