AI-HOPE-PM is a conversational AI platform designed to advance translational precision oncology by enabling users to perform integrative bioinformatics analyses through plain-language queries. The system is built on a retrieval-augmented generation framework—a method that enhances response accuracy by retrieving relevant information from structured datasets—and a fine-tuned biomedical LLM (LLaMA 3). Behind the scenes, the platform uses Python-based scripts to carry out statistical analyses and genomic data processing.

To enable robust analyses, we used curated multimodal datasets from TCGA, AACR Project GENIE, and cBioPortal. These datasets included harmonized clinical, genomic, and demographic variables. In addition, we generated synthetic SDOH variables using a validated Python script, guided by a literature-informed framework. These SDOH features included health care access, financial strain, food insecurity, social support, and health literacy. All datasets were preprocessed into standardized tab-delimited formats with annotated metadata describing each variable type. A full list of variables analyzed—including over 200 clinical, genomic, treatment, and SDOH fields—is publicly available [25], which also contains the source code, example queries, simulated data, and documentation for reproducing all analyses.

Users interact with AI-HOPE-PM via a GUI that accepts plaintext queries. The system interprets these queries using a natural language processing engine to define analytic tasks, including loading a dataset, stratifying cohorts based on genomic or SDOH features, and performing statistical analyses such as survival modeling or odds ratio testing. The resulting structured commands are executed programmatically, streamlining workflows that typically require multiple manual steps or coding expertise.

We evaluated AI-HOPE-PM’s query interpretation accuracy using 100 natural language prompts that reflected diverse real-world research scenarios in clinical genomics and health disparities. A team of expert reviewers established ground truth interpretations for each query to assess system performance. AI-HOPE-PM achieved an overall accuracy of 92.5%, with near-perfect accuracy (99.1%) for single-variable queries and strong performance (88.4%) for more complex, multivariable prompts. Most errors stemmed from ambiguous phrasing (eg, nonspecific end points), syntactic inconsistencies (eg, nested logic), or misalignment between user language and system variable mappings. To address these issues, AI-HOPE-PM integrates built-in clarification prompts and applies a domain-specific ontology to harmonize terminology and guide users toward more structured input. Future development will focus on improving the natural language understanding engine and refining internal mapping algorithms to further enhance accuracy and reproducibility.

To confirm the analytical fidelity of AI-HOPE-PM, we cross-validated its survival analyses, odds ratio outputs, and cohort stratifications against manually performed analyses previously published by our group using similar datasets and variables. These included studies investigating CRC disparities based on TP53, APC, and KRAS mutation status, treatment modality, and SDOH factors across TCGA and cBioPortal cohorts. The results generated by AI-HOPE-PM were consistent with those from traditional statistical pipelines in terms of hazard ratios, P values, and overall survival trends. This validation step supports the platform’s accuracy in replicating established findings and reinforces its reliability as a tool for real-time, natural language–driven bioinformatics analyses.

Although we benchmarked AI-HOPE-PM against established platforms such as cBioPortal and UCSC Xena, it is important to acknowledge that these platforms function through traditional GUIs requiring multistep, click-based interactions. This structural difference makes direct comparisons with AI-HOPE-PM—an intelligent, conversational AI system—challenging. In cBioPortal and Xena, executing multilayered queries or stratified analyses may involve multiple browser windows, dropdown menus, and manual dataset subsetting. In contrast, AI-HOPE-PM enables users to perform similar tasks via a single plain-language prompt, streamlining the process and reducing complexity. While speed remains an advantage for AI-HOPE-PM, we also validated its outputs through comparisons with previously published manual analyses, ensuring consistency and analytical fidelity. This intelligent design is intended to reduce the technical barrier for researchers and support scalable, real-time hypothesis generation in precision medicine Multimedia Appendix 1.

To ensure robustness in handling natural language variability, AI-HOPE-PM incorporates an interactive clarification mechanism that prompts users for additional input when queries are ambiguous or underspecified. Common edge cases include vague end points (eg, “better outcomes”), undefined comparison groups, or syntactic inconsistencies (eg, nested logic). In these instances, the system pauses execution and requests clarification through a structured prompt. Furthermore, AI-HOPE-PM uses a curated biomedical ontology to harmonize synonymous terms and align user inputs with internal variable definitions. These strategies support resilient query interpretation and maintain analytical fidelity across diverse and potentially ambiguous user queries (Multimedia Appendix 2).

