name: alphagem description: > AlphaGEM — genome-scale metabolic modelling aided by protein structure alignment and deep-learning dark-metabolism mining. Use when: (1) Building or refining GEMs beyond sequence-only gap filling, (2) Integrating structure signals into metabolic networks, (3) Comparing to gapseq / VMH workflows. Route via metagenomics-llm; pair with gapseq / vmh for classical GEM context. license: MIT metadata: version: "1.0.1" metagenomics.category: "analysis-tools" metagenomics.tags: "AlphaGEM, GEM, metabolism, structure, deep-learning" metagenomics.upstream: "https://github.com/hongzhonglu/AlphaGEM" metagenomics.stage: "mining" metagenomics.maturity: "reviewed" metagenomics.upstream-version: "snapshot 956b584" metagenomics.last-reviewed: "2026-10-18" metagenomics.command-status: "CLI reviewed against upstream source; not executed here"

AlphaGEM

Upstream: hongzhonglu/AlphaGEM

Citation

Han, W. et al. AlphaGEM enables precise genome-scale metabolic modelling by integrating protein structure alignment with deep-learning-based dark metabolism mining. Nat. Commun. (2026). https://doi.org/10.1038/s41467-026-75549-w

See also docs/references.md.

Analytical thinking

Sequence-only GEM reconstruction misses “dark” metabolism. AlphaGEM combines structure alignment with DL mining to improve genome-scale models. Use on high-quality genomes/MAGs after mag-qc; interpret human/gut context with vmh when relevant.

Models remain computational — not measured fluxes. Pin AlphaGEM version, structure sources (e.g. AFDB), and gap-filling rules.

Inputs and run

Require one high-quality, uniquely named genome FASTA, an appropriate reference-model choice (ecoli, yeast, strco, human, or synechocystis), AlphaGEM's Figshare resources, both supplied environments, and a licensed/configured optimizer. Structure mode additionally requires per-gene structures and a mapping table whose IDs exactly match the FASTA.

set -euo pipefail
id=candidate01
genome=genomes/candidate01.fna
[[ -s "$genome" ]] || exit 2
python AlphaGEM.py --help >/dev/null
# Sequence/PLM route; use structure_alignment only when its structure inputs exist.
python AlphaGEM.py --name "$id" --mode plmsearch --fasta "$genome" \
  --refname ecoli --grothmedium min \
  >"${id}.alphagem.stdout.log" 2>"${id}.alphagem.stderr.log"
test -d "working/$id"

For structure evidence, use --mode structure_alignment --structure DIR --maplist TABLE and preflight mapping-table coverage. Do not mix reference species, medium, structure thresholds, or resource snapshots across models.

Decision tree

Metabolic modelling need?
├─ Classical pathway/GEM from MAG → gapseq
├─ Structure + dark metabolism mining → alphagem
├─ Human/gut knowledge integration → vmh
└─ Annotation distillate only → dram

Validation, recovery, and provenance

Require completed logs and a non-empty model/result set under working/<name>; reconcile input genes, annotations, added reactions, gap-filled reactions, medium and objective. Run mass/charge consistency, blocked-reaction, growth/no-growth and biologically justified phenotype checks with the produced model format. Compare new dark-metabolism reactions to gene/structure evidence; do not accept growth alone as validation.

AlphaGEM writes a named working directory: archive it or rerun under a fresh name after changing parameters. Record repository commit, Figshare-resource checksums, AlphaGEM/DeepEC/ESM/Foldseek/optimizer versions, reference model, medium, thresholds, input checksum, full command, logs and model checksum. Predicted reactions and fluxes remain computational hypotheses.

Related skills

metagenomics-llm · gapseq · vmh · dram · deepblast · mag-qc · microbial-mining · tool-selection