GPTomics/bio-single-cell-metabolite-communication

Version Compatibility

Reference examples tested with: mebocost 1.0+, scanpy 1.10+, anndata 0.10+

Before using code patterns, verify installed versions match. If versions differ:

• Python: pip show <package> then help(module.function) to check signatures

If code throws ImportError, AttributeError, or TypeError, introspect the installed package and adapt the example to match the actual API rather than retrying.

Metabolite-Mediated Cell Communication

"Find which cell types exchange metabolites" -> Infer a metabolite's presence from the cells expressing its synthesizing enzymes, then score communication to cells expressing its sensor or transporter.

• Python: mebocost.create_obj() -> infer_commu() (metabolite-sensor scoring)

Governing Principle

Metabolite-mediated communication is a DOUBLE inference and the most speculative layer of cell-cell communication. scRNA-seq never measures metabolites; their levels are inferred from the expression of synthesizing enzymes, then a chain of further assumptions is stacked: enzyme mRNA -> enzyme protein -> enzyme ACTIVITY -> metabolic FLUX -> intracellular metabolite POOL -> SECRETION/export -> extracellular concentration in space -> import/SENSING by the receiver. Every arrow is an assumption and none is measured. On top of this sit all the ligand-receptor caveats (proxy, no spatial geometry in dissociated data, abundance and depth confounds, ambient RNA), so the output is hypothesis-generation ONLY. A defensible claim is never "cell A produces metabolite X" but "cell A expresses the machinery consistent with producing X". Validation is non-optional and means metabolomics (LC-MS), mass-spectrometry imaging (MALDI/DESI), isotope tracing, or perturbation of the enzyme or sensor - not another expression-based method. State the enzyme->flux->level->sensing chain explicitly whenever reporting a result.

Method Decision Table

| Method | Tests what / null | Use when | Fails when | |--------|-------------------|----------|------------| | MEBOCOST | Metabolite SENDER->RECEIVER communication; estimates an extracellular metabolite level from producing-enzyme expression, scores enzyme->sensor pairs, permutation FDR over shuffled labels | The question is metabolite crosstalk between cell types (enzyme-sensor), analogous to CellPhoneDB for L-R | Synthase mRNA present but substrate/cofactor absent; transporter "sensor" is bidirectional/promiscuous so sender/receiver direction is wrong; no spatial geometry | | scFEA | Per-cell metabolic FLUX through modules; graph neural network solver enforcing flux balance (in approximately out) | The question is relative flux per cell to FEED metabolite reasoning, not direct communication | Read as absolute mol/s (fluxes are relative); needs the matching module/stoichiometry files for the species; not a CCC tool itself | | Compass | Per-cell metabolic STATE via flux-balance analysis; reaction penalty inversely proportional to enzyme expression over Recon2, outputs a score per reaction per cell | Comparing metabolic state between conditions (e.g. pathogenic vs non-pathogenic cells) | Used to claim a SECRETED metabolite communicates (output is reaction favorability, not secretion); assumes steady state, questionable for differentiated non-proliferating cells; heavy compute, micropool first | | NeuronChat | Neurotransmitter/neuromodulator communication; vesicular release machinery and synthesis enzymes vs target receptor abundance | Neural systems specifically (glutamate, GABA, dopamine, serotonin, neuropeptides) | Applied outside neural tissue; same expression-proxy and geometry limits |

Methods and their curated databases evolve; before committing, verify current best practice, the metabolite-sensor database version, and required config/species files against the installed package docs.

