A ML-framework for the discovery of next-generation IBD targets using a harmonized single-cell atlas of patient tissue

This study presents a machine learning framework that leverages a harmonized single-cell atlas of human intestinal tissue to systematically identify, prioritize, and experimentally validate novel, cell type-specific therapeutic targets for IBD, successfully demonstrating the reversal of disease-associated programs through mechanisms distinct from existing biologics.

The Gist

Imagine the human gut as a bustling, complex city. In a healthy city, the police (immune cells), construction crews (fibroblasts), and sanitation workers all work in harmony to keep things running smoothly. But in patients with Inflammatory Bowel Disease (IBD), this city is in chaos. There are riots, fires, and construction gone wrong, leading to painful inflammation and scarring.

For decades, doctors have tried to fix this city by throwing a "net" over the whole area. They use broad-spectrum drugs (like steroids or TNF blockers) that calm down everyone in the city. While this stops the riots, it also shuts down the good guys, leaving the city vulnerable to other problems and causing side effects.

This paper introduces a new, high-tech approach to fix the city with surgical precision. Instead of a net, they built a "Smart City Map" and a "Digital Twin" to find the exact troublemakers and fix them without disturbing the peace elsewhere.

Here is how they did it, broken down into simple steps:

1. Building the "Master Map" (The Harmonized Atlas)

Imagine trying to understand a city by looking at 20 different, blurry, hand-drawn maps from different neighborhoods. It's confusing and full of errors.

• What they did: The researchers took data from 20 different scientific studies (like 20 different mapmakers) and used a super-smart computer system (called AMICA) to stitch them all together into one giant, crystal-clear, high-definition map.
• The Result: They created a "Google Earth" of the human gut containing nearly 1 million individual cells. This map shows exactly which cells are causing trouble, where they are, and what they are saying to each other.

2. The "Digital Twin" Detective (The ML Framework)

Now that they have the map, they needed a detective to find the specific culprits.

• The Tool: They used a machine learning tool called IPR (Immune Patient Representation). Think of this as a "Digital Twin" of a patient's immune system.
• How it works: Instead of just looking at the whole city, the AI looks at specific groups of cells (like the "Lipid-Associated Macrophages," a fancy name for a specific type of fat-loving police officer). It asks: "What is this specific group doing differently in a sick patient compared to a healthy one?"
• The Discovery: The AI found 85 specific "crime patterns" (transcriptional programs) and narrowed them down to 400 potential "suspects" (genes) that could be the root cause of the chaos.

3. The "Wanted List" (Prioritization)

The AI had a list of 400 suspects, but you can't arrest them all. They needed to pick the best ones to test.

• The Filter: They used a checklist to see which suspects were:
  • Actually causing the crime (linked to disease).
  • Easy to catch (druggable).
  • Safe to arrest (won't hurt the heart or liver).
• The Top Picks: This process highlighted two main suspects:
  1. PTGIR: A "bad cop" found in the immune cells (macrophages) that is fueling the fire.
  2. IL6ST: A "bad foreman" found in the construction crew (fibroblasts) that is causing too much scarring (fibrosis).

4. The "Simulation Lab" (Functional Validation)

Before arresting the suspects in real life, they ran a simulation.

• The Experiment: They took real human cells from donors and used gene-editing tools (like molecular scissors) to "knock out" (disable) these two suspects.
• The Result for PTGIR: When they disabled PTGIR in the immune cells, the cells stopped acting angry. They switched from a "riot mode" to a "peacekeeping mode," reducing inflammation. Crucially, this worked differently than current drugs, meaning it could help patients who don't respond to existing treatments.
• The Result for IL6ST: When they disabled IL6ST in the construction cells, the scarring stopped. The cells stopped over-building and started healing.
  • Important Twist: The paper found that if you disabled IL6ST in the immune cells, it actually made things worse! This proves why you need a "smart map"—you have to target the right cell type, or you might accidentally make the disease worse.

5. The "Future Forecast" (Clinical Projection)

Finally, they took the results from their lab simulation and projected them back onto their giant "Master Map" of real patients.

