A pilot evaluation of ChatGPT-4o-assisted hypothesis generation for exploring potential links between fungal biology and cervical cancer

Abstract

The potential relationship between fungal biology and cervical cancer remains poorly understood. This study aimed to assess the ability of ChatGPT-4o to generate research hypotheses concerning possible links between fungal factors, HPV persistence, and cervical carcinogenesis. ChatGPT-4o generated hypotheses in three prespecified domains: microbial dysbiosis, inflammation, and immune modulation. Three author-evaluators with expertise in pathology, oncology, and microbiology independently assigned overall evaluation scores and assessed the scientific reasonableness of experimental-plan components. Targeted literature searches identified publications directly addressing each proposed mechanism. Inter-rater reliability was quantified using a two-way random-effects, absolute-agreement intraclass correlation coefficient (ICC) and pairwise quadratic-weighted Cohen's kappa. No human-generated hypothesis set, alternative language model, or repeated independent AI session was included. ChatGPT-4o generated 55 hypotheses. No directly matching publication was identified for 48 hypotheses under the prespecified search framework; this finding was interpreted as indicating an underexplored direction, not proof of novelty. Based on the predefined composite evaluation rubric, 35 hypotheses (63.6%) received high mean evaluation scores and 20 (36.4%) received moderate mean evaluation scores. Agreement for hypothesis scores was high (ICC (2,1) = 0.975; pairwise weighted kappa = 0.963–1.000). Among 18 experimental-plan components, 12 received high mean scores and 6 received moderate mean scores; agreement was lower but still strong (ICC (2,1) = 0.821; weighted kappa = 0.727–0.914). Representative hypotheses concerned fungal metabolites and oxidative stress, fungal-associated inflammation and oxidative damage, and fungal modulation of immune-checkpoint pathways. This pilot study suggests that ChatGPT-4o can generate diverse and potentially testable hypotheses concerning fungal biology and cervical cancer. However, the generated hypotheses should be regarded as exploratory research directions rather than validated mechanisms and require independent literature verification and experimental validation.

Supplementary Information

The online version contains supplementary material available at https://doi.org/10.1007/s10482-026-02434-3.

Introduction

Cervical cancer remains a major global health concern and is primarily caused by persistent infection with high-risk human papillomavirus (HPV), particularly HPV types 16 and 18, which account for a substantial proportion of cervical cancer cases (Alfaifi 2024; Hampson and Oliver 2024; Thrall et al. 2025; Zhao et al. 2024). Although HPV infection is common and is frequently cleared spontaneously by the host immune system, persistent infection with oncogenic HPV types can induce cervical epithelial abnormalities that may progress from precancerous lesions to invasive cervical cancer (Hampson and Oliver 2024). Multiple host and environmental factors, including smoking, immunosuppression, prolonged oral contraceptive use, and reproductive history, may influence disease progression (Bovo et al. 2023; Bowden et al. 2023). Therefore, understanding the factors that determine HPV persistence and the transition from infection to cervical neoplasia remains an important research priority. Despite advances in HPV vaccination and screening strategies, the biological mechanisms regulating HPV persistence are not completely understood (Gradíssimo and Burk 2017; Perkins et al. 2023).

The cervicovaginal microbiome has emerged as an important factor potentially influencing HPV infection outcomes. This microbial ecosystem consists of complex interactions among microorganisms, epithelial cells, metabolites, and host immune responses. A Lactobacillus-dominated microbiota, particularly communities enriched in Lactobacillus crispatus, is generally associated with vaginal microbial stability, whereas microbial dysbiosis and increased community diversity have been linked to HPV acquisition, persistence, cervical intraepithelial neoplasia (CIN), and cervical cancer (Hu et al. 2024a, b; Kyrgiou and Moscicki 2022; Lebeau et al. 2022; Zhang et al. 2024).These findings suggest that alterations in the cervicovaginal microbial environment may influence HPV persistence through mechanisms involving mucosal homeostasis, inflammation, and host–microbe interactions. However, current investigations have predominantly focused on bacterial communities, while other microbial components remain comparatively understudied.

Fungi represent an important but poorly characterized component of the cervicovaginal microbiome. The vaginal mycobiome comprises diverse fungal taxa, with Candida species being the most extensively studied due to their role in vulvovaginal candidiasis. However, fungal communities may exist beyond clinically apparent infection and may contribute to the ecological balance of the reproductive tract (Cassalia et al. 2024; Iyadorai et al. 2024; Mirahmad et al. 2024). Recent studies have begun to investigate the association between vaginal fungal communities, HPV infection, and cervical disease, suggesting that fungal composition may differ according to HPV status and disease progression and may interact with bacterial members of the cervicovaginal microbiota (Alimena et al. 2022; Godoy-Vitorino et al. 2026; Passarelli et al. 2026). Nevertheless, the biological significance of these fungal alterations remains largely unknown, and whether specific fungal taxa or fungal-associated pathways contribute to HPV persistence requires further investigation.

In other cancer contexts, fungi and fungal-derived products have been implicated in processes relevant to carcinogenesis, including inflammation, cellular signaling, genomic instability, and immune regulation (Cao et al. 2022; Magnussen and Parsi 2013; Wokorach et al. 2022). However, such evidence should not be directly extrapolated to cervicovaginal fungal colonization, because systemic exposure to fungal toxins and local fungal–host interactions represent biologically distinct processes. Therefore, the potential relationship between the cervicovaginal mycobiome, HPV persistence, and cervical disease remains an important but insufficiently explored research question.

The complexity of microbial ecosystems and host–microbe interactions present challenges for conventional hypothesis-driven research. Artificial intelligence (AI), particularly large language models (LLMs), provides a potential approach for integrating heterogeneous biological information and generating research hypotheses from dispersed scientific knowledge (Esteva et al. 2017; Sarode et al. 2022; Yang et al. 2024). AI-based approaches have been increasingly applied in microbiology and cancer research, including microbial pattern recognition, biomarker discovery, and analysis of complex biological interactions (Cao et al. 2024; Fang et al. 2023; Rappoport et al. 2024; Singh et al. 2024). Our previous study further suggested that AI-assisted approaches may facilitate hypothesis generation by connecting existing biomedical evidence and identifying potential research directions (Yao et al. 2025). Such an approach may be particularly valuable for exploring the potential role of the cervicovaginal mycobiome in HPV-related cervical disease, where available evidence remains fragmented.

