Lemonite: identification of regulatory metabolites through data-driven, interpretable integration of transcriptomics and metabolomics data

The paper introduces Lemonite, a data-driven and interpretable framework that integrates transcriptomics and metabolomics data to systematically identify regulatory metabolites and their target gene modules, successfully uncovering novel metabolite-gene relationships in glioblastoma and inflammatory bowel disease without relying on prior differential analysis or complete metabolome annotation.

The Gist

Imagine your body is a massive, bustling city. In this city, the genes are the city planners and architects (the blueprints), the proteins are the construction crews and workers building the roads and houses, and the metabolites are the raw materials, fuel, and supplies (like bricks, wood, and electricity) that keep everything running.

For a long time, scientists thought the city planners (genes) were the only ones in charge. They believed the raw materials (metabolites) were just passive leftovers—stuff that was used up and didn't really influence the plans.

The Big Idea
This paper introduces a new tool called Lemonite (think of it as a "Lemon Tree" for metabolites). The researchers realized that the raw materials actually have a say in how the city is built. Sometimes, a shortage of a specific fuel or an excess of a certain chemical can tell the city planners to change the blueprints entirely.

The problem was that we didn't have a good way to figure out which raw materials were influencing which blueprints, especially when we had thousands of them all at once. Existing tools were either too messy to understand or relied on a "rulebook" that was missing most of the pages.

How Lemonite Works: The Detective Analogy
Imagine you are a detective trying to solve a mystery in a crowded room. You have a list of people (genes) acting strangely and a list of items (metabolites) scattered around. You want to know: Who is influencing whom?

1. The Old Way (The Rulebook): Previous tools tried to solve this by only looking at items they already knew were important. If a weird, unknown chemical was in the room, they ignored it because it wasn't in their rulebook.
2. The Lemonite Way (The Pattern Spotter): Lemonite is a data-driven detective. It doesn't need a rulebook. It looks at the whole room, groups the people who are acting together into "gangs" (gene modules), and then asks, "Which chemicals are present when these gangs are active?"

It uses a clever trick: it builds a giant Knowledge Graph. Think of this as a massive, interconnected map of the city. It connects every known chemical to every known protein and gene. When Lemonite finds a new pattern, it checks this map to see if the connection makes sense biologically.

The Case Studies: Two Crime Scenes
The researchers tested Lemonite on two very different "cities" (diseases):

1. Glioblastoma (A Brain Tumor City)

• The Mystery: Why do some brain tumors grow aggressively and resist treatment?
• The Discovery: Lemonite found that specific chemicals, like myo-inositol and phosphatidylcholines (types of fats), were acting as "bosses" for groups of genes.
• The Twist: These chemicals were telling the immune cells (the city's police force) to stop fighting the tumor and actually help it grow. It turns out the tumor was using these chemicals to trick the police into thinking the tumor was a friend. Lemonite identified the specific "police officers" (immune cells) being tricked and the chemicals doing the tricking.

2. Inflammatory Bowel Disease (A Gut City in Chaos)

• The Mystery: Why is the gut lining inflamed and damaged?
• The Discovery: Lemonite found that chemicals like plasmalogens (another type of fat) and carnitine were regulating genes responsible for keeping the gut wall strong.
• The Proof: To prove Lemonite wasn't just guessing, the researchers went into a lab. They took gut cells and fed them these specific chemicals. Just as Lemonite predicted, the cells changed their behavior! For example, feeding them a coffee compound called trigonelline turned on a gene that helps control the body's internal clock. This confirmed that Lemonite could find new rules that no one knew before.

Why This Matters

• No More Guessing: Lemonite can handle "unknown" chemicals. In the past, if a chemical didn't have a name in the database, it was ignored. Lemonite says, "I don't know what this is called, but I know it's talking to these genes, so let's investigate."
• From Data to Action: Instead of just giving a list of numbers, Lemonite gives a story: "Chemical X is likely causing Gene Y to turn on, which leads to Disease Z." This gives doctors and researchers a clear target for new drugs.
• The Future: This tool helps us understand that our diet, our gut bacteria, and our metabolism are constantly talking to our genes. It's not just about what we eat; it's about how that food changes the instructions our cells are following.

In a Nutshell
Lemonite is a new, smart translator that helps us understand the secret conversations between our body's fuel (metabolites) and our body's instructions (genes). By decoding these conversations, we can find new ways to treat diseases like cancer and inflammatory bowel disease, turning the "unknowns" of our biology into actionable cures.

Technical Summary

1. Problem Statement

Current approaches to understanding gene regulation often treat the metabolome merely as a phenotypic endpoint rather than an active regulatory layer. While metabolites are known to influence protein stability, transcription factor (TF) activity, and chromatin remodeling, there is a significant gap in genome-wide gene regulatory network (GRN) inference that incorporates metabolites. Existing multi-omics integration methods suffer from two main limitations:

• Data-driven methods (e.g., MixOmics, MOFA+): These use dimensionality reduction to find latent factors but often lack biological interpretability and do not explicitly model regulatory mechanisms (i.e., they do not identify which metabolite regulates which gene module).
• Prior knowledge-based methods (e.g., COSMOS+): These rely on curated databases to map interactions. However, they are limited by incomplete metabolite annotations (many metabolites in untargeted studies are unknown or ambiguous) and identifier mismatches across databases, leading to the exclusion of a vast majority of features before analysis.

There is a need for a framework that is data-driven (to handle unannotated metabolites), interpretable (to generate testable hypotheses), and capable of identifying regulatory metabolites acting on co-expressed gene programs without requiring prior differential analysis.

2. Methodology: The Lemonite Framework

The authors developed Lemonite (LemonTree for metabolites), a modular, data-driven pipeline for integrating bulk transcriptomics and metabolomics (including lipidomics) data.

