AI-HOPE-WNT (Artificial Intelligence agent for High-Optimization and Precision Medicine focused on WNT) is a conversational AI platform purpose-built for investigating CRC through the lens of WNT pathway dysregulation. It utilizes a modular architecture composed of LLMs, a natural language-to-code translation engine, and a backend analytics pipeline that enables real-time case–control analysis and hypothesis generation. Upon user input via natural language, the system performs automated query parsing, data retrieval, statistical testing, and visualization (Figure 1). Output includes Kaplan–Meier survival curves, odds ratio metrics, frequency distributions, and narrative summaries contextualized by biomedical literature.

The AI-HOPE-WNT platform follows a structured, end-to-end workflow that transforms natural language queries into rigorous clinical-genomic analyses. Users begin by entering a plain English question—for example, “Compare survival in EOCRC patients with APC mutations versus wild-type receiving FOLFOX treatment.” This query is interpreted by a fine-tuned large language model (LLaMA 3 variant), which converts the input into executable code using a natural language-to-code engine. The system then identifies and subsets relevant patient cohorts from harmonized cBioPortal datasets based on parameters such as mutation status, age, ethnicity, tumor location, MSI status, and treatment exposure. Once cohorts are defined, AI-HOPE-WNT performs statistical analyses appropriate to the query context, including odds ratio calculations, Kaplan–Meier survival estimations, and Cox proportional hazards modeling. Results are returned as structured outputs—such as tables, survival curves, and forest plots—alongside automated narrative summaries contextualized by biomedical literature. Finally, the platform generates high-resolution visualizations and structured reports using Matplotlib and Plotly, enabling reproducible, publication-ready outputs. This integrated workflow enhances usability, transparency, and analytical rigor, establishing AI-HOPE-WNT as a robust tool for pathway-specific precision oncology research.

The core language processing component of AI-HOPE-WNT is built on a variant of the LLaMA 3 large language model, selected for its balance of efficiency, scalability, and open architecture. This LLM was fine-tuned with domain-specific prompts and biomedical context to improve performance in interpreting clinical-genomic queries. The model operates within a retrieval-augmented generation (RAG) framework, which allows it to cross-reference structured biomedical knowledge sources during natural language-to-code translation. Prompt engineering techniques were applied to align model outputs with statistical and bioinformatics standards, reduce hallucination risk, and preserve analytical accuracy. Additionally, structured prompting ensures reproducibility, while the modular design allows for future upgrades or substitution with newer LLMs, ensuring continued adaptability as foundational AI models evolve.

To evaluate cohort differences and mutation enrichment in AI-HOPE-WNT, we employed several standard statistical metrics. For categorical comparisons (e.g., mutation prevalence by subgroup), we used odds ratios (ORs), mathematically defined as:

where a and b represent the number of individuals with and without the mutation in group 1, and c and d represent the same in group 2. Confidence intervals (95%) for ORs were computed using the Woolf method.

For time-to-event data, we used Kaplan–Meier survival estimates, with group differences assessed using the log-rank test, where the test statistic follows a chi-squared distribution:

with O₁ as observed events, E₁ as expected events, and V₁ as the variance. For multivariable survival analysis, we used Cox proportional hazards models, defined as:

where h(t∣X) is the hazard function at time t given covariates X, and β values are estimated log hazard ratios. Model significance was determined via likelihood ratio tests and Wald statistics.

These evaluation metrics are implemented programmatically within AI-HOPE-WNT’s statistical backend to ensure consistency, interpretability, and reproducibility across analyses.

AI-HOPE-WNT interfaces with harmonized CRC datasets curated from cBioPortal and The Cancer Genome Atlas (TCGA), with emphasis on WNT-relevant gene mutations (e.g., APC, CTNNB1, RNF43, AXIN1, AXIN2). Clinical variables such as age, gender, tumor location, stage, microsatellite instability (MSI) status, and treatment regimen (e.g., FOLFOX) were included. Genomic annotations and mutation profiles were preprocessed into standardized, tab-delimited formats with unified sample IDs. Metadata were harmonized using established ontologies (e.g., OncoTree, Disease Ontology) to ensure interoperability. Data validation included cross-referencing gene mutation calls, consistency checks, and enrichment annotations for WNT-associated genes.