Therefore, this study aimed to investigate whether ChatGPT could serve as a hypothesis-generation tool for exploring potential relationships between fungal communities, the cervicovaginal microbial ecosystem, HPV infection, and cervical disease. This AI-assisted framework aimed to identify biologically plausible associations, candidate fungal taxa, and testable hypotheses regarding potential interactions between fungal communities, the cervicovaginal microbiome, and HPV-associated cervical disease, thereby providing directions for future experimental validation.

Results

AI-assisted generation and author-evaluator assessment of hypotheses linking fungal communities with HPV-associated cervical disease

ChatGPT generated 55 hypotheses concerning possible relationships between fungi and HPV-associated cervical disease. These included 15 hypotheses related to microbial dysbiosis, 20 related to inflammation, and 20 related to immune modulation (refer Tables 1, 2 and 3; Supplementary Materials 2–4). Following hypothesis generation, ChatGPT was prompted to select one hypothesis from each domain and generate a corresponding experimental plan. The selected hypotheses and experimental plans were subsequently evaluated by three author-evaluators who performed their assessments independently.

Table 1.

ChatGPT and author-evaluator score of hypotheses related to microbial dysbiosis

No.	Hypothesis	ChatGPT overall score	Keywords	Publications identified	E1	E2	E3	Group
1	Chronic vaginal fungal infections lead to a persistent pro-inflammatory state that facilitates HPV persistence and cervical cancer	4—High	Chronic inflammation, HPV persistence, fungal infection, cervical cancer	None identified	4	4	4	4.00
2	Specific fungal species within the vaginal microbiome disrupt epithelial integrity, enhancing HPV integration and promoting cancer	4—High	Fungal species, epithelial integrity, HPV integration, carcinogenesis	None identified	4	5	4	4.33
3	Fungal dysbiosis reduces Lactobacillus populations, creating conditions that favor HPV persistence and cervical cancer	3—Moderate	Fungal dysbiosis, Lactobacillus, HPV persistence, microbiome	6 publications	3	3	3	3.00
4	Fungal biofilms in the cervicovaginal environment protect HPV particles from immune detection, increasing cancer risk	4—High	Fungal biofilms, immune evasion, HPV persistence, cervical cancer	None identified	3	3	3	3.00
5	The interaction between specific fungi and HPV alters gene expression in cervical cells, promoting oncogenic transformation	4—High	Fungal-HPV interaction, gene expression, oncogenic transformation	None identified	4	4	4	4.00
6	Long-term antifungal treatment disrupts the microbiome, leading to bacterial overgrowth that enhances HPV-mediated carcinogenesis	3—Moderate	Antifungal treatment, bacterial overgrowth, HPV, cervical cancer	None identified	3	3	3	3.00
7	Fungi modulate cytokine and chemokine production, enhancing HPV-infected cell oncogenic potential	3—Moderate	Fungal modulation, cytokines, chemokines, HPV oncogenesis	None identified	4	4	4	4.00
8	Fungal metabolites induce oxidative stress in HPV-infected cells, accelerating cervical cancer progression	4—High	Fungal metabolites, oxidative stress, HPV, cervical cancer	None identified	4	4	4	4.00
9	Fungal colonization interferes with cervical epithelial desquamation, increasing HPV persistence and cancer risk	4—High	Fungal colonization, epithelial desquamation, HPV persistence	None identified	4	4	4	4.00
10	Vaginal mycobiome differences predispose women to higher HPV persistence and cervical cancer risk	3—Moderate	Vaginal mycobiome, HPV persistence, cervical cancer	None identified	3	3	3	3.00
11	Fungal proteases degrade immune factors in the cervix, reducing defense against HPV and promoting cancer	4—High	Fungal proteases, immune degradation, HPV, cervical cancer	None identified	3	3	3	3.00
12	Fungal presence influences HPV DNA methylation, affecting viral gene expression and cancer risk	4—High	Fungal influence, DNA methylation, HPV gene expression	None identified	4	4	4	4.00
13	Fungal infections induce epithelial-mesenchymal transition (EMT) in HPV-infected cells, driving metastasis	4—High	Fungal infections, EMT, HPV, metastasis	None identified	4	4	4	4.00
14	Antifungal-resistant fungi exert selective pressure that enhances HPV oncogenic potential	3—Moderate	Antifungal resistance, selective pressure, HPV oncogenesis	None identified	3	3	3	3.00
15	Co-infection with fungi and HPV creates microbial niches that enhance viral persistence and carcinogenesis	4—High	Co-infection, microbial niches, HPV persistence, cervical cancer	None identified	4	4	4	4.00

E1, E2, and E3 denote Evaluators 1–3 (pathology, oncology, and microbiology, respectively). Group is the arithmetic mean of E1–E3, reported to two decimal places. 'None identified' means that no directly relevant publication was retrieved under the predefined strategy

Table 2.