• The Check: They asked, "If we treat a real patient with this new drug, will their gut look like a healthy gut?"
• The Answer: Yes. The simulation showed that targeting these specific genes would reverse the disease patterns found in real patients, and it would do so in a way that is distinct from current drugs.

The Big Picture

This paper is like a blueprint for precision medicine.

• Old Way: "Let's calm down the whole city with a loud siren." (Broad drugs, side effects).
• New Way: "Let's use a high-tech map to find the one specific person starting the fire, and only neutralize them." (Targeted drugs, fewer side effects, better results).

By combining a massive, unified map of human cells with artificial intelligence, the researchers have found two new, highly specific ways to treat IBD that could help patients who currently have no other options. They aren't just guessing anymore; they are navigating the city with a GPS.

Technical Summary

1. Problem Statement

Despite advances in treating Inflammatory Bowel Disease (IBD), including Crohn's disease (CD) and ulcerative colitis (UC), a significant unmet clinical need remains due to high rates of non-response, relapse, and adverse events.

• Limitations of Current Discovery: Traditional target discovery relies heavily on Genome-Wide Association Studies (GWAS). While GWAS has identified over 200 genetic loci, these often point to broad, pleiotropic pathways (e.g., TNF, JAK) leading to therapies with systemic immune suppression and metabolic complications.
• Data Challenges: Single-cell RNA sequencing (scRNA-seq) offers the necessary cellular resolution to identify cell-type-specific disease pathways. However, existing scRNA-seq data is fragmented, characterized by small cohorts, technical variability, and a lack of harmonization, making it difficult to derive robust, generalizable therapeutic targets.
• Translational Gap: There is a disconnect between in silico discovery and in vitro validation, often due to the use of preclinical models that fail to recapitulate the specific cellular states found in human patient tissues.

2. Methodology

The authors developed an integrated framework combining a harmonized single-cell atlas, a machine learning (ML) representation model, and an agentic reasoning system to guide target discovery and validation.

A. Data Harmonization and Atlas Construction

• Platform: Utilized AMICA DB™, Immunai's harmonized database of proprietary and public scRNA-seq datasets.
• Scale: Integrated 20 public studies comprising 530 intestinal samples and 994,206 cells.
• Pipeline: Raw FASTQ files were processed (10x Genomics Cell Ranger), filtered for quality, and integrated using batch correction. Cells were annotated into 129 distinct subtypes/states using a proprietary tissue reference and ontology.
• Validation: An independent cohort of 163 samples (520,969 cells) from four new studies was harmonized to serve as an external validation set.

B. Immune Patient Representation (IPR) Framework

• Concept: An ML framework that learns a latent space representation of patient immune profiles. It aggregates gene expression by cell type within a sample (pseudobulk) to capture sample-level biological signals.
• Execution:
  1. Pseudobulk Construction: Aggregated gene expression by annotated cell type.
  2. Dimensionality Reduction: Applied linear dimensionality reduction to identify latent components capturing major biological variation (e.g., inflamed vs. non-inflamed, responder vs. non-responder).
  3. Feature Extraction: Identified 85 disease-associated transcriptional programs (IPR features) specific to cell types (e.g., lipid-associated macrophages, fibroblasts).
• Interpretation: Used AMICA-Reason™, an agentic LLM-driven framework, to interpret these statistical signals into mechanistic biological modules (e.g., linking gene sets to calprotectin pathways).

C. Target Prioritization

• Candidate Generation: Extracted 4,036 candidate genes correlated with IPR features.
• Scoring & Ranking: Genes were scored and ranked based on:
  1. IPR Metrics: Effect size, linkage to inflammation/non-response, and differential expression.
  2. Orthogonal Data: Genetic linkage (GWAS), network biology (PPI), druggability (Surfaceome), and safety liabilities (expression in heart/liver).
• Filtering: Applied a Target Product Profile (TPP) to filter for cell-type specificity, tractability, and novelty, resulting in a final list of 19 high-priority candidates (from a pool of 343 novel candidates).

