Enabling the clinical application of artificial intelligence in genomics: a perspective of the AMIA Genomics and Translational Bioinformatics Workgroup

Nephi A. Walton, Radhakrishnan Nagarajan, Chen Wang, Murat Sincan, Robert R. Freimuth, David B. Everman, Derek C. Walton, Scott P. McGrath, Dominick J. Lemas, Panayiotis V. Benos, Alexander V. Alekseyenko, Qianqian Song, Ece D. Gamsiz Uzun, Casey Overby Taylor, Alper Uzun, Thomas N. Person, Nadav Rappoport, Zhongming Zhao, Marc S. WilliamsJ. Am. Medical Informatics Assoc 2024 · OpenReview

TL;DR

This paper, authored by the AMIA Genomics and Translational Bioinformatics Workgroup, outlines critical factors needed to bridge the gap between AI's potential in genomics and its actual clinical application. It emphasizes the need for larger, more diverse datasets, robust data standards, and careful consideration of ethical implications to ensure AI benefits all patients equitably.

What Problem Does This Paper Address?

The paper addresses the significant challenge of translating the immense potential of Artificial Intelligence (AI) in genomics into tangible clinical benefits. While AI is rapidly transforming other fields like art, law, and education, its integration into genomics for diagnostics, personalized treatment, and risk prediction faces substantial hurdles. These include the inherent complexity and vast dimensionality of genomic data, the lack of comprehensive functional understanding of many genetic variants, and the need for AI models that are not only accurate but also interpretable, unbiased, and ethically sound, especially concerning patient privacy and health equity.

Methods

The authors, representing the AMIA Genomics and Translational Bioinformatics (GenTBI) Workgroup, developed their perspective through a consensus-building process involving multiple in-person and online meetings, alongside a review of pertinent publications. They identified and summarized key factors crucial for advancing the adoption and clinical use of AI in genomics, categorizing them into five main areas: Data Availability and Accessibility, Technical Aspects of Genomic Data, Ethical and Analytical Considerations, Stakeholder Engagement and Education, and Regulatory Frameworks. This approach allowed for a comprehensive, multi-faceted assessment of the ecosystem required for successful AI integration in clinical genomics.

Datasets

The paper does not focus on specific datasets used for training or evaluation of AI models. Instead, it discusses the need for larger, more diverse datasets. It references initiatives like the "All of Us" project and TOPMed, which aim to collect genomic data that better reflects population diversity, as crucial for developing more generalizable AI models and minimizing health disparities. The authors also highlight that current AI models in other domains, like ChatGPT and image generation models, are trained on vastly larger datasets (billions of text files, millions of images) compared to the available genomic data, underscoring the data gap in this field.

Key Insights and Ideas

  • Data is King, but Diversity is Queen: While the cost of genomic sequencing is decreasing, leading to more data, the sheer volume is still orders of magnitude smaller than what AI models in other domains are trained on. Crucially, existing genomic datasets often lack diversity, risking AI models that exacerbate existing health disparities.
  • Beyond the "Black Box": For AI to be clinically adopted, it must be interpretable and transparent. Clinicians, patients, and regulators need to understand how AI models arrive at their conclusions, not just what the conclusion is.
  • Standards are Essential for Scale: Standardizing genomic data representation and knowledge is critical for developing accurate, portable AI models that can be reliably used across different institutions and health systems.
  • Ethical AI is Non-Negotiable: AI in genomics must actively mitigate biases related to socio-economic status, race, and ethnicity. Patient privacy and data security are paramount, requiring robust safeguards and potentially novel approaches like federated learning.
  • Regulation Needs to Catch Up: Current regulatory frameworks, particularly from the FDA, are not fully equipped to handle adaptive AI approaches. A balance is needed between ensuring patient safety and fostering innovation.

Key Results

This paper is a perspective piece and does not present quantitative results from AI model development or evaluation. Instead, its "results" are the identification and articulation of key factors and challenges. The core findings are the five categories of factors identified by the AMIA GenTBI Workgroup as essential for enabling clinical AI in genomics:

  1. Data Availability and Accessibility: Growth of genomic data is promising, but still lags behind other AI domains; democratization of data through sharing initiatives is vital.
  2. Technical Aspects of Genomic Data: Functional genomics knowledge and standardized data formats (like GA4GH and FHIR) are crucial for AI model development and integration into EHRs.
  3. Ethical and Analytical Considerations: Privacy, accuracy, and bias are critical concerns that must be addressed through careful algorithm design and data handling.
  4. Stakeholder Engagement and Education: Clinician education on AI's capabilities and limitations, alongside transparency and interpretability, is key to trust and adoption.
  5. Regulatory Frameworks: Existing regulations need to adapt to the dynamic nature of AI, balancing safety with innovation.

Limitations and Future Work

The authors acknowledge that substantial informatics research and development are still needed. They emphasize the necessity of developing larger, more diverse datasets and establishing responsible frameworks for data sharing and analysis. Future work should focus on creating best-practice guidelines for AI in genomics, ensuring that AI applications are evaluated rigorously in clinical environments, are transparent, reproducible, and generalizable. A critical area for future work is the development of new regulatory frameworks that can accommodate adaptive AI technologies while safeguarding patient safety and fostering innovation. The paper also implicitly points to the need for advanced computing infrastructure to handle the scale of data required for effective AI in genomics.

Summary updated

Abstract

Given the importance AI in genomics and its potential impact on human health, the American Medical Informatics Association—Genomics and Translational Biomedical Informatics (GenTBI) Workgroup developed this assessment of factors that can further enable the clinical application of AI in this space.
BibTeX
@article{walton2024enabling,
  author = {Nephi A. Walton and Radhakrishnan Nagarajan and Chen Wang and Murat Sincan and Robert R. Freimuth and David B. Everman and Derek C. Walton and Scott P. McGrath and Dominick J. Lemas and Panayiotis V. Benos and Alexander V. Alekseyenko and Qianqian Song and Ece D. Gamsiz Uzun and Casey Overby Taylor and Alper Uzun and Thomas N. Person and Nadav Rappoport and Zhongming Zhao and Marc S. Williams},
  title = {{Enabling the clinical application of artificial intelligence in genomics: a perspective of the AMIA Genomics and Translational Bioinformatics Workgroup}},
  journal = {J. Am. Medical Informatics Assoc.},
  year = {2024},
  url = {https://openreview.net/forum?id=aLZvQZTktA},
}
Figure