ChatGPT and author-evaluator score of hypotheses related to inflammation

No.	Hypothesis	ChatGPT overall score	Keywords	Publications identified	E1	E2	E3	Group
1	Chronic vaginal fungal infections trigger a sustained inflammatory response that disrupts immune surveillance, allowing persistent HPV infection and promoting cervical cancer	4—High	Chronic inflammation, immune surveillance, HPV persistence, cervical cancer	None identified	4	4	4	4.00
2	Fungal-induced production of pro-inflammatory cytokines and chemokines in the cervix enhances HPV-infected cell proliferation, increasing cancer risk	4—High	Pro-inflammatory cytokines, chemokines, cell proliferation, HPV, cervical cancer	None identified	4	4	4	4.00
3	Specific fungal species in the vaginal microbiome modulate inflammation, accelerating the transition of HPV-infected cells to high-grade CIN and cancer	3—Moderate	Fungal species, inflammation, CIN, HPV, cervical cancer	Nicolò et al. 2023	4	4	4	4.00
4	Fungal infections cause chronic inflammation that upregulates HPV oncogenes (E6 and E7), enhancing cervical cancer development	4—High	Chronic inflammation, HPV oncogenes, cervical cancer	29 publications	4	4	4	4.00
5	Inflammatory mediators released in response to fungal colonization disrupt epithelial barrier function, creating a microenvironment conducive to HPV persistence and cervical cancer	4—High	Inflammatory mediators, epithelial barrier, HPV persistence, cervical cancer	None identified	4	4	4	4.00
6	Fungal biofilms in the cervix trigger localized inflammation that impairs the clearance of HPV-infected cells, increasing cancer risk	4—High	Fungal biofilms, localized inflammation, HPV clearance, cervical cancer	None identified	3	3	3	3.00
7	Fungal-derived pro-inflammatory molecules interact with HPV pathways, synergistically enhancing inflammation and accelerating cervical cancer	4—High	Pro-inflammatory molecules, HPV pathways, synergistic inflammation, cervical cancer	None identified	4	4	4	4.00
8	Recurrent fungal infections lead to chronic inflammation in the cervix that promotes angiogenesis, supporting HPV-transformed cell growth and invasion	3—Moderate	Recurrent infections, chronic inflammation, angiogenesis, HPV, cervical cancer	Gupta et al. 2018	3	3	3	3.00
9	Chronic inflammation from fungal infections activates NF-kB signaling, upregulating survival genes and promoting cervical cancer	4—High	Chronic inflammation, NF-kB signaling, survival genes, cervical cancer	None identified	4	4	4	4.00
10	Fungal-induced inflammation increases ROS and NO production in the cervix, causing DNA damage in HPV-infected cells and enhancing carcinogenesis	4—High	Inflammation, ROS, NO, DNA damage, HPV, cervical cancer	Ohnishi et al. 2013	4	4	4	4.00
11	Fungal infections cause inflammation that releases MMPs, degrading extracellular matrix and facilitating HPV-infected cell invasion into cervical cancer	4—High	Inflammation, MMPs, extracellular matrix, cell invasion, HPV, cervical cancer	Libra et al. 2009	4	4	4	4.00
12	Fungal-induced inflammation alters cytokine profiles, skewing immune response towards a Th2 profile, promoting HPV persistence and cervical cancer	3—Moderate	Inflammation, cytokine profiles, Th2, HPV persistence, cervical cancer	None identified	4	4	4	4.00
13	Chronic inflammation from fungal infections induces epigenetic modifications in host genes, creating a pro-tumorigenic environment in HPV-infected cells	4—High	Chronic inflammation, epigenetic modifications, pro-tumorigenic environment, HPV, cervical cancer	None identified	3	3	3	3.00
14	Fungal infections stimulate the release of pro-inflammatory lipid mediators, enhancing HPV oncogene expression and driving cervical cancer	4—High	Inflammation, lipid mediators, HPV oncogene expression, cervical cancer	Tan et al. 2016	4	4	4	4.00
15	Fungal components interact with TLRs on cervical cells, triggering chronic inflammation that upregulates HPV oncogenes and promotes carcinogenesis	4—High	Fungal components, TLRs, chronic inflammation, HPV oncogenes, cervical cancer	None identified	4	4	4	4.00
16	Fungal-induced inflammation disrupts cervical microenvironment homeostasis, activating STAT3 pathways and driving HPV-associated cancer	4—High	Inflammation, microenvironment disruption, STAT3, HPV-associated cancer	None identified	4	4	4	4.00
17	Chronic inflammation from fungal antigens promotes Treg expansion, suppressing immune response against HPV and increasing cervical cancer risk	3—Moderate	Chronic inflammation, Tregs, immune suppression, HPV, cervical cancer	None identified	4	4	4	4.00
18	Fungal infections induce chronic inflammation that leads to adhesion molecule expression, facilitating the retention of HPV-infected cells and promoting cancer	4—High	Chronic inflammation, adhesion molecules, HPV retention, cervical cancer	None identified	3	3	3	3.00
19	Fungal-induced inflammation promotes the release of extracellular vesicles (EVs) containing HPV oncogenes, enhancing cell invasiveness and cancer progression	4—High	Inflammation, extracellular vesicles, HPV oncogenes, cancer progression	None identified	3	3	3	3.00
20	Chronic inflammation from fungal infections leads to fibrosis and tissue remodeling, creating an environment supportive of HPV-driven cervical cancer	4—High	Chronic inflammation, fibrosis, tissue remodeling, HPV-driven cancer	None identified	3	3	3	3.00

E1, E2, and E3 denote Evaluators 1–3 (pathology, oncology, and microbiology, respectively). Group is the arithmetic mean of E1–E3, reported to two decimal places. 'None identified' means that no directly relevant publication was retrieved under the predefined strategy

Table 3.