Confounds That Mimic Communication

| Confound | How it manufactures a fake signal | Mitigation | |----------|-----------------------------------|------------| | Compounded inference | Enzyme mRNA is a poor proxy for metabolite concentration (post-transcriptional control, substrate availability, allostery, compartmentalization), so a "secreted" metabolite may never be made | State the enzyme->flux->level->sensing chain; require metabolomics/MSI/tracing before any production claim | | Bidirectional transporters | A transporter labeled a "sensor" may export rather than import, and many move several metabolites, so sender/receiver direction can be inverted | Treat transporter-based calls as lower-confidence than dedicated-receptor calls; check transport directionality literature | | Ambient RNA | Soup of highly expressed transcripts inflates enzyme/sensor "expression" in clusters that do not transcribe them | Decontaminate (SoupX/DecontX/CellBender) before inference | | Cell-type abundance and depth | Larger clusters tighten the permutation null and deeper cells detect more genes, inflating significance independent of biology | Down-sample, run on integrated counts, do not compare raw counts across conditions | | No spatial geometry | Metabolites diffuse and degrade, but dissociated data has no coordinates, so a "communication" may be between cells never co-located | Validate proximity with spatial metabolomics/MSI; do not claim neighbor exchange from dissociated data |

Run MEBOCOST

Goal: Score metabolite sender->receiver communication between cell types with permutation significance.

Approach: Build a MEBOCOST object from a log-normalized AnnData with cell-type labels and a config file pointing at the metabolite-sensor database, then run permutation inference; results carry both the communication score and an FDR.

from mebocost import mebocost
import scanpy as sc

adata = sc.read_h5ad('adata_annotated.h5ad')   # log-normalized, gene SYMBOLS not Ensembl IDs

# config_path points to mebocost.conf listing the metabolite-enzyme-sensor database paths
# cutoff_prop=0.15: a gene must be expressed in >=15% of a group to count (dropout floor)
# species MUST match the data: mouse data against the human enzyme/sensor DB returns almost nothing
mebo = mebocost.create_obj(adata=adata, group_col='cell_type', condition_col=None,
                           met_est='mebocost', config_path='./mebocost.conf', species='human',
                           cutoff_exp='auto', cutoff_met='auto', cutoff_prop=0.15,
                           sensor_type='All', thread=8)

# n_shuffle=1000: label-permutation null for FDR; min_cell_number=10 drops tiny groups
commu_res = mebo.infer_commu(n_shuffle=1000, seed=12345, Return=True,
                             min_cell_number=10, pval_method='permutation_test_fdr',
                             pval_cutoff=0.05, thread=None)

Filter and Summarize Results

Goal: Extract the significant, defensible metabolite communications.

Approach: Filter on the permutation FDR (not the raw p-value), then summarize by metabolite and by sender->receiver pair; column names are capitalized in the result table.

sig = commu_res[commu_res['permutation_test_fdr'] < 0.05].copy()

# Result columns: Sender, Receiver, Metabolite_Name, Sensor, Annotation (Transporter/Enzyme),
# Commu_Score, Norm_Commu_Score, met_in_sender, sensor_in_receiver, permutation_test_fdr
sig['pair'] = sig['Sender'] + ' -> ' + sig['Receiver']
top_metabolites = sig['Metabolite_Name'].value_counts().head(10)
top_pairs = sig['pair'].value_counts().head(10)

# Transporter-based sensors are lower-confidence (bidirectional); separate them
transporter_calls = sig[sig['Annotation'] == 'Transporter']

Compare Conditions

Goal: Find metabolite communications that differ between conditions (e.g. tumor vs normal).

Approach: Either pass condition_col to a single object, or run MEBOCOST separately per condition subset and compare the significant sets; never compare raw interaction counts across conditions without controlling for cell number and depth.

results = {}
for cond in adata.obs['condition'].unique():
    sub = adata[adata.obs['condition'] == cond].copy()
    obj = mebocost.create_obj(adata=sub, group_col='cell_type', met_est='mebocost',
                              config_path='./mebocost.conf', species='human',
                              cutoff_prop=0.15, thread=8)
    results[cond] = obj.infer_commu(n_shuffle=1000, seed=12345, Return=True,
                                    min_cell_number=10, pval_cutoff=0.05)
# A "differential" metabolite call (significant in one condition only) is a HYPOTHESIS for metabolomics