To assess usability and speed, we benchmarked AI-HOPE-PM against existing platforms including cBioPortal and UCSC Xena. Biomedical researchers were asked to complete tasks such as dataset loading, filtering based on genomic or SDOH attributes, and initiating analyses. Task durations were measured using stopwatch protocols. AI-HOPE-PM consistently outperformed traditional tools in terms of execution time and ease of use, owing to its automation and intuitive language-driven interface.

To evaluate the capacity of AI-HOPE-PM to integrate and analyze SDOH alongside clinical and genomic data, we developed a set of simulated SDOH variables. These variables—including financial strain, food insecurity, social support, health literacy, and insurance access—were generated using a Python-based simulation framework informed by published epidemiological distributions and associations relevant to cancer outcomes. The simulation approach was designed to mirror the variability and prevalence observed in real-world populations [26], thereby enabling realistic case–control stratifications and hypothesis testing. Although these SDOH variables are simulated, they serve as a pragmatic proxy in the absence of widely available, high-quality, individual-level SDOH data within public genomic datasets. For full transparency and reproducibility, the simulation scripts are publicly available [25]. Future validation studies using empirical SDOH data from institutional and community-linked datasets are planned to further refine and expand the platform’s capabilities.

The platform supports several statistical methods commonly used in cancer genomics, including Kaplan-Meier survival analysis with log-rank testing, Cox proportional hazards regression, and odds ratio calculations for categorical comparisons. Output includes plots such as survival curves and forest plots, accompanied by narrative summaries that describe the findings in context. All outputs are backed by reproducible Python code logs, which are stored internally and can be exported for validation or inclusion in publications [13-15].

A formal usability study is underway to evaluate AI-HOPE-PM’s effectiveness and accessibility for biomedical researchers. Participants are comparing its interface, output quality, and query interpretation capabilities with those of GUI-based tools and other AI-driven platforms. While we did not perform head-to-head comparisons with generative systems such as CellAgent or AutoBA due to differing scopes, AI-HOPE-PM’s unique ability to integrate SDOH, clinical, and genomic data positions it as a novel tool for equitable and scalable precision medicine research.

To preliminarily assess usability, we conducted a small-scale case study involving six non-bioinformatician users, including oncology fellows and public health researchers. Participants were asked to perform common research tasks using AI-HOPE-PM—such as loading datasets, selecting cohorts by genomic and social variables, and running survival analyses—using only natural language queries. All users completed the tasks successfully, with positive feedback highlighting the intuitive interface, rapid execution, and elimination of the need for coding expertise. These findings provide initial validation of the platform’s accessibility to diverse research users.

As with any LLM-based system, AI-HOPE-PM is susceptible to biases and potential hallucinations, particularly when interpreting complex or underspecified queries. To mitigate these risks, the system integrates domain-specific ontologies and harmonized variable dictionaries to reduce misinterpretation and support consistent query resolution. Additionally, the platform’s built-in clarification prompts serve as a real-time validation mechanism, prompting users to confirm or refine ambiguous instructions. While simulated SDOH-genomic interactions provide a useful testing framework, future efforts will emphasize empirical validation using real-world datasets to reduce confounding.

To address privacy concerns when working with sensitive real-world SDOH variables—such as insurance status, ethnicity, and income—AI-HOPE-PM is designed to operate as a secure, local AI system deployed within institutional infrastructures. Unlike cloud-based models that may transmit data externally, AI-HOPE-PM processes all data on-site, minimizing the risk of exposure or unauthorized access. This local deployment model supports compliance with data protection regulations, including the Health Insurance Portability and Accountability Act and the General Data Protection Regulation, where applicable. In future iterations, we plan to integrate customizable privacy modules and access controls to align with institutional review board protocols and ensure ethical handling of sensitive population-level health data.

To mitigate the risk of hallucinations and enhance the reliability of AI-HOPE-PM’s outputs, the platform incorporates several ethical safeguards. First, the system leverages domain-specific biomedical ontologies to align user queries with validated clinical and genomic concepts, reducing the likelihood of misinterpretation. Second, AI-HOPE-PM includes built-in prompts that clarify ambiguous user input, supporting more accurate query resolution. We also plan to implement human-in-the-loop verification workflows and bias detection modules, which will allow researchers to review, confirm, and flag generated outputs prior to downstream use. These strategies collectively enhance interpretability, accountability, and user trust in the AI-driven analytical process.