ChatGPT and author-evaluator score of hypotheses related to immune modulation

No.	Hypothesis	ChatGPT overall score	Keywords	Publications identified	E1	E2	E3	Group
1	Fungal infections in the cervix alter the local immune landscape by suppressing antigen-presenting cells (APCs), reducing the immune system's ability to recognize and clear HPV-infected cells, thereby promoting cervical cancer progression	4—High	Fungal infections, APC suppression, immune evasion, HPV, cervical cancer	None identified	3	3	3	3.00
2	Fungal components mimic HPV antigens, leading to an immune tolerance that diminishes the immune response to HPV-infected cells and increases the risk of cervical cancer	4—High	Fungal components, HPV antigen mimicry, immune tolerance, cervical cancer	None identified	3	3	3	3.00
3	Chronic fungal infections in the cervix drive the expansion of regulatory T cells (Tregs), which suppress cytotoxic T cell activity against HPV-infected cells, facilitating the development of cervical cancer	4—High	Chronic fungal infections, Tregs, immune suppression, HPV, cervical cancer	None identified	4	4	4	4.00
4	Fungal-induced immune modulation shifts the balance of Th1/Th2 responses in the cervix, leading to a Th2-dominant environment that supports the persistence of HPV and promotes cervical carcinogenesis	3—Moderate	Fungal-induced modulation, Th1/Th2 balance, immune environment, HPV persistence, cervical cancer	None identified	4	4	4	4.00
5	Fungal infections lead to the production of immune-modulatory cytokines, such as IL-10 and TGF-β, which inhibit the anti-tumor immune response, allowing HPV-infected cells to evade immune surveillance and progress to cervical cancer	4—High	Fungal infections, cytokines, IL-10, TGF-β, immune evasion, HPV, cervical cancer	None identified	4	4	4	4.00
6	Fungal biofilms in the cervix create an immune-privileged niche, shielding HPV-infected cells from immune detection and contributing to the persistence of infection and eventual cervical cancer development	4—High	Fungal biofilms, immune privilege, HPV persistence, cervical cancer	None identified	3	3	3	3.00
7	Fungal pathogens alter the expression of immune checkpoint molecules (e.g., PD-L1) on cervical epithelial cells, reducing the effectiveness of the immune response against HPV and increasing the risk of cervical cancer	4—High	Fungal pathogens, immune checkpoint, PD-L1, immune response, HPV, cervical cancer	None identified	4	4	4	4.00
8	Chronic fungal colonization in the cervix leads to the recruitment of myeloid-derived suppressor cells (MDSCs), which suppress anti-HPV immune responses and promote the progression of HPV-associated cervical lesions to cancer	4—High	Chronic fungal colonization, MDSCs, immune suppression, HPV, cervical cancer	None identified	4	4	4	4.00
9	Fungal infections in the cervix trigger the production of extracellular vesicles (EVs) that carry immune-suppressive molecules, which inhibit the activation of HPV-specific T cells and facilitate cervical cancer progression	4—High	Fungal infections, extracellular vesicles, immune suppression, HPV-specific T cells, cervical cancer	None identified	4	4	4	4.00
10	Fungal infections in the cervix induce a state of chronic low-grade inflammation that exhausts local immune cells, leading to an impaired immune response to HPV and increased susceptibility to cervical cancer	4—High	Fungal infections, chronic low-grade inflammation, immune cell exhaustion, HPV, cervical cancer	None identified	3	3	3	3.00
11	Fungal-induced modulation of the local microbiome alters the production of short-chain fatty acids (SCFAs), which in turn influence the immune response to HPV and contribute to cervical carcinogenesis	3—Moderate	Fungal-induced modulation, microbiome, SCFAs, immune response, HPV, cervical cancer	None identified	3	3	3	3.00
12	Fungal pathogens manipulate the local immune response by upregulating inhibitory receptors on natural killer (NK) cells, reducing their ability to target and destroy HPV-infected cells, thereby increasing the risk of cervical cancer	4—High	Fungal pathogens, NK cells, inhibitory receptors, immune response, HPV, cervical cancer	None identified	4	4	4	4.00
13	Fungal metabolites alter the cytokine profile in the cervical microenvironment, promoting an immunosuppressive state that facilitates the persistence of HPV and progression to cervical cancer	4—High	Fungal metabolites, cytokine profile, immunosuppression, HPV persistence, cervical cancer	None identified	4	4	4	4.00
14	Chronic fungal infections in the cervix lead to the secretion of immunomodulatory fungal proteins that interfere with dendritic cell maturation, impairing the initiation of effective HPV-specific immune responses and promoting cervical cancer	4—High	Chronic fungal infections, immunomodulatory proteins, dendritic cell maturation, HPV, cervical cancer	None identified	3	3	3	3.00
15	Fungal infections cause the activation of pattern recognition receptors (PRRs) on cervical epithelial cells, leading to an immune response that paradoxically promotes HPV persistence and carcinogenesis	4—High	Fungal infections, PRRs, immune response, HPV persistence, cervical cancer	None identified	4	4	4	4.00
16	Fungal pathogens enhance the secretion of pro-tumorigenic cytokines by cervical epithelial cells, leading to a local immune environment that supports the survival and growth of HPV-infected cells and accelerates cervical cancer development	4—High	Fungal pathogens, pro-tumorigenic cytokines, immune environment, HPV, cervical cancer	None identified	4	4	4	4.00
17	Fungal-induced disruption of the mucosal barrier in the cervix allows for increased HPV infection and persistence, while simultaneously modulating the local immune response to favor carcinogenesis	4—High	Fungal-induced disruption, mucosal barrier, immune response modulation, HPV persistence, cervical cancer	None identified	4	4	4	4.00
18	Fungal infections in the cervix modulate the local production of chemokines, leading to altered immune cell trafficking that favors the persistence of HPV and the development of cervical cancer	4—High	Fungal infections, chemokines, immune cell trafficking, HPV persistence, cervical cancer	None identified	3	3	3	3.00
19	Fungal pathogens in the cervix induce the secretion of soluble factors that downregulate co-stimulatory molecules on antigen-presenting cells, leading to a weakened immune response against HPV and an increased risk of cervical cancer	4—High	Fungal pathogens, soluble factors, co-stimulatory molecules, immune response, HPV, cervical cancer	None identified	3	3	3	3.00
20	Chronic fungal infections in the cervix result in the epigenetic reprogramming of local immune cells, reducing their ability to mount effective anti-HPV responses and increasing the likelihood of cervical cancer progression	4—High	Chronic fungal infections, epigenetic reprogramming, immune response, HPV, cervical cancer	None identified	4	4	4	4.00

E1, E2, and E3 denote Evaluators 1–3 (pathology, oncology, and microbiology, respectively). Group is the arithmetic mean of E1–E3, reported to two decimal places. 'None identified' means that no directly relevant publication was retrieved under the predefined strategy

The targeted literature search retrieved no publication directly matching 48 of the 55 proposed hypothesis-level relationships. Seven hypotheses were associated with at least one related publication: two were associated with multiple publications and five with a single publication. “None identified” therefore indicates that no directly matching publication was retrieved using the applied search strategy; it does not establish that the corresponding hypothesis is definitively novel or biologically valid.

Using the five-point composite evaluation criteria presented in Supplementary Material 1, ChatGPT assigned high overall scores to 44 hypotheses and moderate scores to 11. Based on the arithmetic mean of the three evaluator scores, 35 hypotheses (63.6%) received high overall scores and 20 (36.4%) received moderate scores. The mean scores assigned by E1, E2, and E3 were 3.64, 3.65, and 3.64, respectively, with a median score of 4 for each evaluator. Mean author-evaluator scores were 3.62 for microbial dysbiosis, 3.70 for inflammation, and 3.60 for immune modulation.

Inter-rater reliability across the 55 hypotheses was ICC (2,1) = 0.975 for individual evaluator ratings and ICC (2,3) = 0.992 for the mean ratings of the three evaluators. Pairwise quadratic-weighted kappa values ranged from 0.963 to 1.000. These statistics indicate a high level of consistency among the three evaluators but do not demonstrate the correctness, biological validity, or superiority of the AI-generated hypotheses.