Users interact with the platform via plain English queries that are interpreted by an LLM (LLaMA 3) and translated into structured data operations. AI-HOPE-WNT allows for flexible cohort definitions, enabling stratification by mutation status, clinical features, treatment exposure, or demographic characteristics. For example, users can query “Compare survival between EOCRC patients with APC mutations vs. wild-type in FOLFOX-treated tumors” or “Identify co-occurrence between AXIN2 and RNF43 in rectal tumors among Hispanic/Latino patients.” The system prompts clarifications when inputs are ambiguous, enhancing transparency and preventing spurious analyses.

Analyses are conducted via Python-based pipelines integrated into the backend. For categorical variables, the system computes frequency distributions, chi-square or Fisher’s exact tests, and odds ratios with 95% confidence intervals. Time-to-event data are analyzed using Kaplan–Meier survival estimation and log-rank tests; Cox proportional hazards regression is applied for multivariate analyses. WNT-focused genes were prioritized for mutation enrichment, co-mutation, and prognostic modeling. The platform supports subgroup exploration by age (<50 vs. ≥50), ethnicity (Hispanic/Latino vs. non-Hispanic White), and tumor location (colon vs. rectal).

AI-HOPE-WNT integrates a retrieval-augmented generation (RAG) engine that references a structured biomedical knowledge base to support contextual accuracy and limit hallucinations. Structured prompting ensures adherence to bioinformatics conventions and enables reproducibility across analytical outputs. To validate platform performance, we replicated key findings from our previously published studies (Song et al., 2023; Youssef et al., 2025; Cerami et al., 2012), including RNF43 and AXIN2 mutation differences in EOCRC H/L and survival trends associated with WNT pathway alterations.

To benchmark the performance of AI-HOPE-WNT, we conducted a task-based comparison with two widely used platforms—cBioPortal and UCSC Xena—which served as baseline tools. Performance was evaluated across three core dimensions: (1) execution time, measured from query input to analysis output; (2) workflow complexity, defined by the number of steps or interface interactions required to complete a query; and (3) subgroup analysis flexibility, assessed by the ability to stratify cohorts across multiple demographic and clinical variables (e.g., age, ethnicity, treatment, mutation status). For example, a survival analysis stratified by APC mutation status and FOLFOX treatment in early-onset Hispanic/Latino patients required only a single query in AI-HOPE-WNT, compared to 6–8 manual steps in cBioPortal and Xena. Quantitatively, AI-HOPE-WNT reduced average task completion time by 72% and enabled nested subgroup comparisons that were not natively supported in baseline tools. These results underscore the platform’s superior efficiency, usability, and analytic scope—particularly in real-time, natural language-driven settings focused on pathway-specific precision oncology.

Post-analysis, the platform generates structured reports including summary statistics, mutation frequency tables, survival plots, and forest plots. Narrative interpretations are auto-generated, linking statistical findings to literature-derived insights. Visualizations were rendered using Matplotlib and Plotly within the backend, ensuring high-resolution graphics suitable for scientific reporting.

AI-HOPE-WNT was developed to transform natural language prompts into automated clinical-genomic analyses, allowing real-time exploration of WNT signaling dysregulation in CRC. The system integrates structured datasets with natural language queries, enabling users to dynamically stratify patient cohorts based on tumor location, age, sex, mutation status, microsatellite instability (MSI), and treatment exposure. In response to these queries, the platform executes statistical procedures such as Kaplan–Meier survival analysis, odds ratio testing, and mutation frequency comparisons, and presents the results through high-resolution visualizations and narrative summaries. This functionality allows non-coding users—including clinicians and translational researchers—to interactively generate and test hypotheses in an intuitive and reproducible manner. The results presented here demonstrate AI-HOPE-WNT’s capacity to both replicate previously published findings and uncover novel associations across EOCRC populations.

