Shadowgraph
About
Sign in

Sitgttish/etap-clf

ETAP-CLF: an ESM3-based classifier for regulated cell-death protein prediction (ferroptosis, pyroptosis, senescence).

0

stars

1

commits

Jupyter Notebook

primary language

Sep 5, 2026

updated

README

ETAP-CLF

Code for “ETAP-CLF: an ESM3-based transformer attention framework for binary protein classification.”

Preprint: bioRxiv 10.64898/2026.08.16.745129

ETAP-CLF keeps a pretrained ESM3 backbone (esm3_sm_open_v1) frozen, projects its per-residue embeddings into a 256-dimensional latent space, contextualizes them with four pre-layer-normalized transformer encoder layers, and aggregates them with a learned single-query attention-pooling head before a binary classifier — 2,503,682 trainable parameters. The same architecture and hyperparameters are trained separately for ferroptosis, senescence, and pyroptosis.

Ferroptosis is the primary benchmark and is evaluated twice: on an internal random sequence split, and on a gene-disjoint external cohort whose genes are entirely absent from model development.

Layout

etap/            the ETAP-CLF package — model, data, training, evaluation, CLI
ferroptosis/     the ferroptosis model notebook and cohort description
figures/         figure generators and published vector panels
pyroptosis/      pyroptosis split assignment and training history
paper/           the manuscript
*.ipynb          task notebooks (see the map below)

Where each result comes from

PaperProduced by
Figure 1A — architecturefigures/fig2_architecture.py
Figure 1B–D — workflowfigures/fig1_workflow.py
Figure 2A–E — pooling UMAPsfigures/fig_umap3d_five.py, vectors from ESM3_FerroCLF_pooling_ablation.ipynb
Figure 4A–D — senescence & pyroptosisfigures/fig4_portability.py
Figure 5A–F — recovery & pathway enrichmentfigures/fig5_pathways.py
Table 1 — dataset compositionferroptosis/ETAP_Ferroptosis.ipynb, ETAP_Senescence.ipynb, ETAP_Pyroptosis.ipynb
Table 2 — ferroptosis performanceferroptosis/ETAP_Ferroptosis.ipynb, Baselines_Published_Predictors.ipynb, ETAP_FeroConCap_comparison.ipynb
Table 3 — senescence & pyroptosisETAP_Senescence.ipynb, ETAP_Pyroptosis.ipynb

Figure filenames in figures/ predate the manuscript's final numbering; this table is authoritative.

Reproducing

  1. pip install -r requirements.txt, and install MMseqs2 for redundancy reduction.
  2. Set an HF_TOKEN Colab secret (ESM3 is gated). No token is hard-coded.
  3. Run ferroptosis/ETAP_Ferroptosis.ipynb for the primary result, then the baseline and portability notebooks.
  4. Generate figures from figures/. The Figure 4 and Figure 5 scripts are plotting-only and take source-data CSVs as command-line arguments.

Sequence data, ESM3 embeddings, and model weights are not stored here — see Data availability in the manuscript.

Contributors

jr4658

1 commits

Similar repos

facebookresearch/esm

Evolutionary Scale Modeling (esm): Pretrained language models for proteins

4,169

Biohub/esm

2,947

evolutionaryscale/esm

2,947

chao1224/ProteinCLAP_pretrain_EBM_NCE_downstream_property_prediction

Model

0

michele1993/protGPT3

2

Jinyeop3110/Post_Training_Protein_LLM

0

microsoft/protnote

ProtNote is a multimodal deep learning model that leverages free-form text to enable both…

64

DessimozLab/ESM3di

using ESM2 to translate amino acids to 3di sequences to create foldseek

1

Languages

Jupyter Notebook

56.2%

Python

43.8%

Sitgttish/etap-clf

ETAP-CLF: an ESM3-based classifier for regulated cell-death protein prediction (ferroptosis, pyroptosis, senescence).

0

stars

1

commits

Jupyter Notebook

primary language

Sep 5, 2026

updated

README

ETAP-CLF

Code for “ETAP-CLF: an ESM3-based transformer attention framework for binary protein classification.”

Preprint: bioRxiv 10.64898/2026.08.16.745129

ETAP-CLF keeps a pretrained ESM3 backbone (esm3_sm_open_v1) frozen, projects its per-residue embeddings into a 256-dimensional latent space, contextualizes them with four pre-layer-normalized transformer encoder layers, and aggregates them with a learned single-query attention-pooling head before a binary classifier — 2,503,682 trainable parameters. The same architecture and hyperparameters are trained separately for ferroptosis, senescence, and pyroptosis.

Ferroptosis is the primary benchmark and is evaluated twice: on an internal random sequence split, and on a gene-disjoint external cohort whose genes are entirely absent from model development.

Layout

etap/            the ETAP-CLF package — model, data, training, evaluation, CLI
ferroptosis/     the ferroptosis model notebook and cohort description
figures/         figure generators and published vector panels
pyroptosis/      pyroptosis split assignment and training history
paper/           the manuscript
*.ipynb          task notebooks (see the map below)

Where each result comes from

PaperProduced by
Figure 1A — architecturefigures/fig2_architecture.py
Figure 1B–D — workflowfigures/fig1_workflow.py
Figure 2A–E — pooling UMAPsfigures/fig_umap3d_five.py, vectors from ESM3_FerroCLF_pooling_ablation.ipynb
Figure 4A–D — senescence & pyroptosisfigures/fig4_portability.py
Figure 5A–F — recovery & pathway enrichmentfigures/fig5_pathways.py
Table 1 — dataset compositionferroptosis/ETAP_Ferroptosis.ipynb, ETAP_Senescence.ipynb, ETAP_Pyroptosis.ipynb
Table 2 — ferroptosis performanceferroptosis/ETAP_Ferroptosis.ipynb, Baselines_Published_Predictors.ipynb, ETAP_FeroConCap_comparison.ipynb
Table 3 — senescence & pyroptosisETAP_Senescence.ipynb, ETAP_Pyroptosis.ipynb

Figure filenames in figures/ predate the manuscript's final numbering; this table is authoritative.

Reproducing

  1. pip install -r requirements.txt, and install MMseqs2 for redundancy reduction.
  2. Set an HF_TOKEN Colab secret (ESM3 is gated). No token is hard-coded.
  3. Run ferroptosis/ETAP_Ferroptosis.ipynb for the primary result, then the baseline and portability notebooks.
  4. Generate figures from figures/. The Figure 4 and Figure 5 scripts are plotting-only and take source-data CSVs as command-line arguments.

Sequence data, ESM3 embeddings, and model weights are not stored here — see Data availability in the manuscript.

Similar repos

facebookresearch/esm

Evolutionary Scale Modeling (esm): Pretrained language models for proteins

4,169

Biohub/esm

2,947

evolutionaryscale/esm

2,947

chao1224/ProteinCLAP_pretrain_EBM_NCE_downstream_property_prediction

Model

0

michele1993/protGPT3

2

Jinyeop3110/Post_Training_Protein_LLM

0

microsoft/protnote

ProtNote is a multimodal deep learning model that leverages free-form text to enable both…

64

DessimozLab/ESM3di

using ESM2 to translate amino acids to 3di sequences to create foldseek

1

Contributors

jr4658

1 commits

Languages

Jupyter Notebook

56.2%

Python

43.8%