Evaluation of AI-generated hypotheses related to microbial dysbiosis

ChatGPT generated 15 hypotheses related to microbial dysbiosis (refer Table 1; Supplementary Material 2). Fourteen hypotheses had no directly matching publications identified through the targeted literature search. The remaining hypothesis, which proposed that fungal dysbiosis reduces Lactobacillus populations and creates conditions favoring HPV persistence, was associated with six related publications.

ChatGPT assigned high overall evaluation scores to 10 hypotheses and moderate scores to five. Based on the mean ratings of E1–E3, nine hypotheses received high overall scores and six received moderate scores, with a mean author-evaluator score of 3.62. The ChatGPT and author-evaluator score categories agreed for 12 of the 15 hypotheses (80.0%) and differed for three (20.0%) (refer Fig. 1).

Fig. 1.

Overall evaluation scores for AI-generated hypotheses related to microbial dysbiosis. ChatGPT represents the model’s self-assigned overall evaluation score. E1, E2, and E3 represent the scores assigned independently by the three author-evaluators. Group represents the arithmetic mean of the E1–E3 scores for each hypothesis. Scores ranged from 1 (very low) to 5 (very high)

Among the 15 hypotheses, Hypothesis 2— “Specific fungal species within the vaginal microbiome disrupt epithelial integrity, enhancing HPV integration and promoting cancer”—received the highest overall evaluation score from the three author-evaluators (4.33). However, before the author-evaluator assessment, ChatGPT had selected Hypothesis 8— “Fungal metabolites induce oxidative stress in HPV-infected cells, accelerating cervical cancer progression”— for experimental-plan generation because it linked fungal metabolites with oxidative stress and HPV-associated cervical carcinogenesis. This difference reflects the separate AI selection and subsequent author-evaluator assessment stages.

Evaluation of AI-generated hypotheses related to inflammation

ChatGPT generated 20 hypotheses related to inflammation (refer Table 2; Supplementary Material 3). Fourteen hypotheses had no directly matching publications identified through the targeted literature search, whereas six were associated with related publications. Hypothesis 4, which proposed that fungal-induced chronic inflammation upregulates HPV E6 and E7 oncogenes, was associated with 29 publications. Each of the other five hypotheses was associated with one publication.

ChatGPT assigned high overall evaluation scores to 16 hypotheses and moderate scores to four. Based on the mean ratings of E1–E3, 14 hypotheses received high overall scores and six received moderate scores, with a mean author-evaluator score of 3.70. The ChatGPT and author-evaluator score categories agreed for 12 of the 20 hypotheses (60.0%) and differed for eight (40.0%) (refer Fig. 2).

Fig. 2.

Overall evaluation scores for AI-generated hypotheses related to inflammation. ChatGPT represents the model’s self-assigned overall evaluation score. E1, E2, and E3 represent the scores assigned independently by the three author-evaluators. Group represents the arithmetic mean of the E1–E3 scores for each hypothesis. Scores ranged from 1 (very low) to 5 (very high)

Among the hypotheses receiving high overall evaluation scores, Hypothesis 10— “Fungal-induced inflammation increases ROS and NO production in the cervix, causing DNA damage in HPV-infected cells and enhancing carcinogenesis”—was selected for experimental-plan generation. This hypothesis received a mean author-evaluator score of 4.00.

Evaluation of AI-generated hypotheses related to immune modulation

The 20 immune-modulation hypotheses addressed antigen presentation, regulatory T cells, Th1/Th2 balance, immunomodulatory cytokines, immune checkpoints, myeloid-derived suppressor cells, natural killer cell regulation, extracellular vesicles, and immune-cell trafficking (refer Table 3, Supplementary Material 4). Under the predefined literature-search criteria, no directly relevant publications were retrieved for any of the 20 hypotheses.

ChatGPT assigned high overall evaluation scores to 18 hypotheses and moderate overall evaluation scores to 2. Based on the mean ratings of the three author-evaluators, 12 hypotheses received high overall evaluation scores and 8 received moderate scores, with a mean author-evaluator score of 3.60. The ChatGPT and author-evaluators agreed for 12 of the 20 hypotheses (60.0%) and differed for 8 (40.0%) (refer Fig. 3). Category agreement was lower than that observed for microbial dysbiosis (80.0%) but equal to that observed for inflammation (60.0%), indicating some divergence between ChatGPT’s self-evaluation and the author-evaluator assessment in this domain.

Fig. 3.

Overall evaluation scores for AI-generated hypotheses related to immune modulation. ChatGPT represents the model’s self-assigned overall evaluation score. E1, E2, and E3 represent the scores assigned independently by the three author-evaluators. Group represents the arithmetic mean of the E1–E3 scores for each hypothesis. Scores ranged from 1 (very low) to 5 (very high)

Among the hypotheses receiving high overall evaluation scores, Hypothesis 7— “Fungal pathogens alter the expression of immune checkpoint molecules (e.g., PD-L1) on cervical epithelial cells, reducing the effectiveness of the immune response against HPV and increasing the risk of cervical cancer”—was selected for experimental-plan generation. It received an overall evaluation score of 4.00. The hypothesis was selected because it proposed a specific and measurable immune-regulatory mechanism that could be examined under defined fungal-exposure conditions. Its selection represents prioritization for experimental evaluation and should not be interpreted as evidence that cervicovaginal fungal colonization causes clinically meaningful immune-checkpoint activation.

Generation and assessment of AI-generated experimental plans

Following hypothesis generation, ChatGPT was prompted to select one hypothesis from each mechanistic domain and generate a detailed experimental plan: (i) fungal metabolites induce oxidative stress in HPV-infected cells; (ii) fungal-associated inflammation increases ROS and NO production, resulting in DNA damage in HPV-infected cells; and (iii) fungal exposure alters immune-checkpoint pathways, such as PD-L1, potentially impairing immune responses against HPV-infected cells. The complete prompts and unedited AI-generated outputs are provided in Supplementary Materials 2–4.

Each experimental plan contained six components: background, rationale, experimental design, expected outcomes, potential pitfalls, and alternative approaches. The three evaluators subsequently rated these 18 components using the five-point component-specific criteria provided in Supplementary Material 1. Of the components, 12 (66.7%) received high mean scores (≥ 4.00), while six (33.3%) received moderate mean scores (3.00–3.99); none received low or very low scores. Component-level means ranged from 3.00 to 4.67, and the overall mean across all 18 components was 3.76 (refer Table 4).