Threshold and Permutation Rationale

| Parameter | Value | Rationale | |-----------|-------|-----------| | cutoff_prop | 0.15 | A chosen dropout floor (MEBOCOST tutorials use 0.15-0.25, not a fixed package default): a gene expressed in <15% of a group is mostly dropout, but real low-abundance signaling is also discarded, so tune per dataset | | cutoff_exp / cutoff_met | 'auto' | MEBOCOST data-derived thresholds for calling a gene/metabolite present; set manually only with a documented reason | | n_shuffle | 1000 | Stable label-permutation FDR; the FDR is about label shuffling, not actual metabolite flux | | min_cell_number | 10 | Groups under ~10 cells give unstable mean expression and inflated scores | | permutation_test_fdr cutoff | 0.05 | Filter on the FDR, not the raw permutation p-value; significance is statistical, not a measured concentration |

Common Errors

| Symptom | Cause | Fix | |---------|-------|-----| | KeyError: 'metabolite' / 'pval' | Result columns are capitalized (Sender, Receiver, Metabolite_Name, Commu_Score, permutation_test_fdr) | Use the exact column names from commu_res.columns | | AttributeError: module 'mebocost' has no attribute 'create_obj' | Wrong import; create_obj lives in the submodule | from mebocost import mebocost then mebocost.create_obj(...) | | Almost no metabolites detected | Genes are Ensembl IDs, data is not log-normalized, config_path database is missing, or species does not match the data (mouse data run against the human enzyme/sensor DB) | Convert to gene symbols, log-normalize, point config_path at a valid mebocost.conf, set species to match the organism | | A cell type "secretes" a metabolite implausibly | Synthase mRNA present but substrate/cofactor absent, or ambient RNA inflated the enzyme | Decontaminate ambient RNA; treat as "machinery consistent with", validate with metabolomics | | Sender/receiver direction looks reversed | Sensor is a bidirectional/promiscuous transporter | Check Annotation == 'Transporter' calls separately; confirm transport direction | | More communications in condition B than A | Counts scale with cell number and depth | Compare score magnitudes or matched subsets, not raw counts |

Related Skills

• single-cell/cell-communication - Ligand-receptor CCC; the single-inference counterpart this skill mirrors at one extra remove
• single-cell/cell-annotation - Cell-type labels define metabolite senders and receivers
• single-cell/preprocessing - Log-normalization, gene-symbol mapping, and ambient-RNA decontamination happen here, before inference
• metabolomics/pathway-mapping - Places inferred metabolites in pathway context and informs which to prioritize
• metabolomics/isotope-tracing - Orthogonal flux validation that a producing cell actually makes the metabolite
• systems-biology/flux-balance-analysis - Genome-scale FBA underlying Compass-style per-cell metabolic state

References

• Zheng R, et al. MEBOCOST maps metabolite-mediated intercellular communications using single-cell RNA-seq. Nucleic Acids Res 53(12):gkaf569 (2025). PMID 40568942.
• Alghamdi N, et al. A graph neural network model to estimate cell-wise metabolic flux using single-cell RNA-seq data [scFEA]. Genome Res 31(10):1867-1884 (2021).
• Wagner A, et al. Metabolic modeling of single Th17 cells reveals regulators of autoimmunity [Compass]. Cell 184(16):4168-4185 (2021).
• Zhao W, et al. Inferring neuron-neuron communications from single-cell transcriptomics through NeuronChat. Nat Commun 14(1):1128 (2023).
• Dimitrov D, et al. Comparison of methods and resources for cell-cell communication inference from single-cell RNA-Seq data. Nat Commun 13:3224 (2022). [discordance framing]
• Young MD, Behjati S. SoupX removes ambient RNA contamination from droplet-based single-cell RNA sequencing data. GigaScience 9(12):giaa151 (2020).