cdskit localize peroxisome head

The CDSKIT-trained weights are MIT-licensed; the model release includes LICENSE and SHA256SUMS. The license update leaves the checkpoint bytes and legacy runtime requirements unchanged. Source datasets retain their original terms.

p_peroxisome is available in the experimental prerelease model cdskit-localize-targeting5-perox-deeploc21-et-v1.pt, registered under the alias targeting5-perox-deeploc21-et-v1. When that alias is used with cdskit localize --model, CDSKIT downloads the model from GitHub Releases into the local model cache if it is not already present and verifies its SHA-256 checksum.

cdskit localize \
  --seq_file proteins.faa \
  --seq_type protein \
  --model targeting5-perox-deeploc21-et-v1 \
  --organism_group non_plant \
  --report localize.tsv

The peroxisome head is a CPU-runtime scikit-learn ExtraTrees classifier trained on DeepLoc21 Swiss-Prot train/validation rows with sequence-level features, including C-terminal PTS-like features. Its p_peroxisome output is a learned sequence-label score, not an independently calibrated biological probability. The historical evaluation below does not establish broad peroxisome-associated localization accuracy, and a PTS motif match alone does not prove targeting.

Release asset

Use the pretrained runtime setup. This artifact contains legacy scikit-learn 1.5.2 estimators. Current CDSKIT supports their known loss layout on scikit-learn 1.9.0 and has been compared against the 1.5.2 reference on a retained synthetic fixture.

• GitHub model release
• Asset: cdskit-localize-targeting5-perox-deeploc21-et-v1.pt
• SHA-256: d0998df8819d975b4392342ab78dccc0dd95cf301e4d2df8f38c73d0b5aab445

Fairness checks

• External accession and exact-sequence overlap are excluded or reported.
• A constant-zero baseline and a regex PTS baseline are reported.
• MMseqs homology subsets and cluster out-of-fold evaluation are used to expose similarity-driven optimism.
• HPA is kept as a difficult broad-localization stress test.

Candidate comparison

The current candidate uses ExtraTrees because it improved the independent external and cluster-OOF checks compared with the earlier HGB candidate, while keeping inference CPU-only.

Candidate	UniProt external AUPRC	UniProt external F1	Cluster OOF AUPRC	Cluster OOF F1	HPA stress AUPRC	HPA stress F1
HGB, perox_sequence_v1	0.239	0.293	0.191	0.212	0.007	0.000
HGB, broad_localize_v1	0.209	0.302	0.125	0.142	-	-
ExtraTrees, perox_sequence_v1	0.256	0.370	0.246	0.277	0.149	0.250

Evaluation snapshot

These are historical development results, not metrics rerun for the current CDSKIT version. The validation threshold was selected on that validation set; use the independent external and cluster evaluations to assess generalization.

Model / evaluation	Rows	Positives	AUPRC	AUROC	F1	Notes
DeepLoc21 validation, perox head	5,462	53	0.482	0.895	0.583	threshold tuned on DeepLoc21 fold_id=4
DeepLoc21 validation, regex PTS	5,462	53	0.080	0.765	0.217	simple PTS1/PTS2 signal baseline
UniProt experimental CC external, perox head	11,395	138	0.256	0.847	0.370	no accession/exact-sequence overlap with DeepLoc21 training
UniProt experimental CC external, regex PTS	11,395	138	0.041	0.642	0.156	same external rows
UniProt experimental CC cluster OOF, perox head	11,395	138	0.246	0.854	0.277	MMseqs clusters at 30% identity / 80% coverage
HPA external stress test, DeepLoc21-trained perox head	1,717	7	0.149	0.665	0.250	broad peroxisome-associated task; no positive homology hits

These results support replacing the previous constant-zero p_peroxisome placeholder for signal-like use cases, but they do not yet support advertising the head as a broad peroxisome-localization model.

Reproduce

These are research commands that need prepared inputs, optional ML dependencies, and an mmseqs executable on PATH. Data files below are not included in the package. Download/prepare the DeepLoc tables with python -m cdskit.deeploc_benchmark --download yes --prepare yes --benchmark no and supply the indicated UniProt snapshot separately. A new live UniProt query does not reproduce a frozen historical dataset exactly.