Table 4.

Author-evaluator score of the components of three ChatGPT-4o generated experimental plans

Hypothesis	E1	E2	E3	Group
1. Fungal metabolites and oxidative stress
Background	3	3	3	3.00
Rationale	4	5	5	4.67
Experimental Design	4	4	4	4.00
Expected Outcomes	4	4	4	4.00
Potential Pitfalls	4	4	4	4.00
Alternative Approaches	4	4	4	4.00
2. Fungal-associated inflammation and ROS/NO production
Background	4	4	4	4.00
Rationale	4	4	4	4.00
Experimental Design	3	3	3	3.00
Expected Outcomes	3	3	3	3.00
Potential Pitfalls	4	5	4	4.33
Alternative Approaches	4	4	4	4.00
3. Fungal modulation of immune-checkpoint pathways
Background	4	4	4	4.00
Rationale	4	4	4	4.00
Experimental Design	3	3	3	3.00
Expected Outcomes	4	4	4	4.00
Potential Pitfalls	3	4	4	3.67
Alternative Approaches	3	3	3	3.00

E1, E2, and E3 denote Evaluators 1–3. Group is the arithmetic mean of E1–E3, reported to two decimal places. Scores range from 1 (very low scientific reasonableness) to 5 (very high scientific reasonableness)

The mean experimental-plan scores assigned by E1, E2, and E3 were 3.67, 3.83, and 3.78, respectively. Inter-rater reliability was ICC (2,1) = 0.821 for individual evaluator ratings and ICC (2,3) = 0.932 for the mean ratings of the three evaluators. Pairwise quadratic-weighted kappa values ranged from 0.727 to 0.914. These results indicate consistency among the evaluators in their assessment of the AI-generated plans. However, the plans were not experimentally implemented or validated in this study.

Evaluation of the experimental plan for fungal metabolites and oxidative stress

The first AI-generated experimental plan was based on the hypothesis that “fungal metabolites induce oxidative stress in HPV-infected cells, accelerating cervical cancer progression” (refer Supplementary Material 2). The proposed plan combined cellular and animal approaches to evaluate oxidative stress, DNA damage, cellular proliferation, HPV oncogene expression, and tumor progression following exposure to selected fungal metabolites. The complete experimental design is provided in Supplementary Material 2.

Among the six components, the background received a moderate mean score of 3.00, whereas the rationale received the highest mean score of 4.67. The experimental design, expected outcomes, potential pitfalls, and alternative approaches each received a mean score of 4.00 (refer Table 4). Thus, five components received high mean scores, and one received a moderate mean score.

The AI-generated plan identified three principal potential pitfalls: uncertainty regarding physiologically relevant metabolite concentrations, cross-species differences between humans and mice, and the complexity of the vaginal microbiome. The proposed alternative approaches included preliminary dose–response studies, humanized mouse models, gnotobiotic mouse models, and microbiome sequencing.

Experimental plan for fungal-induced inflammation and oxidative/nitrosative stress

The second AI-generated experimental plan was based on the hypothesis that “fungal-induced inflammation increases ROS and NO production in the cervix, causing DNA damage in HPV-infected cells and enhancing carcinogenesis” (refer Supplementary Material 3). The background and rationale each received a mean evaluator score of 4.00. The experimental design and expected outcomes each received a mean score of 3.00, while the potential pitfalls and alternative approaches received mean scores of 4.33 and 4.00, respectively (refer Table 4). Thus, four components received high mean scores and two received moderate mean scores.

The proposed plan combined in vitro and in vivo approaches to examine whether fungal-associated inflammation increases oxidative and nitrosative stress, DNA damage, and tumor progression in HPV-related cervical models. It included fungal co-culture and cytokine-stimulation experiments, measurement of ROS, NO, and DNA damage, and an intervention component using antioxidants or NO inhibitors. The complete experimental design is provided in Supplementary Material 3.

The AI-generated plan identified three principal potential pitfalls: limited representativeness of the selected fungal species, sensitivity and specificity limitations in ROS and NO measurements, and differences between human and mouse immune responses. The proposed alternative approaches included testing multiple clinically relevant fungal species, using complementary ROS and NO assays, and applying humanized mouse models or organotypic cervical-tissue cultures.

Experimental plan for fungal modulation of immune-checkpoint pathways

The third AI-generated experimental plan was based on the hypothesis that “fungal pathogens alter the expression of immune checkpoint molecules (e.g., PD-L1) on cervical epithelial cells, reducing the effectiveness of the immune response against HPV and increasing the risk of cervical cancer” (refer Supplementary Material 4). The proposed plan combined cellular and animal approaches to examine fungal-associated changes in immune-checkpoint expression. T-cell-mediated cytotoxicity, cytokine production, immune-cell infiltration, tumor progression, and responses to PD-L1 blockade. The complete experimental design is provided in Supplementary Material 4.

Among the six components, the background, rationale, and expected outcomes each received a high mean score of 4.00. The experimental design and alternative approaches each received a moderate mean score of 3.00, while the potential pitfalls received a mean score of 3.67 (refer Table 4). Thus, three components received high mean scores and three received moderate mean scores.

The AI-generated plan identified limited representativeness of the selected fungal species, differences between human and mouse immune responses, and difficulties in detecting subtle changes in PD-L1 expression as its principal potential pitfalls. The proposed alternative approaches included screening a broader range of fungal species, using humanized mouse or patient-derived xenograft models, and applying single-cell RNA sequencing or multiplex immunofluorescence to assess immune-checkpoint expression.

Materials and methods

AI-assisted hypothesis generation

ChatGPT-4o (OpenAI, San Francisco, CA, USA) was used to generate candidate hypotheses concerning potential relationships between fungi and HPV-associated cervical disease. Three domain-specific prompts addressed microbial dysbiosis, inflammation, and immune modulation. The complete prompts and unedited AI-generated outputs are provided in Supplementary Materials 2–4.

ChatGPT-4o was accessed through the ChatGPT interface. Although the complete prompts and retained outputs were preserved, detailed session metadata, including the access date, browsing configuration, and regeneration history, were unavailable for retrospective reporting.