Given the sensitivity of SDOH data, especially variables such as insurance status, ethnicity, or socioeconomic conditions, AI-HOPE-PM is currently designed as a locally deployed system to prevent data exposure through external servers or third-party services. This architecture ensures that no identifiable information is shared beyond institutional firewalls. For future deployments that may involve real-world SDOH data, we plan to incorporate privacy-preserving methods including data deidentification, access controls, and secure computation protocols. All future iterations will comply with established data protection regulations such as the Health Insurance Portability and Accountability Act and the General Data Protection Regulation, ensuring responsible and ethical use of sensitive population-level data.

This study presents the development and application of AI-HOPE-PM, a conversational AI system designed to integrate clinical, genomic, and SDOH data for precision oncology research. AI-HOPE-PM addresses key limitations in existing bioinformatics tools by enabling users to pose natural language queries that are automatically translated into executable workflows. This allows for case-control stratification and hypothesis testing that include both molecular and nonmolecular variables.

In multiple CRC case studies, AI-HOPE-PM demonstrated the ability to reveal associations between genomic alterations (eg, TP53 and APC mutations), treatment exposures (eg, FOLFOX chemotherapy), and SDOH variables such as financial strain, food insecurity, health care access, and social support. These findings underscore the importance of contextualizing genomic data within broader socioeconomic and behavioral frameworks to better understand cancer disparities and inform population-relevant strategies.

Traditional tools such as cBioPortal and UCSC Xena have facilitated broad access to public cancer genomic datasets, yet they require manual, multistep filtering and operate within fixed analytical frameworks. These platforms typically lack support for SDOH integration and require a certain level of technical expertise, limiting their accessibility for noncomputational researchers and clinicians. More recent tools like CellAgent [17] and AutoBA [18] have begun to explore the use of LLMs in biomedical contexts, but their scope is generally limited to genomic analysis and does not extend to the integration of clinical or social variables essential for advancing health equity.

Our group’s prior work introduced AI-HOPE, a closed-system, LLM-driven conversational agent designed to enable integrative clinical and genomic data analyses through natural language interactions [23]. AI-HOPE allows users to perform association studies, prevalence assessments, and survival analyses on locally stored datasets while maintaining data security and interpretability. It demonstrated its capabilities by identifying well-documented associations in TCGA CRC datasets, such as the enrichment of TP53 mutations in late-stage CRC and the association of KRAS mutations with poor progression-free survival in FOLFOX-treated patients. While AI-HOPE addressed the integration of clinical and genomic data, it was not explicitly designed to handle population-level equity metrics or SDOH variables.

AI-HOPE-PM builds on and significantly extends this foundation by incorporating SDOH dimensions—such as financial strain, health care access, food insecurity, and health literacy—into its analytical framework. This addition allows researchers to study cancer outcomes in a more holistic context, bridging molecular findings with real-world social environments. Furthermore, AI-HOPE-PM expands the scope of natural language query handling to accommodate multimodal stratification involving genomic, clinical, and social parameters, which is essential for addressing health disparities. By doing so, it complements AI-HOPE’s functionality while introducing new capabilities that are critical for equity-focused translational research.

A key strength of AI-HOPE-PM is its ability to perform integrative, user-defined analyses through natural language queries without requiring programming expertise. This significantly reduces technical barriers for clinician-scientists and public health researchers. Importantly, the platform enables the inclusion of SDOH variables—such as financial strain, health care access, and social support—that are often absent from traditional bioinformatics workflows. Its modular architecture supports rapid cohort definition, survival modeling, and odds ratio testing across large, harmonized datasets, allowing for real-time hypothesis generation and exploratory analysis.

