Skip to content

cdskit localize multilabel integrated v1

Jump to bottom
Kenji Fukushima edited this page Sep 11, 2026 · 3 revisions

Integrated ten-label localization model v1

This published CDSKIT model combines a CNN with ten sequence-feature specialists. It predicts nucleus, cytoplasm, extracellular, mitochondrion, cell_membrane, endoplasmic_reticulum, chloroplast, golgi_apparatus, lysosome_vacuole and peroxisome. Multiple labels can be predicted for one protein.

Install and download

Install CDSKIT >=0.29.0 and PyTorch using the integrated runtime setup. Inference runs on CPU without scikit-learn, Transformers or an ESM download.

Download the .pt checkpoint from the model release. The release also includes MODEL_CARD.md, evaluation.json, LICENSE and SHA256SUMS. For example, in a POSIX shell:

curl --fail --location --output cdskit-localize-multilabel-integrated-v1.pt \
  https://github.com/kfuku52/cdskit/releases/download/localize-multilabel-integrated-v1/cdskit-localize-multilabel-integrated-v1.pt
shasum -a 256 cdskit-localize-multilabel-integrated-v1.pt

The checkpoint SHA-256 must be:

e9b35ead3eca4dcdf833e469f18f1d67e05a3de274b00861cbf08bf138d01621

Predict

cdskit localize --seq_file proteins.faa --seq_type protein \
  --model cdskit-localize-multilabel-integrated-v1.pt \
  --threads 1 --report localization.tsv

For CDS input, use --seq_type dna (the default). There is no short download alias for this model; use the downloaded file path. Keep the default safe loading setting. After installation and download, inference can run offline.

Output contains seq_id, semicolon-separated predicted_labels, ten p_LABEL columns and perox_signal_type. The model uses a separate threshold for each label. Its default legacy decision policy forces at least one label; an explicit --decision_policy safe-v1 supports input abstention while preserving the checkpoint’s label-selection setting for valid sequences. A .json report contains the same row objects; predicted_labels remains a string.

Evaluation and limits

Five-fold development evaluation across three seeds gave mean macro/micro F1 0.5040/0.5965, compared with 0.4028/0.5357 for the baseline. On the previously inspected 1,717-protein HPA set, macro F1 was effectively unchanged (0.2195 to 0.2203), while micro F1 improved (0.4974 to 0.5327). Peroxisome F1 was zero on HPA's seven positives. On 115 additional experimental UniProt positives, recall was 25/115 versus baseline 31/115; that positive-only panel cannot estimate precision or F1.

These results do not establish universal improvement or calibrated biological confidence. The reported evaluations used the legacy decision policy. Current CDSKIT supports --decision_policy safe-v1 for input abstention and explicit --taxonomy_id 9606 to exclude chloroplast labels for human proteins. --organism_group does not suppress chloroplast predictions in this ten-label model. These runtime overrides were not used in the historical metrics above. See the full experiment and audit and the release model card for split design, homology checks and limitations.

License

The CDSKIT-trained weights are MIT-licensed. The release's model card preserves UniProt, DeepLoc and HPA attribution; source datasets retain their own terms. The MIT license update changed no checkpoint bytes.

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.
  • 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.

Clone this wiki locally