Containerized Nextflow pipeline for generating knowledge graph embeddings and link predictions using PyKEEN for multiple biomedical KGs (Hetionet, BioKG, OpenBioLink, PrimeKG).

• Workflow: main.nf (process embedding)
• Global config: nextflow.config
• Per-dataset profiles: conf/hetionet.config, conf/biokg.config, conf/openbiolink.config, conf/primekg.config
• Container build: Dockerfile, requirements.txt
• Deployment pipeline: .github/workflows/docker-deploy.yml
• Terraform (public ECR): terraform/main.tf, terraform/providers.tf, terraform/backend.hcl
• Ignore rules: .gitignore
• License: LICENSE

The Nextflow process embedding loads a selected dataset via PyKEEN, merges training/validation/testing triples, and runs pipeline() with user-supplied hyperparameters from a remote YAML config. Results are saved to config["save"]["path"].

save:
  path: /output/dir
model:
  name: TransE
  embedding_dim: 256
seed: 42
train:
  num_epoch: 50
  num_negative: 32
optimizer:
  class: Adam
  lr: 0.0005

• params.dataset (one of: hetionet, biokg, openbiolink, primekg)
• params.config (S3 or local path to YAML)
• params.outdir (publish directory / S3 prefix)
• params.max_time (wall-time hint)

Profiles supply dataset + config path (see conf/*.config files).

Use a profile (recommended):

nextflow run main.nf -profile hetionet

Override output dir:

nextflow run main.nf -profile openbiolink --outdir s3://bucket/path/

Direct parameter usage (without profile):

nextflow run main.nf --dataset hetionet --config s3://bucket/configs/hetionet.yaml

GPU container is defined in nextflow.config (uses image pushed to public ECR).

Build locally:

docker build -t artemis-kgs-embeddings:local -f Dockerfile .

The CI workflow .github/workflows/docker-deploy.yml auto-tags images with either:

• Git tag (without leading v)
• Commit short SHA

Public ECR repository name is created via Terraform.

Initialize (adjust bucket/table in terraform/backend.hcl):

cd terraform
terraform init -reconfigure -backend-config=backend.hcl
terraform apply

Outputs:

• public_image_uri_latest

Resources:

• Repository: terraform/main.tf
• Provider setup: terraform/providers.tf

Each profile file (e.g. conf/openbiolink.config) sets:

• params.dataset
• params.config (S3 path to YAML)
• Optional resource overrides (cpus, memory)

Global defaults in nextflow.config:

• process.container points to public.ecr.aws/alethiotx/artemis-kgs-embeddings:latest
• process.containerOptions enables --gpus all

The script writes cuda_version.txt after allocating a CUDA tensor to assert GPU availability.

pipeline_result.save_to_directory(config["save"]["path"]) produces:

• Model artifacts
• Embeddings
• Evaluation metrics

Directory path is controlled by save.path in YAML.

• Wrong dataset name: ensure it matches profile.
• Missing GPU: container must run with --gpus all.
• Config path issues: verify S3 permissions and YAML keys.

MIT License in LICENSE.

nextflow run main.nf -profile hetionet

Dockerfile
requirements.txt
main.nf
nextflow.config
conf/hetionet.config
conf/biokg.config
conf/openbiolink.config
conf/primekg.config
.github/workflows/docker-deploy.yml
terraform/main.tf
terraform/providers.tf
terraform/backend.hcl
.gitignore
LICENSE

• Public knowledge graph providers (Hetionet, BioKG, OpenBioLink, PrimeKG)
• PyKEEN, scikit-learn, and Nextflow communities
• Portions of this codebase were assisted using GitHub Copilot (Claude Sonnet 4.5) for code generation, refactoring, cleaning and documentation. The authors reviewed, modified, and validated all AI-assisted code. Responsibility for the correctness, performance, and reproducibility of the code rests entirely with the authors. No AI tools were used to generate scientific conclusions or interpretations in this study.