Main candidate model and UniProt experimental CC external benchmark:

python -m cdskit.perox_benchmark \
  --training_tsv data/localize_bench/deeploc21/deeploc21_localization_train_validation.tsv \
  --external_test_tsv data/localize_bench/eukaryota_full_with_lineage.tsv \
  --external_format uniprot_exp_cc \
  --feature_profile perox_sequence_v1 \
  --model_kind extra_trees \
  --base_model targeting5 \
  --model_out data/localize_bench/perox_deeploc21_et_v1/cdskit-localize-targeting5-perox-deeploc21-et-v1.pt \
  --out_json data/localize_bench/perox_deeploc21_et_v1/perox_benchmark_uniprot_exp_external.json \
  --out_md data/localize_bench/perox_deeploc21_et_v1/perox_benchmark_uniprot_exp_external.md \
  --predictions_prefix data/localize_bench/perox_deeploc21_et_v1/perox_predictions_uniprot_exp \
  --homology_check yes \
  --homology_threads 4 \
  --cluster_oof yes \
  --cluster_oof_source external \
  --cluster_oof_folds 5 \
  --cluster_oof_method mmseqs

HPA stress test for the same DeepLoc21-trained candidate:

python -m cdskit.perox_benchmark \
  --training_tsv data/localize_bench/deeploc21/deeploc21_localization_train_validation.tsv \
  --external_test_tsv data/localize_bench/deeploc21/deeploc21_hpa_test.tsv \
  --external_format prepared \
  --feature_profile perox_sequence_v1 \
  --model_kind extra_trees \
  --homology_check yes \
  --homology_threads 4 \
  --out_json data/localize_bench/perox_deeploc21_et_v1/perox_benchmark_hpa_external.json \
  --out_md data/localize_bench/perox_deeploc21_et_v1/perox_benchmark_hpa_external.md \
  --predictions_prefix data/localize_bench/perox_deeploc21_et_v1/perox_predictions_hpa

Feature and decision versions

See the localization feature and decision contract for the corrected nine-residue PTS2 definition, legacy model compatibility, --decision_policy safe-v1, v2 output and explicit taxonomy constraints. Published artifacts retain their historical defaults; new staged pipeline models default to safe inference with ensure_one_label=False. Scores remain unverified as calibrated biological probabilities.

References

Cite the method or resource actually used, in addition to the CDSKIT version and command. Papers below do not validate newly trained CDSKIT models or new options.

• Thumuluri V et al. (2022). DeepLoc 2.0: multi-label subcellular localization prediction using protein language models. Nucleic Acids Research 50:W228–W234. Paper — Localization labels, sorting-signal data and protein-language-model methodology where used.
• Ødum M et al. (2024). DeepLoc 2.1: multi-label membrane protein type prediction using protein language models. Nucleic Acids Research 52:W215–W220. Paper — DeepLoc 2.1 data/partition provenance; citing the dataset does not mean CDSKIT executes the DeepLoc predictor.
• Geurts P, Ernst D, Wehenkel L (2006). Extremely randomized trees. Machine Learning 63:3–42. Paper — ExtraTrees method used for the peroxisome/specialist head.
• The UniProt Consortium (2025). UniProt: the Universal Protein Knowledgebase in 2025. Nucleic Acids Research 53:D609–D617. Paper — Protein sequences and annotations; also report the downloaded snapshot/query and evidence filters.
• Thul PJ et al. (2017). A subcellular map of the human proteome. Science 356:eaal3321. Paper — Human Protein Atlas Cell Atlas annotation resource for the HPA evaluations; report the actual snapshot and label mapping.
• Steinegger M, Söding J (2017). MMseqs2 enables sensitive protein sequence searching for the analysis of massive data sets. Nature Biotechnology 35:1026–1028. Paper — Cite when MMseqs2 is used for homology filtering or clustering in the research workflow.
• Flynn CR, Mullen RT, Trelease RN (1998). Mutational analyses of a type 2 peroxisomal targeting signal that is capable of directing oligomeric protein import into tobacco BY-2 glyoxysomes. The Plant Journal 16:709–720. Paper — Experimental PTS2 nonapeptide evidence; the CDSKIT motif/window detector is a heuristic.
• Gonzalez NH et al. (2011). A Single Peroxisomal Targeting Signal Mediates Matrix Protein Import in Diatoms. PLOS ONE 6:e25316. Paper — Evidence that PTS2-mediated import is not universal across taxa; motif matches alone do not prove localization.