A. Core Algorithm

• Input: Coupled transcriptomics and metabolomics data.
• Gene Clustering: Uses the LemonTree ensemble method (based on model-based Gibbs samplers) to infer consensus co-expression modules from transcriptomics data.
• Regulator Assignment: For each gene module, Lemonite infers regulatory programs using an ensemble of decision trees.
  • Predictors: Transcription Factor Activities (TFAs, inferred via DecouplR and CollecTri) and metabolite/lipid abundances.
  • Output: A consensus score for each regulator (TF or metabolite) predicting the module's expression.
• Prioritization: Modules and regulators are filtered and ranked based on:
  • Module coherence (co-expression strength).
  • Differential expression between sample groups.
  • Functional enrichment (GO pathways).
  • Connectivity to known Protein-Protein Interactions (PPIs).
  • Regulator scores (high-scoring regulators targeting multiple modules).

B. The Lemonite Knowledge Graph (KG)

To enhance interpretability and validate predictions, the authors constructed a comprehensive Knowledge Graph integrating:

• Sources: HMDB, BioGRID, UniProt, IntAct, chEMBL, LINCS, STITCHdb, Human1-GEM (genome-scale metabolic model), and MetalinksDB.
• Scale: ~370,000 unique metabolite-gene interactions and ~2.1 million protein-protein interactions.
• Classification: Interactions are annotated as Causal (LINCS, chEMBL), Metabolic Pathway (Human1-GEM), or Other (physical associations).
• Function: The KG is used for in silico validation of data-driven predictions, contextualizing regulators via PPIs (e.g., connecting a metabolite to a TF via a single intermediate protein), and prioritizing modules enriched for known interactions.

C. Implementation

• Available as a NextFlow pipeline on GitHub.
• Includes a web interface (www.lemonite.ugent.be) for interactive exploration.

3. Key Results

A. Benchmarking Existing Methods

The authors re-analyzed two cohorts using MixOmics (DIABLO), MOFA+, and COSMOS+:

• GBM (Glioblastoma, n=99): DIABLO and MOFA+ identified latent factors distinguishing subtypes (e.g., IDH-mutant vs. others) but failed to provide mechanistic links between specific metabolites and gene programs. COSMOS+ could only map a tiny fraction of metabolites due to database limitations.
• IBD (Inflammatory Bowel Disease, n=75): Similar limitations were observed; latent factors were identified, but direct regulatory mechanisms remained obscure.

B. Application to Glioblastoma (GBM)

Lemonite identified 63 modules, with 46 being coherent and 30 enriched for PPIs.

• Key Findings:
  • Mesenchymal-like Immune Programs: Modules associated with the mesenchymal subtype were regulated by IRF6 (TF) and myo-inositol and phosphatidylcholines (metabolites).
  • Cell-Type Specificity: Integration with single-cell RNA-seq revealed these modules were primarily expressed in Tumor-Associated Macrophages (TAMs) and monocytes.
  • IDH Mutations: A specific module (Module 18) linked 2-hydroxyglutaric acid (2-HG) and L-methionine to DNA/histone methylation pathways via the enzyme BCAT1. The KG contextualized this by showing how 2-HG inhibits demethylases while methionine feeds the methylation cycle.
  • Hub Regulators: Identified Creatinine, 2-HG, Myo-inositol, and Homocysteine as major metabolic hubs.

C. Application to Inflammatory Bowel Disease (IBD)

Lemonite identified 97 modules, with 86 retained after filtering.

• Key Findings:
  • Ion Transport: Module 49 (downregulated in UC) was enriched for ion transport genes and regulated by plasmalogens.
  • Hub Regulators: Plasmalogens, carnitine species, sphingomyelins, and putrescine were identified as key metabolic hubs.
  • Novel Predictions: The method prioritized trigonelline (a coffee compound) as a regulator of PER3 (circadian rhythm gene).

D. Experimental Validation

The authors experimentally validated novel predictions in the HT29 colonic epithelial cell line:

• C2-carnitine: Significantly upregulated ME1 and BHLHE40 expression.
• Trigonelline: Upregulated PER3 expression.
• $\alpha$-glycerophosphocholine: Upregulated NAAA expression.
• Significance: None of these specific metabolite-gene pairs existed in the prior knowledge graph, proving Lemonite's ability to discover novel, data-driven regulatory interactions.

4. Key Contributions

1. Lemonite Framework: A novel, interpretable, data-driven method that jointly infers gene modules and their regulators (TFs and metabolites) without requiring prior differential analysis or complete metabolite annotation.
2. Lemonite Knowledge Graph: A unified resource integrating >370k metabolite-gene interactions and >2M PPIs, resolving identifier mismatches and classifying interaction types (causal vs. metabolic).
3. Discovery of Regulatory Metabolites: Successfully identified specific metabolites (e.g., myo-inositol, trigonelline) acting as regulators of gene programs in complex diseases, moving beyond the view of metabolites as mere endpoints.
4. Experimental Validation: Provided in vitro evidence for novel metabolite-gene interactions that were not previously documented in literature or databases.

5. Significance

This work addresses a critical bottleneck in systems biology: the integration of the metabolome into gene regulatory networks. By demonstrating that metabolites can be systematically identified as regulatory drivers of gene expression, Lemonite shifts the paradigm from viewing metabolism as a passive readout to an active regulatory layer. The ability to handle unannotated metabolites and generate experimentally testable hypotheses makes this tool highly valuable for studying complex diseases like cancer and IBD, where metabolic reprogramming is a hallmark. The open-source nature of the pipeline and the interactive web resource further democratize access to multi-omics integration.