← Back to Projects

01 · Computational Biology · Microbiome Dynamics

Inferring Microbial
Interactions.

Reconstruction of the ecological interaction network of the human skin microbiome in atopic dermatitis patients, from raw sequencing counts to a hierarchical Bayesian dynamical model.

Context Summer Internship 2026
Model Generalised Lotka–Volterra
Tools MIMIC · Decontam · Grimer
01 / Objective

From sequencing counts to ecology.

The human skin hosts a diverse microbial community that protects against pathogens and maintains skin homeostasis, yet it remains far less studied computationally than the gut microbiome. This project aimed to infer the microbial interactions driving skin microbiome dynamics in atopic dermatitis (AD) patients, applying the generalised Lotka–Volterra (gLV) framework to a longitudinal dataset.

The report covers preliminary abundance analysis, model fitting across several frameworks, contaminant detection, and an extension to hierarchical Bayesian modelling.

02 / Dataset

Body sites, cohorts, timepoints.

Samples were roughly balanced across four body sites — volar forearm, calf, antecubital fossa and sebaceous forehead — and split between a monthly cohort (1,422 samples) and a weekly cohort (1,156 samples).

Only patients with more than 10 timepoints were retained for model fitting: 30 patients for the sebaceous forehead region and 26 for the three others. The weekly cohort, limited to three AD patients, was ultimately too sparse to support reliable inference and was mainly kept as a cross-check.

03 / Preliminary Analysis

Composition and diversity.

Cutibacterium acnes and Staphylococcus aureus dominate every region, together accounting for roughly 70% of mean relative abundance in the monthly cohort. No clear temporal trend was observed across the multi-year sampling window.

The Shannon diversity index was consistently highest on the volar forearm and calf, and lowest on the sebaceous forehead — consistent with those sites having a smaller share of abundance concentrated in the top species.

04 / Absolute Abundance

Correcting for gene copy number.

Relative abundances were converted to absolute abundances using qPCR-derived total DNA load and a genus-level 16S rRNA copy-number correction sourced from the rrnDB database.

Absolute abundance
\[ A_{i,j} = r_{i,j} \times q_j \times \frac{1}{c_i} \]

Here \(r_{i,j}\) is the relative abundance of genus \(i\) in sample \(j\), \(q_j\) the sample's qPCR DNA load, and \(c_i\) its assigned copy number. Comparing raw, per-sample normalised and centred-log-ratio (CLR) summaries showed that a small number of high-qPCR-load samples can otherwise dominate raw rankings.

05 / Modelling

The generalised Lotka–Volterra model.

Microbial dynamics are described by a system of coupled growth equations, where each taxon's growth rate is modulated by the abundance of every other taxon it interacts with:

gLV dynamics
\[ \frac{dx_i}{dt} = x_i \left( \mu_i + \sum_{j} M_{ij}x_j \right) \]
\(x_i\) Absolute abundance of taxon \(i\) at time \(t\).
\(\mu_i\) Intrinsic growth rate of taxon \(i\).
\(M_{ij}\) Effect of taxon \(j\) on the growth of taxon \(i\) — the interaction matrix the project set out to estimate.

Three complementary approaches

01

Regularised regression

Ridge, Lasso and ElasticNet fits (via MIMIC), with regularisation strength chosen by Leave-One-Patient-Out cross-validation and robustness assessed through Meinshausen–Bühlmann stability selection.

Frequentist
02

Bayesian inference

A Bayesian gLV fit with Gaussian priors on \(M_{ij}\), sampled by MCMC, treating an interaction as credible when its 95% interval excludes zero.

MCMC
03

Hierarchical Bayesian

Per-patient interaction terms drawn from a shared population-level distribution, with partial pooling and an optional horseshoe prior to shrink spurious off-diagonal interactions toward zero.

Partial pooling
06 / Contaminants

Separating biology from reagent noise.

Reagent-derived signals can be misread as genuine community members, producing false interaction terms. Four complementary detection methods were combined — Salter, Salter-CLR, Decontam, and a newly added tool, Grimer, which cross-references each genus against published contamination studies and MGnify biome occurrence data.

A genus flagged by any single method was treated as a contamination candidate and excluded, favouring a conservative pipeline over one prone to spurious ecological signal.

Follow-up finding. After decontamination, Anaerococcus showed a consistent negative effect on Staphylococcus growth across every regression method and the Bayesian model — a pattern consistent with published reports that Gram-positive anaerobic cocci limit S. aureus colonisation, suggesting a genuine rather than artefactual signal.
07 / Results

Diagonals hold, off-diagonals don't.

Across every method — Ridge, Lasso, ElasticNet, Bayesian and hierarchical Bayesian — self-regulation (diagonal) terms were consistently strong and stable, while cross-species (off-diagonal) interaction terms were weak and unstable.

Leave-One-Patient-Out RMSE stayed low and stable on the monthly cohort (0.06–0.12) across all three regression methods. Under a horseshoe prior designed to shrink noise-driven coefficients, essentially all off-diagonal terms collapsed toward zero, while diagonal estimates were unaffected — reinforcing that the off-diagonal structure is not well identified by the current data.

08 / Implementation

Pipeline and tooling.

The core pipeline was built around MIMIC for gLV fitting (regression and Bayesian) and Grimer for contaminant screening. Two additional tools were evaluated but ultimately dropped: MDSINE2, which failed to produce a reliable fit due to NaN issues in its trajectory initialisation and clustering collapsing all taxa into one group at this sample size, and MBPert, a gradient-descent ODE fitting method that proved too computationally expensive to pursue alongside the rest of the project.

Re-running the previous pipeline's copy-number correction from scratch surfaced discrepancies with earlier hardcoded values, the pipeline logic itself was validated as correct once the same copy numbers were substituted back in.
09 / Summary

From raw counts to ecological structure.

The project connects several stages of quantitative microbiome analysis:

01 Preliminary relative-abundance and diversity analysis across body sites and cohorts.
02 Conversion to absolute abundance via qPCR load and rrnDB copy-number correction.
03 gLV model fitting through regularised regression and Bayesian inference (MIMIC).
04 Multi-method contaminant screening, including the newly integrated Grimer tool.
05 Extension to hierarchical Bayesian modelling with per-patient interaction terms.
06 A consistent conclusion: self-regulation is robustly identifiable, cross-species interactions are not, given current data.
← Back to Projects