D. Experimental Validation & Clinical Projection

• Model Optimization: Used SystemMatch (an ML pipeline) to screen >30 polarization conditions for human mononuclear phagocytes (MNPs) to find the condition most transcriptionally similar to disease-relevant cells in the atlas.
• Functional Assays: Performed CRISPR-Cas9 endonuclease-mediated knockout (KO) of top candidates (PTGIR in MNPs; IL6ST in fibroblasts) in primary human cells.
• Readouts: Multi-omics profiling including bulk RNA-seq, spectral flow cytometry, and multiplexed protein assays.
• Clinical Projection: Projected in vitro KO gene signatures back onto the IBD atlas and external clinical cohorts to verify if the perturbation reversed disease-associated transcriptional programs.

3. Key Results

A. Atlas and IPR Insights

• The harmonized atlas successfully resolved 129 cell states, revealing increased infiltration of monocytes and recruited macrophages in inflamed tissue, consistent with known IBD biology.
• The IPR framework identified 85 distinct transcriptional programs. Pathway analysis confirmed enrichment in TNF signaling, cytotoxicity, and known IBD clinical signatures.
• Benchmarking: The framework successfully recovered ~50% of marketed IBD drug targets and 35% of Phase II/III targets within the top 400 ranked genes, validating its ability to capture established biology while prioritizing cell-type-specific targets (e.g., deprioritizing broad IL-23 due to extra-intestinal expression).

B. Validation of PTGIR (Macrophage Target)

• Discovery: PTGIR (Prostaglandin I2 Receptor) was identified as a top target in lipid-associated macrophages (LAMs) and inflammatory macrophages.
• Mechanism: In vitro KO of PTGIR in human MNPs (polarized to mimic LAMs) reversed pro-inflammatory states induced by PTGIR agonists.
• Phenotype: PTGIR KO led to:
  • Downregulation of pro-inflammatory and pro-fibrotic mediators (VEGF-A, CCL7).
  • Upregulation of oxidative metabolism and stress recovery pathways (TCA cycle).
  • Increased surface expression of regulatory markers (CD36, CD200R).
• Differentiation: The mechanism of action was orthogonal to TNF blockade (Infliximab) and anti-integrin therapy, showing distinct pathway modulation. It showed overlap with JAK inhibition but only in treatment responders, suggesting a potential to mimic JAKi efficacy with better safety.
• Clinical Projection: The PTGIR KO signature reversed disease-associated IPR features in patient macrophages, confirming translational relevance.

C. Validation of IL6ST (Fibroblast Target)

• Discovery: IL6ST (gp130) was a top target in fibroblasts but not in myeloid cells.
• Cell-Type Specificity:
  • In Fibroblasts: IL6ST KO (with OSM stimulation) caused a robust anti-fibrotic shift, downregulating TNF/TGFβ responses and extracellular matrix production (VEGF-A, PCOLCE).
  • In Macrophages: Conversely, IL6ST KO in MNPs triggered a pro-inflammatory phenotype (upregulation of IFNγ pathways and CXCL3/TNF).
• Significance: This highlights the critical importance of cell-type-specific targeting; systemic inhibition of IL6ST could exacerbate inflammation, whereas fibroblast-specific targeting could treat fibrosis.

4. Significance and Impact

• Scalable Precision Medicine: The paper demonstrates a scalable, modular framework that moves beyond broad genetic loci to pinpoint specific genes, cell types, and states driving pathology.
• Bridging the Translational Gap: By anchoring in silico discovery in a high-resolution human atlas and using ML to optimize in vitro models (SystemMatch), the framework significantly improves the predictive validity of preclinical studies.
• Novel Therapeutic Mechanisms: The identification of PTGIR and IL6ST as cell-type-specific targets offers new avenues for IBD treatment that are distinct from current biologics (TNF, JAK, Integrin blockers), potentially addressing non-responders and reducing off-target toxicity.
• Agentic Reasoning: The integration of AMICA-Reason™ demonstrates how LLMs can be used to interpret complex single-cell data, transforming statistical signals into coherent biological hypotheses.
• Future Outlook: The framework is designed to integrate proprietary data and more advanced representation learning models, positioning it as a foundational tool for next-generation drug discovery in immune-mediated diseases.

In conclusion, this work establishes a robust, data-driven pipeline that successfully translates high-dimensional single-cell data into validated, cell-type-specific therapeutic targets, offering a new paradigm for precision drug discovery in IBD.