Using Artificial Intelligence to Unlock Hidden Insights in Genomic Research - with Dr. Mark Kiel of Genomenon

Dr. Kiel establishes the framework for understanding real world data versus evidence versus insights, explaining how clinical literature contains pre-aggregated, pre-vetted patient journeys from subject matter experts. He clarifies that published case reports represent real patients with demographic data, genotypes, laboratory values, treatments, and outcomes—particularly valuable for rare disease and precision oncology.

• •Real world data (raw patient information) becomes real world evidence when organized into databases, which then yields actionable insights
• •Clinical literature contains de-identified patient journeys painstakingly documented by clinicians who saw them
• •This pre-vetted information has applications in both pharmaceutical drug development and clinical diagnostics
• •The literature represents decades of peer-reviewed insight across all human diseases

Discussion of the fundamental challenges pharmaceutical companies face: the sheer volume of genomic literature, heterogeneous data patterns across different article types and databases, and the complexity of human language describing genetic variants. Manual review creates bottlenecks that prevent teams from asking and answering critical questions at the scale needed for comprehensive disease understanding.

• •Volume of literature represents decades of collected, peer-reviewed patient insights across all diseases
• •Heterogeneous data patterns include different article types, supporting databases, and varied human language descriptions
• •Manual processes don't scale—creating a fundamental barrier to accessing available knowledge
• •AI capabilities are essential to overcome these bottlenecks and make comprehensive analysis possible

Dr. Kiel explains why standard GPT models are insufficient for genomic literature and how Genomenon's specialized 'genomic language processing' works. The system first identifies and extracts genes, variants, and patient data, then uses LLMs with RAG to understand entities in context and reconcile them to standard ontologies. Knowledge graphs connect relationships across data points for both direct recovery and indirect discovery.

• •Off-the-shelf language models cannot handle the complexity of genetic variant descriptions and cross-reference organization
• •Custom NLP identifies and extracts genes, variants, demographic data, phenotypes, treatments, and outcomes
• •LLMs with RAG reconcile extracted entities to standard ontologies for data interoperability across diseases
• •Knowledge graphs enable both data recovery (finding all relevant information) and making causal linkages for new discoveries
• •System is amortized across all human diseases and the full breadth of medical literature

Dr. Kiel makes the case that expert curation should be considered an AI capability rather than separate from it. Genomenon's curators train models, iterate improvements, and review outputs for high-stakes applications like regulatory submissions and multibillion-dollar drug development decisions. The level of human review is fit-for-purpose based on the use case and risk level.

• •Expert curators develop training datasets, test models, and iteratively improve AI performance
• •Human review is calibrated to use case—regulatory submissions and major investment decisions require full curation
• •Curators are trained in genetic curation and clinical information annotation
• •AI outputs must be accurate, traceable, and defensible from source references for pharmaceutical applications

Beyond time and cost savings, the real value is enabling previously impossible analyses. Genomenon can catalog every human disease, every patient ever described, with granular demographic, clinical, laboratory, treatment, and outcome data. The scalability allows customization for each disease's unique characteristics—different lab values, treatments, and endpoints—making comprehensive rare disease analysis economically feasible.

• •AI enables cataloging of every patient ever described across all human diseases—inconceivable with manual methods
• •Captures longitudinal patient journeys with granular detail: demographics, phenotypes, genotypes, treatments, outcomes
• •Customizable extraction for disease-specific idiosyncrasies (unique lab values, tests, treatments, endpoints)
• •Scalability transforms what would be economically impossible into routine analysis
• •Number of patients extracted, data granularity, and customizability are key measurable benefits

Dr. Kiel addresses the challenge of automating drug development by breaking it into component parts. While full automation is as complex as a Mars mission, many information gathering and assessment tasks are routinizable: identifying optimal trial endpoints, defining inclusion criteria based on genetic biomarkers, and estimating patient prevalence. Automating these aspects dramatically reduces costs and makes rare disease drug development economically viable.

• •Drug development can't be fully automated yet, but specific components involving information gathering are routinizable
• •Key automatable tasks: identifying trial endpoints, cataloging genetic biomarker inclusion criteria, estimating patient prevalence
• •Information assessment challenges are fundamentally different from wet lab empirical work
• •Automation reduces costs enough to make rare disease programs economically feasible where they previously weren't
• •Combined with newborn sequencing initiatives, this enables finding patients faster and prioritizing drug development for invisible rare indications

Discussion of why specialized AI models outperform general-purpose foundational models for genomics. The key is training on the right substrate (specific articles and content) and asking context-specific questions. Similar to how call center AI needs fine-tuning despite general language understanding, genomic intelligence requires knowing the field of view, scope, and how to extract meaningful information from domain-specific content.

• •Foundational models understand broad language but make poor domain-specific tools without fine-tuning
• •Success requires two elements: right substrate (specific article scope) and context-specific questions
• •AI must know its field of view and what content is relevant before extracting meaningful information
• •Offsetting one AI model with another for verification is becoming standard practice
• •Human judgment remains essential for deciding how to structure AI tool silos and verification systems

The episode concludes with three critical takeaways: treating literature as an organized data asset, using purpose-built AI tools that pair extraction with expert curation for decision-ready evidence, and building data foundations that enable automation of repeatable development steps. As more patient journeys enter the published record, teams with strong infrastructure will estimate populations more accurately and choose programs with greater confidence.

• •Clinical literature holds deep real world patient data but requires structure to use effectively
• •Purpose-built AI extracts variants/phenotypes, aligns to ontologies, maps relationships, and pairs with curation
• •Long-term gain is scalability—automating repeatable steps while maintaining accuracy for high-stakes decisions
• •Strong data foundations enable better patient population estimates and more confident program selection
• •Human value shifts to judgment tasks: structuring AI systems, verification, and strategic decision-making