However, several limitations should be acknowledged. First, while this study used harmonized and simulated SDOH variables to demonstrate the platform’s functionality, the availability and quality of real-world, longitudinal SDOH data remain limited in many health care systems. This may affect the generalizability of findings and the real-world applicability of the platform. Future efforts will require integration with validated, longitudinal SDOH datasets—potentially through partnerships with clinical institutions and population health data repositories. Second, AI-HOPE-PM’s current design is optimized for structured, publicly available datasets such as TCGA, cBioPortal, and AACR GENIE. As such, its adaptability to unstructured clinical data or eHealth records is limited. While this design choice enhances reproducibility and alignment with standardized biomedical ontologies, future work should explore interoperability with clinical informatics platforms and natural language extraction from eHealth records to expand usability in health care settings. Third, this study focused exclusively on CRC datasets. As a result, findings and workflows may not be immediately generalizable to other cancer types without retraining or additional customization of the AI system. Although the architecture is designed to be adaptable, validation on other tumor types and disease areas will be essential for broader adoption. Fourth, while benchmarking analyses demonstrated strong performance compared to tools like cBioPortal and UCSC Xena, formal usability testing and prospective validation in real-world clinical and research environments were not conducted. These are planned as part of future development phases and will be critical for refining the user interface, evaluating human-AI collaboration, and assessing clinical impact. By acknowledging and addressing these limitations, future iterations of AI-HOPE-PM can be improved to better support equitable, scalable, and clinically relevant precision medicine research.

A notable limitation of the current study is the use of simulated SDOH variables rather than real-world data. While these simulated features were generated to reflect established patterns from peer-reviewed literature and public health datasets, they cannot fully replicate the variability, context-dependence, or missingness typical of empirical SDOH data collected in clinical or community settings. This limitation may impact the external validity of some findings and restrict generalizability. To address this, we are actively pursuing collaborations with health systems and community-based data partners to incorporate validated, longitudinal SDOH datasets into future deployments of AI-HOPE-PM. This planned integration will enable more accurate assessment of equity-relevant outcomes and enhance the platform’s application in real-world clinical research.

While AI-HOPE-PM achieved a high query interpretation accuracy of 92.5% during internal evaluation, several error modes were identified that merit consideration. The most frequent issues involved ambiguity in natural language input—particularly when users provided imprecise criteria for cohort selection or omitted critical parameters. Additionally, complex nested queries and nonstandard phrasing occasionally led to misinterpretation or partial execution. In a minority of cases, errors stemmed from misalignment between user terminology and the platform’s internal ontology, particularly for less common clinical or SDOH variables. To address these challenges, AI-HOPE-PM integrates clarification prompts that guide users toward more precise query formulation and supports synonym recognition for common variable names. Ongoing improvements include refining the language model’s domain specificity and expanding the internal ontology to better accommodate diverse user inputs. These enhancements are essential for improving reproducibility and user experience in real-world settings.

A key limitation of this study is the use of simulated SDOH variables rather than real-world data. While simulation allowed us to prototype and evaluate the functionality of AI-HOPE-PM under controlled conditions, it does not fully capture the complexity, heterogeneity, or potential missingness often present in real clinical and social datasets. To address this limitation, we have developed and released an open-source Python script [25] that transparently outlines our simulation methodology. Additionally, we are actively working on the integration of real-world SDOH data through ongoing projects [32], which is sequencing and characterizing tumors from 500 Hispanic/Latino patients in the Los Angeles catchment area. These datasets will allow us to test AI-HOPE-PM’s performance in real clinical environments and refine its capacity to analyze authentic, population-specific SDOH variables in future iterations.

To address this limitation, we acknowledge that the current evaluation of AI-HOPE-PM using 100 natural language queries—while carefully curated by physician-scientists, public health researchers, biostatisticians, and bioinformaticians to reflect real-world clinical and translational scenarios—represents an early validation phase. These queries were intentionally designed to ensure clinical accuracy, relevance, and internal consistency. However, we recognize the importance of expanding evaluation to include a broader and more diverse group of end users across different levels of expertise. Future iterations of AI-HOPE-PM will incorporate structured feedback from clinicians, public health researchers, and community health stakeholders. This participatory approach will help identify diverse interaction patterns, reduce potential biases, and enhance the platform’s interpretive capacity over time.

Future development of AI-HOPE-PM will focus on several enhancements. First, expanding support for additional omics layers [32], including spatial biology [33] and single-cell [34-36], could improve the platform’s applicability to emerging areas in systems oncology. Integration with federated learning frameworks may also enable secure, institution-specific model updates without compromising patient privacy. Moreover, enhancing the system’s ability to handle longitudinal data, including treatment timelines and SDOH trajectories, will be critical for supporting causal inference and policy-relevant research [37-42].