To validate the analytical robustness of the platform, AI-HOPE-WNT was applied to replicate previously reported associations between WNT pathway mutations and clinical outcomes in EOCRC patients. Using survival modeling, the system reproduced findings from a published CRC WNT study (Monge et al., 2024), comparing outcomes among EOCRC patients with and without WNT pathway mutations. In Hispanic/Latino (H/L) patients, WNT-altered tumors were associated with significantly improved survival (p = 0.0167), and a similar trend was observed in non-Hispanic White (NHW) patients, where survival differences were even more pronounced (p = 0.0007) (Figure 2). These findings confirm prior evidence that WNT pathway activation, in certain molecular and demographic contexts, may have prognostic relevance in EOCRC.

Additionally, AI-HOPE-WNT reproduced previously reported trends in ancestry-specific mutation enrichment. Odds ratio testing showed that RNF43 mutations were more prevalent in H/L patients compared to NHW, with an odds ratio of 1.31 (Supplementary Figure S1). Although the platform’s dataset included a smaller number of H/L samples than the original study, the observed enrichment supports the reproducibility and ancestry-awareness of the platform’s analytical workflow. Together, these validations highlight the platform’s capability to replicate established WNT pathway associations through fully automated, natural language–driven analysis.

Exploratory analysis using AI-HOPE-WNT revealed a significant survival advantage among EOCRC patients treated with FOLFOX who harbored mutations in the APC gene. In patients under the age of 50, APC mutations were associated with significantly improved survival compared to wild-type cases (p = 0.0027) (Figure 3). These findings suggest that APC mutation status may influence treatment response to FOLFOX in younger CRC patients and highlight the potential utility of pathway-specific AI agents for evaluating molecular markers within therapeutic contexts.

The platform was also used to explore survival outcomes based on tumor stage among RNF43-mutated CRC patients. AI-HOPE-WNT identified a significant survival disadvantage in patients with Stage IV (metastatic) tumors compared to those with Stage I–III (non-metastatic) disease (p < 0.0001) (Figure 4). This result underscores the prognostic importance of tumor stage in RNF43-mutant CRC and demonstrates how AI-HOPE-WNT can uncover nuanced, mutation-specific insights through stratified cohort comparisons.

Using natural language queries, AI-HOPE-WNT investigated co-mutation patterns between AXIN1 and APC by tumor location. Odds ratio testing revealed differential enrichment of APC co-mutations in colon versus rectal tumors. However, Kaplan–Meier survival analysis showed no statistically significant survival differences between these anatomical subgroups (p = 0.2216) (Figure 5). These results suggest that while co-mutation patterns may vary by location, their prognostic impact may be limited—or require larger datasets for confirmation—highlighting the platform’s potential for guiding further inquiry.

Sex-based survival differences were evaluated among AXIN2-mutated patients, stratified by MSI status. A trend toward worse progression-free survival was observed in male patients relative to females, though the difference did not reach statistical significance (p = 0.1274). Nevertheless, odds ratio analysis revealed MSI-stable enrichment in both male and female subgroups (Supplementary Figure S2), indicating potential interactions between gender, microsatellite status, and WNT signaling. These findings illustrate how AI-HOPE-WNT can facilitate exploration of under-investigated biological interactions using clinical-genomic data.

AI-HOPE-WNT was used to assess the relationship between APC mutation prevalence and age in primary CRC tumors. Among 3,396 APC-mutant and 1,252 wild-type cases, odds ratio testing showed that APC mutations were significantly more prevalent in EOCRC patients (<50 years). Further, Kaplan–Meier survival analysis demonstrated significantly improved outcomes in APC-mutated tumors (p = 0.00001) (Supplementary Figure S3). This supports the prognostic relevance of APC mutations in younger CRC populations and illustrates how AI-HOPE-WNT can stratify patients for biomarker discovery.

Collectively, these results showcase the analytical range and depth of AI-HOPE-WNT. The platform successfully replicated prior WNT-related findings while generating new hypotheses linking gene alterations to survival outcomes across treatment types, tumor stages, demographics, and molecular subgroups. By enabling real-time, natural language–driven analysis of harmonized clinical-genomic data, AI-HOPE-WNT offers a novel, scalable approach to translational discovery in colorectal cancer. Its ability to handle complex stratification and return visual and statistical outputs makes it a powerful tool for precision oncology research, education, and clinical hypothesis testing.