RNASeq-Synthetic-LLM Framework

A key component of this study was the development of RNASeq-Synthetic-LLM, a Generative AI-assisted framework for transcriptomic data augmentation and biological feature discovery.

Publication:

Generative AI-augmented transcriptomic and microbiome analysis across inflammatory and fibrotic disease states in Crohn's disease.
Frontiers in Artificial Intelligence (2026).
DOI: https://doi.org/10.3389/frai.2026.1881820

Overview

We utilized OpenAI's GPT-4.1-mini model to conditionally generate biologically informed gene expression profiles based on patient metadata and literature-derived molecular rules.

The framework leverages demographic and clinical covariates to simulate realistic transcriptomic variation while preserving known biological relationships observed in inflammatory bowel disease (IBD).

Synthetic Transcriptomic Data Generation

A structured prompting strategy was designed to simulate RNA-seq profiles for specific patient characteristics, for example:

• Age
• Sex
• Disease phenotype
• Fibrotic status
• Inflammatory status

Example patient profile:

65-year-old male with Crohn's disease

Prompts incorporated established biological evidence including:

• Age-associated inflammatory signaling
• Sex-specific immune regulation
• Disease-related cytokine activation
• Fibrosis-associated transcriptional programs

The model generated multiplicative gene-level adjustment factors which were applied to template CPM expression profiles to create biologically plausible synthetic samples.

Syn-LLM-Mod Pipeline

The synthetic transcriptomics workflow integrated:

1. Bulk RNA-seq datasets
2. Normalized CPM matrices
3. Patient metadata
4. Literature-derived biological priors
5. Large Language Model-guided gene modulation

GPT-derived outputs were used to adjust template gene expression levels according to known biological mechanisms.

SPsimSeq Integration

Synthetic RNA-seq samples were further generated using SPsimSeq, a semi-parametric simulation framework for RNA sequencing studies.

Input variables included:

• Age
• Sex
• Disease Label

SPsimSeq-generated expression profiles were subsequently refined through LLM-guided biological adjustments to enhance disease-specific transcriptomic realism.

Biological Knowledge Augmentation

Literature-based biological rules were encoded into the pipeline, including:

• Increased IL6 expression with aging
• Increased TNF signaling during active inflammation
• Immune-response modulation by sex
• Fibrosis-associated extracellular matrix remodeling pathways
• Crohn's disease-specific inflammatory signatures

Applications

The RNASeq-Synthetic-LLM framework was used to:

• Improve feature stability
• Support multi-omics integration
• Prioritize biologically relevant biomarkers
• Characterize inflammatory and fibrotic disease states
• Assist interpretation of transcriptomic signatures identified in Crohn's disease

Workflow

Patient Metadata
       │
       ▼
Literature-derived Biological Rules
       │
       ▼
GPT-4.1-mini Conditional Modulation
       │
       ▼
Gene-level Adjustment Factors
       │
       ▼
Template CPM Profiles
       │
       ▼
Synthetic RNA-seq Generation
       │
       ▼
Multi-Omics Integration
       │
       ▼
Machine Learning & Biomarker Discovery