Future iterations of AI-HOPE-PM will prioritize the integration of more inclusive and representative genomic datasets to enhance the platform’s utility across diverse patient populations. While the current analyses rely on publicly available sources such as TCGA and cBioPortal—which are known to underrepresent racial and ethnic minorities—there have been encouraging advances in improving dataset diversity, particularly in CRC cohorts submitted by major US cancer centers. Notably, several ongoing initiatives aim to sequence and characterize tumors from historically underrepresented populations, including Hispanic/Latino patients with CRC [32]. These datasets, once publicly released, will be incorporated into AI-HOPE-PM to improve its generalizability and relevance in addressing cancer health disparities. This aligns with our overarching mission to develop equity-focused precision oncology tools that are responsive to the needs of all communities.

In this study, benchmarking primarily focused on task completion time—measuring the duration to execute standard bioinformatics queries across AI-HOPE-PM, cBioPortal, and UCSC Xena. While AI-HOPE-PM demonstrated superior efficiency due to its natural language automation, we acknowledge that this assessment does not encompass analytical output comparison. Future benchmarking studies will evaluate not only speed but also reproducibility and concordance of statistical results, including survival curves, odds ratios, and subgroup analyses. This expanded evaluation will ensure that AI-HOPE-PM delivers results comparable in accuracy and robustness to established platforms, further supporting its utility for translational cancer research.

A preliminary usability assessment was conducted during an internal pilot deployment involving five clinician-scientists and three public health researchers. Participants were asked to complete common clinical-genomic queries using AI-HOPE-PM and provide structured feedback on system usability, interpretability of outputs, and ease of query formulation. Feedback indicated that users found the natural language interface intuitive and appreciated the automation of statistical analyses without coding. Suggestions for improvement included refining terminology prompts and expanding visualization customization. These insights have been incorporated into the current version of AI-HOPE-PM, and a formal usability study with a larger and more diverse cohort is currently underway to systematically evaluate accessibility, performance, and user satisfaction.

To enhance accessibility and promote broader adoption, particularly in resource-constrained environments, we are actively exploring deployment strategies that reduce local infrastructure requirements. Although the current AI-HOPE-PM system benefits from graphics processing unit acceleration for large-scale genomic analyses, the core functionalities—including query interpretation, basic statistical modeling, and report generation—can be executed on standard central processing unit-based systems. Additionally, we are developing a lightweight web-hosted version of the platform with backend support on scalable cloud infrastructure, enabling institutions with limited computational resources to access AI-HOPE-PM through a browser without the need for specialized hardware. Future iterations will also offer modular processing options that allow users to select compute-intensive features based on available resources.

User-centered evaluations—including usability studies with diverse researchers and clinicians—are planned to better understand the platform’s accessibility and impact in real-world settings. Additionally, collaborations with community-based research initiatives may help validate AI-HOPE-PM’s role in addressing health disparities and improving equity in precision medicine.

AI-HOPE-PM was developed with scalability and accessibility in mind, including potential deployment in resource-constrained settings. The system can be installed and executed locally, eliminating the need for high-bandwidth internet or continuous cloud access. While graphics processing unit acceleration can enhance performance for large-scale queries, the platform’s modular backend supports central processing units–only configurations for smaller datasets and standard analyses. Ongoing optimization efforts aim to further reduce computational overhead through lightweight LLM variants and model compression techniques. These features support broader adoption across diverse institutional environments, including low-resource clinical and research settings.

A key consideration for the broader adoption of AI-HOPE-PM is the potential for language bias and variability in natural language queries. While the current version of the platform is optimized for English-language input, this may limit accessibility for nonnative English speakers or introduce semantic variability that could affect interpretation. To mitigate this, AI-HOPE-PM employs a domain-specific ontology with synonym recognition and structured clarification prompts that guide users toward standardized, interpretable input. These features reduce the likelihood of misinterpretation and increase the reliability of query processing. Nonetheless, we recognize the importance of supporting diverse linguistic backgrounds in biomedical research. Future iterations of the platform will integrate multilingual capabilities and undergo structured usability evaluations in non–English-speaking populations to ensure equitable utility and minimize language-related inequities in research engagement.