Targeted literature search

Targeted literature searches were conducted using keywords derived from the fungal factor, proposed mechanism, HPV-related outcome, and cervical-disease context of each hypothesis. The hypothesis-specific keywords are presented in Tables 1, 2 and 3.

A publication was considered directly relevant when it explicitly connected a fungal factor with the proposed mechanism and an HPV-related or cervical-disease outcome. Publications addressing only general fungal carcinogenesis, HPV-associated inflammation without a fungal component, or fungal effects at another anatomical site were considered contextually related rather than directly matching the complete hypothesis.

“None identified” indicates that no directly matching publication was retrieved using the applied search strategy. The database coverage, exact search dates, and complete search strings were not archived contemporaneously and could not be fully reconstructed. The search was therefore treated as exploratory evidence mapping rather than a systematic review or definitive verification of novelty.

Generation and selection of AI-derived hypotheses

After generating hypotheses for each domain, ChatGPT received a follow-up prompt asking it to select the “best hypothesis” and generate a structured experimental plan. Each plan included six components: background, rationale, experimental design, expected outcomes, potential pitfalls, and alternative approaches.

Selection was performed by ChatGPT before assessment by the author-evaluators and therefore not based on their subsequent scores. The selected hypotheses and complete unedited experimental plans are included in Supplementary Materials 2–4.

Author-evaluator assessment of AI-generated hypotheses

Three author-evaluators (authors Y.L., T.G., and L.M.), with expertise in pathology, oncology, and microbiology, respectively, independently reviewed the AI-generated hypotheses and corresponding experimental plans. Prior to evaluation, all evaluators received a standardized briefing describing the study objectives, the three biological themes (microbial dysbiosis, inflammation, and immune modulation), and the predefined evaluation criteria (refer Supplementary Materials 1). The evaluators were instructed to assess the provided materials independently, without discussion or consultation with other evaluators.

Each hypothesis received a single overall evaluation score based on the combined consideration of novelty, potential for success and biological plausibility, impact and implications, and feasibility; these dimensions were not scored separately. The overall evaluation score was recorded on a five-point scale: 1, very low; 2, low; 3, moderate; 4, high; and 5, very high. Detailed definitions for each scoring category are provided in Supplementary Materials 1.

In addition to quantitative scores, evaluators provided qualitative comments describing the rationale for their assessments, including the strengths, limitations, and potential challenges associated with each hypothesis and experimental design. These evaluations were used to compare the perceived scientific value and practical feasibility of the AI-generated research proposals.

Statistical analysis of evaluator agreement

Inter-rater reliability was assessed separately for the 55 overall hypothesis-evaluation scores and the 18 experimental-plan component scores. A two-way random-effects, absolute-agreement intraclass correlation coefficient was calculated for individual evaluator ratings, ICC (2,1), and for the mean ratings of the three evaluators, ICC (2,3). Because the five-point ratings were ordinal, pairwise quadratic-weighted Cohen’s kappa coefficients were also calculated for E1 versus E2, E1 versus E3, and E2 versus E3. Mean author-evaluator scores were calculated as the arithmetic mean of E1–E3 and reported to two decimal places. Mean scores of 4.00–5.00 were categorized as high, 3.00–3.99 as moderate, 2.00–2.99 as low, and below 2.00 as very low. ChatGPT self-assigned scores were not included in the inter-rater reliability analyses. Analyses were performed using R version 4.2.1 and the “irr” package. Reliability statistics were interpreted as measures of evaluator consistency rather than evidence of hypothesis validity. No inferential comparison with human-generated hypotheses or another language model was performed.

Discussion

This pilot study explored the use of ChatGPT-4o for generating hypotheses concerning potential relationships among fungal communities, HPV persistence, and cervical disease. The model generated 55 hypotheses across three biological themes—microbial dysbiosis, inflammation, and immune modulation—and developed structured experimental plans for three selected hypotheses. These findings suggest that ChatGPT-4o can rapidly integrate concepts from different biomedical fields and broaden the initial hypothesis space for an underexplored research question (Datt et al. 2024; Huang et al. 2024; Yao et al. 2025).

The AI-generated outputs were evaluated consistently by the three author-evaluators. Agreement was high for the overall hypothesis-evaluation scores, with ICC (2,1) of 0.975 and ICC (2,3) of 0.992, and remained substantial for the experimental-plan components, with ICC (2,1) of 0.821 and ICC (2,3) of 0.932. The comparatively lower agreement for experimental plans may reflect the greater domain-specific judgment required to assess models, controls, assays, and feasibility. These results indicate that the outputs were sufficiently structured for consistent assessment by the author-evaluators, although inter-rater agreement does not establish their biological validity.

ChatGPT-4o and the author-evaluators did not always prioritize the same hypotheses. Within the microbial-dysbiosis category, ChatGPT-4o selected the fungal-metabolite/oxidative-stress hypothesis, whereas the highest author-evaluator mean was assigned to the hypothesis concerning epithelial-barrier disruption and HPV integration. This difference illustrates the complementary roles of ChatGPT and author evaluation. The barrier-disruption hypothesis addresses an important event in cervical carcinogenesis, but barrier damage alone does not imply viral integration, and commonly used HPV-positive cancer cell lines cannot model the initial integration process. Thus, AI-generated hypotheses require review by the author-evaluators to distinguish conceptual interest from mechanistic specificity and experimental testability.

The literature-search findings should also be interpreted cautiously. No directly matching publication was identified for 48 of the 55 hypotheses, suggesting that the proposed relationships may be underexplored. However, non-retrieval does not prove novelty because it may reflect database coverage, search sensitivity, or differences in terminology (Altman and Bland 1995; Rethlefsen et al. 2021). Moreover, evidence concerning mycotoxin-associated carcinogenesis at other anatomical sites does not establish a direct role for cervicovaginal fungi in cervical cancer (Cao et al. 2022; IARC Working Group on the Evaluation of Carcinogenic Risks to Humans 2012; Kensler et al. 2011; Liu and Wu 2010; Magnussen and Parsi 2013; Marasas et al. 2004; Sun et al. 2007; Wokorach et al. 2022). The most relevant biological context is provided by studies of the cervicovaginal mycobiome, epithelial immunity, and HPV-associated microbial ecology (Alimena et al. 2022; Godoy-Vitorino et al. 2026; Hu et al. 2024a, b; Kyrgiou and Moscicki 2022; Lebeau et al. 2022; Passarelli et al. 2026; Zhang et al. 2024). The generated hypotheses should therefore be regarded as candidates for further investigation rather than established mechanisms.

