Generative AI boosted the performance of models hunting for fibrosis signals, but the researchers say the synthetic findings need real-world validation.
Generative AI has helped researchers tease out molecular signals associated with intestinal fibrosis in Crohn’s disease, but the experimental approach remains a long way from predicting which patients will develop irreversible bowel damage.
The study, published in Frontiers in Artificial Intelligence, combined transcriptomic and microbiome data with machine learning and synthetic gene-expression data to investigate the differences between inflammatory and fibrotic Crohn’s disease.
Researchers from the University of Birmingham and University of Birmingham Dubai analysed publicly available datasets comprising tissue from 176 non-IBD controls, 187 people with Crohn’s disease, and 85 with fibrotic Crohn’s disease.
“Intestinal fibrosis is a major complication of Crohn’s disease (CD), a subtype of inflammatory bowel disease (IBD) driven by chronic inflammation and resulting in irreversible structural damage requiring surgery,” the researchers wrote.
“However, the molecular differences between inflammatory and fibrotic CD remain poorly defined.”
Current treatments largely target inflammation and have limited direct effects on the mechanisms driving fibrosis, the researchers said.
“Once fibrosis develops, it is generally considered difficult to reverse,” they wrote.
“Notably, up to 70% of patients with transmural inflammation develop stricturing complications after 10 years of diagnosis.
“At present, surgical intervention remains a primary treatment option for established intestinal fibrosis, although it does not prevent recurrence.”
The researchers initially identified 94 differentially expressed genes shared between inflammatory and fibrotic Crohn’s disease and controls. Machine-learning models narrowed these to 43 genes consistently identified by three different algorithms.
Those genes fell into patterns associated with persistent innate immune activation, loss of epithelial metabolic function, epithelial stress, and impaired barrier integrity.
Some genes involved in nutrient absorption and epithelial metabolism, including MGAM, ALDOB, and FABP6, were elevated in inflammatory Crohn’s disease but decreased in fibrotic disease.
In contrast, genes including CHI3L1, CXCL1, S100A8, TREM1, and IL1RN became progressively more highly expressed, peaking in fibrotic Crohn’s disease.
The findings led the researchers to suggest that fibrotic Crohn’s disease may not represent a completely distinct molecular state, but rather a “reconfigured inflammatory condition” in which persistent immune activation is accompanied by epithelial dysfunction and altered host-microbiome interactions.
The microbiome findings added another layer to that picture, they said.
Analysis of 80 microbiome samples found depletion of short-chain fatty acid-producing genera including Faecalibacterium, Anaerostipes, Coprococcus, and Ruminococcus, alongside enrichment of Bilophila and Bacteroides.
However, the microbiome and transcriptomic datasets came from different patient cohorts, meaning the researchers could not directly integrate the two at the individual patient level.
The researchers then turned to generative AI to tackle another problem – the relatively small number of fibrosis samples.
Synthetic RNA-sequencing profiles were generated using a hybrid approach combining the SPsimSeq simulation tool with GPT-4.1-mini.
The language model was given structured biological rules drawn from the literature and patient metadata – including age, sex, and disease status – to modify simulated gene-expression profiles.
The synthetic data appeared to preserve the overall statistical characteristics and correlation structure of the original gene-expression data.
It also improved classification performance. In one comparison, a Random Forest classifier achieved a mean AUC of 0.79 using the original dataset, 0.80 using SPsimSeq-generated data, and 0.94 with the LLM-generated dataset.
The augmented analysis also surfaced fibrosis-associated genes including IL23R, TGFB, and TNF, which already have established roles in inflammation and fibrosis.
The researchers said there were potentially important clinical implications from the study.
“Although this study is based on retrospective public datasets and does not include prospective clinical validation, the findings suggest several potential directions for future translational research,” they wrote.
Related
“One of the major unmet challenges in clinical practice is distinguishing active inflammation from established fibrosis, as these processes frequently coexist and cannot always be reliably differentiated using current endoscopic or cross-sectional imaging techniques alone.”
But the results came with a substantial caveat, the researchers said, as the AI-generated profiles were simulations, not measurements from additional Crohn’s patients.
They also acknowledged that their synthetic augmentation depended on inferred expression patterns and may not reproduce the full biological variability of real tissue. None of the identified genes or microbial features were validated in an independent patient cohort.
The study was also cross-sectional, preventing the researchers from tracking the transition from inflammation to fibrosis in individual patients over time. Missing information on factors including medication use, disease duration, BMI, race, and diet could also have confounded the results.
Differences in sequencing platforms and protocols across the source datasets presented another limitation, while the relatively small number of fibrosis samples may have reduced the predictive accuracy of the models.
The researchers said larger, well-characterised longitudinal cohorts and independent external validation would be needed before the approach could have a clinical role.
“When considered collectively, these findings suggest that intestinal fibrosis is unlikely to represent a distinct endpoint, but rather reflects a reconfigured inflammatory state driven by ongoing immune activation, loss of epithelial integrity, and interactions between the host and the microbiome,” they concluded.
