BGPT: Paper Review: Inferring Gene Presence in Incomplete Data via Phylogenetic Occupancy Modeling

Explore by Goal

Quick Explanation Copied

Fast verdict (skeptical, evidence-based)

This draft proposes a phylogenetic occupancy (belief-network) model that simultaneously infers true gene presence and genome completeness, using a latent binary “occupancy” state evolving on a fixed tree and an observation model governed by completeness. The approach is claimed to improve core-genome inference and enable ancestral-state reconstruction under incomplete genomes.

Key strengths are methodological coherence with ecology occupancy models and exact marginalization via Felsenstein-style pruning on trees . Key risks are that the model assumes a particular evolutionary coupling (two-state CTMC-like dynamics with mixture rate heterogeneity) and uses ML point estimates, potentially under-representing uncertainty if real data violate these assumptions.

Long Explanation

Paper Review (draft): Inferring Gene Presence in Incomplete Data via Phylogenetic Occupancy Modeling

DOI: 10.64898/2026.02.27.708499 (March 02, 2026).

Authors: John S. A. Mattick; Wesley C. Demontigny; Charles F. Delwiche.

1) What the model is (visual first)

A latent gene presence state z evolves across genomes on a phylogenetic belief network, while observed gene detection x is generated by an occupancy observation model governed by genome completeness p.

2) Core scientific idea

The observation model uses an occupancy formulation where, for each gene i and genome j, the observed presence/absence x_ij is an imperfect detection of latent occupancy z_ij, with completeness p_j controlling the probability of observing absence despite true presence.
The “coupling across genomes” is achieved by conditional independence on a phylogenetic tree: latent occupancies for a given gene are coupled by evolution along branches, with pruning-style dynamic programming for exact marginalization.
Rate heterogeneity is incorporated via a mixture of rate multipliers for branch length t (10-component mixture, discretized vs log-normal mixture noted), intended to reflect different “transience” levels among gene families.

3) Simulation results (what’s actually supported by numbers we have)

The manuscript reports precision–recall behavior across simulated completeness profiles and genome counts, and gives one quantitative anchor: “around 90% precision at 40% recall” under many completeness profiles.

Skeptical note: this is an anchor statement, not a full PR curve with uncertainty; without the full plotted data values, we can’t reconstruct calibration/variance from this excerpt alone.

4) Empirical evaluation scope (what populations, what “ground truth” means)

Proteobacterial datasets: α-proteobacteria (442 genomes) and γ-proteobacteria (694 genomes) are sampled from NCBI RefSeq with stringent criteria (reference genomes, type material, excluding atypical genomes and metagenome assemblies).

Phylogenies are inferred from concatenated orthogroups selected to be present in ≥95% of genomes (single-copy preferred; two copies resolved by taking longest ortholog), with alignment via MAFFT and trimming via TrimAl, and tree inference via IQTree2.

Ground truth is operational: for evaluation, “observed” orthogroups are artificially depleted according to completeness profiles and treated as if that depletion produces missingness.

5) Asgard ancestral reconstruction results (numbers you can directly inspect)

Asgard analysis uses 739 Asgard taxa for downstream analysis (741 total minus 2 excluded for extreme gaps in the phylogeny alignment), with 206 TACK outgroup genomes.

They focus on eukaryotic signature proteins (ESPs) mapped to orthogroups using eggNOG-based designations, then perform joint reconstruction to infer ancestral ESP presence/absence counts at major nodes.

The draft also claims that no Asgard group claims more than 40% of identified Asgard ESPs in the common ancestor.

6) Methodological strengths (what seems logically well-grounded)

The approach is conceptually consistent with occupancy modeling: latent presence states are coupled to a detection/observation process, mapping naturally onto incomplete assemblies where gene absence may reflect non-detection rather than true absence.
Using sum-product for marginal posterior probabilities and max-product for MAP reconstructions follows established probabilistic graphical model practice on tree structures.
Maximum-likelihood parameter estimation with ML decoding can scale, and the draft reports practical runtimes (minutes on an NVIDIA L4 GPU per dataset), which is important for genome-scale gene inventories.

7) Critical weaknesses & epistemic risks (what could break the claims)

Fixed phylogeny assumption: the model conditions on a fixed tree and does not propagate uncertainty in topology/branch lengths into gene-presence posteriors, yet the draft acknowledges that real biology may violate model specifications.
ML point estimates under uncertainty: the draft reports maximum likelihood estimates rather than fully characterizing posterior uncertainty (it notes HMC posterior concentration but computational cost). If uncertainty is underestimated, marginal probabilities may be overconfident.
Observational model tied to “completeness” only: the observation probability uses completeness p_j, but in real MAGs/mixed assemblies, non-detection can depend on gene length, GC, local coverage, assembly fragmentation, and clustering errors. This could make “p_j” insufficient.
Orthogroup assignment and alignment pipeline sensitivity: gene presence/absence is defined by orthogroups derived from eggNOG/emapper; orthology errors and alignment/tree errors can propagate into occupancy inference.

8) Reproducibility audit (based on what’s provided in the draft)

The draft claims code availability at a GitHub repository and supplemental materials on Zenodo.
However, this provided full-text excerpt does not include parameter defaults, random seeds, or exact hyperparameter priors beyond the described 10-component mixture and 1,000 optimization epochs; those details would be needed for exact replication.

9) Comparison anchor: what they claim vs baseline methods

The manuscript positions its method as improving on mOTUpan, arguing that at deeper evolutionary depths the assumption that “core genes” are those predictable purely from genome completeness becomes problematic.

10) What would most decisively change my confidence

If the method’s inferred gene-present posteriors showed systematic miscalibration on independent datasets where completeness is measured differently (not just “artificial depletion”), then the implied detection model might be misspecified.
If occupancy inference is highly sensitive to phylogeny errors, then incorporating tree uncertainty would be necessary for robust inference; the authors explicitly list this as future work.

Author reviews (clickable)

Feedback:

Updated: April 11, 2026