The author-evaluators also identified limitations in the proposed experimental models. Ect1/E6E7 is not an HPV-negative control because it was immortalized using HPV16 E6 and E7, while CaSki and SiHa cells contain integrated HPV16 DNA and cannot model the initial process of viral integration (Cui et al. 2023; Hu et al. 2024a, b). Similarly, K14-HPV16 mice represent transgenic oncogene expression rather than natural HPV infection. More physiologically relevant models, including primary cervical keratinocytes, cervical organoids, or systems maintaining episomal HPV genomes, would be preferable for studying early HPV-related events. Before implementation, the proposed studies would also require clearly defined fungal-exposure conditions and appropriate vehicle, fungal-only, mammalian-cell-only, cell-viability, fungal-burden, and mechanism-specific controls. Assays such as DCFDA, Griess, and MTT would also require more specific or orthogonal measurements to support mechanistic interpretation (Ruijter et al. 2024).

We recognize that there are limitations in our study. The evaluation involved three author-evaluators, and the composite scoring rubric did not quantify novelty, plausibility, feasibility, testability, and clinical relevance as separate outcomes. Detailed session and literature-search metadata were not fully available for retrospective reporting, and the generation process was not repeated across documented independent sessions. In addition, the study did not include matched human-generated hypotheses or another language model, and the proposed biological mechanisms were not experimentally validated. Consequently, the relative performance of ChatGPT compared with conventional expert-led hypothesis generation remains to be established.

Despite these limitations, this study provides proof-of-concept that ChatGPT-4o can serve as a rapid and structured tool for exploratory biomedical hypothesis generation. The model generated a broad range of mechanistically diverse hypotheses and translated selected ideas into experimental frameworks that could be systematically evaluated by author-evaluators. The high consistency among evaluator ratings further indicates that AI-generated outputs can be subjected to structured and consistent human appraisal. Importantly, the study identified potentially underexplored relationships among fungal biology, HPV persistence, and cervical disease and illustrated the importance of critical human review in identifying assumptions and methodological weaknesses. ChatGPT should therefore be viewed as a complementary tool that can broaden the initial hypothesis space and support, rather than replace, expert scientific reasoning.

Future studies can build on this framework by incorporating multiple independent AI sessions, fully archived model and search metadata, separate evaluation dimensions, blinded assessment, and matched human and AI comparator groups. Experimental validation of the most biologically plausible hypotheses will be essential to determine whether AI-assisted hypothesis generation can ultimately contribute to new mechanistic insights into HPV-associated cervical disease.

Conclusion

This pilot study suggests that ChatGPT-4o can rapidly generate diverse and structured hypotheses linking fungal communities with HPV-associated cervical disease. The high inter-rater agreement indicates that these outputs were sufficiently coherent for consistent scoring by the author-evaluators, while the methodological limitations identified in the experimental plans highlight the continuing need for critical human review. Although the present study does not establish biological causality or superiority over human-led hypothesis generation, it supports the potential use of ChatGPT-4o as a complementary tool for broadening the initial hypothesis space and identifying potentially underexplored research directions. Further studies incorporating repeated AI sessions, matched comparators, independent blinded evaluation, and experimental validation are required to determine the reproducibility and scientific value of this approach.

Supplementary Information

Below is the link to the electronic supplementary material.

Author contributions

Conceptualization, X.L., W.G., and J.M.; methodology, Y.L., T.G., L.M., X.L., W.G., and J.M.; investigation and literature searching, Y.L., T.G., L.M., L.Y., M.L., and S.H.; formal analysis, Y.L., T.G., L.M., W.G., and J.M.; writing—original draft preparation, Y.L. and T.G.; writing—review and editing, all authors; supervision, X.L., W.G., and J.M. All authors have read and approved the final manuscript.

Funding

This work was partially supported by funding from the University of Tennessee Health Science Center (R073290109) to WG in Memphis, TN, USA, and partially supported by funding from the Second Affiliated Hospital of Zhengzhou University (202304121) to JM in Zhengzhou, Henan, China, and partially supported by funding from Qiqihar City Joint Guidance Project of Science and Technology Program (LSFGG-2024101) to YL in Qiqihar.

Data availability

The data generated and evaluated in this study are included in the manuscript and its Supplementary Materials. The evaluation criteria, prompts, retained unedited ChatGPT outputs, generated hypotheses, and author-evaluator scores are provided in the manuscript and Supplementary Materials 1–4. No patient-level, clinical, or publicly sourced dataset was analyzed in this study. Supplementary Materials 1–4 are intended to remain accessible through the journal’s online supplementary-material system upon publication. No separate public-repository deposition is currently planned. Additional information supporting the reported analyses is available from the corresponding author upon reasonable request. No datasets were generated or analysed during the current study.

Materials availability

No biological samples, cell lines, animals, plasmids, or other physical research materials were generated or used in this study. The materials evaluated consisted of AI-generated textual outputs and author-evaluator scoring records, which are provided in the manuscript and its Supplementary Materials.

Code availability

The inter-rater reliability analyses were performed using R version 4.2.1. The analysis code is available from the corresponding author upon reasonable request.

Declarations

Conflict of interest

All authors declare no competing financial interests.

Ethics approval and consent to participate

Not applicable. This study did not involve human participants, patient-level data, animals, or biological experiments; it evaluated AI-generated text, literature-search results, and author-evaluator scores.

Consent for publication

Not applicable.

Declaration of generative and AI-assisted technologies in the writing process

During the conduct of this study, the authors used ChatGPT-4o (OpenAI) as the AI tool under evaluation for exploratory literature synthesis, hypothesis generation, and generation of preliminary experimental plans in response to author-defined prompts. The generated outputs were reviewed and evaluated by the three author-evaluators and were not treated as established scientific evidence. ChatGPT-4o was also used during manuscript revision to assist with language clarity and organization. The authors critically reviewed, verified, and revised all AI-assisted content and take full responsibility for the accuracy, interpretation, and conclusions of the published article. ChatGPT-4o was not assigned authorship and did not make decisions regarding hypothesis validity, author-evaluator scoring, statistical interpretation, or final manuscript content.