Analysis outputs

<p dir="ltr">Updated winnow analysis outputs (tabular results).</p><p dir="ltr"><b>Models and training data (Hugging Face, pinned revisions):</b></p><ul><li><a href="huggingface.co/datasets/InstaDeepAI/winnow-ms-datasets">InstaDeepAI/winnow-ms-datasets</a> @ 659802319d618a359de5ab90ec6b0195681e94a6 (training/evaluation spectra and InstaNovo predictions; DOI 10.57967/hf/6610)</li><li><a href="huggingface.co/InstaDeepAI/winnow-general-model">InstaDeepAI/winnow-general-model</a> @ e2089330dd59adb9685e5b3d7d61f0cd69a3bbb0 (calibrator for general_results/, feature_importance/, and related analyses; config key train_extra_small_mass_error_da; DOI 10.57967/hf/6611)</li><li><a href="huggingface.co/InstaDeepAI/winnow-helaqc-model">InstaDeepAI/winnow-helaqc-model</a> @ d56542b961eac7d896e51bf0716a242fc394ab1f (InstaNovo HeLa calibrator (helaqc_results/instanovo/); DOI 10.57967/hf/6612)</li></ul><p dir="ltr"><b>Additional calibrators (Figshare, same project/collection):</b></p><ul><li>Additional HeLa Single Shot Models (<a href="doi.org/10.6084/m9.figshare.32744946.v2" target="_blank" rel="noreferrer">10.6084/m9.figshare.32744946.v2</a>) (helaqc_results/casanovo/ and helaqc_results/primenovo/)<br>Folders: casanovo_helaqc/, primenovo_helaqc/</li><li>Hold-one-out generalisation models (<a href="doi.org/10.6084/m9.figshare.30147364">10.6084/m9.figshare.30147364.v2</a>) (companion to generalisation/calibrator_generalisation_results.csv)<br>Folders: trained_on_gluc/, trained_on_helaqc/, trained_on_herceptin/, trained_on_immuno/, trained_on_sbrodae/, trained_on_snakevenoms/, trained_on_tplantibodies/, trained_on_woundfluids/</li></ul><p dir="ltr"><b>Outputs are organised into folders:</b></p><ul><li><b>ablations/</b> — Feature-ablation evaluation: aggregated metrics across withheld feature groups (ablation_summary.csv).</li><li><b>fdr_overlap/</b> — Winnow vs database-search overlap at 1 %, 5 %, and 10 % nominal FDR: retained PSM and unique-peptide counts, discordance categories, and per-project overlap summaries.</li><li><b>feature_importance/</b> — Feature-importance analysis for the general model on PXD014877 (C. elegans): permutation importance (perm_importance.pkl) and SHAP values (shap_values.pkl).</li><li><b>general_results/</b> — General-model evaluation on external benchmark datasets from InstaDeepAI/winnow-ms-datasets (general_model_evaluation/). labelled/ holds database-search reference runs; full/ holds full-search predictions.

Each project folder contains metadata.csv and preds_and_fdr_metrics.csv.</li><li><b>generalisation/</b> — Leave-one-source-out calibrator generalisation metrics (calibrator_generalisation_results.csv). Corresponding calibrator checkpoints are in Figshare article 30147364 (trained_on_*/ folders, one model per held-out training source).</li><li><b>helaqc_results/</b> — HeLa QC benchmark (PXD044934): InstaNovo, Casanovo, and PrimeNovo prediction outputs. instanovo/ uses InstaDeepAI/winnow-helaqc-model; casanovo/ and primenovo/ use the HeLa calibrators in Figshare article 32744946 (casanovo_helaqc/, primenovo_helaqc/).