القائمة
Use when when you have access to annotated MS/MS spectra from a specific ionization mode (e.g., negative ESI) or adduct class (e.
التثبيت باستخدام Codex أو Claude انسخ هذا Prompt والصقه في Codex أو Claude أو مساعد آخر ليراجع صفحة Skill ويثبّتها لك.
استنادا إلى تصنيف SOC المهني
Use when when you have a neural network or machine learning model with multiple tunable hyperparameters (layer size, regularization strength, dropout) or design choices (e.
Use when when building a visualization library that must support multiple plotting backends (e.g., matplotlib, bokeh, plotly) and multiple data types (e.g., chromatograms, spectra, peak maps) without duplicating core logic or configuration handling across backend–plot-type combinations.
Use when after peaks have been assigned to heteroatom classes (e.g., CHO, CHON, CHOS, CHOP) and you need to compare molecular composition across samples, classes, or time series.
Use when after running inference on a trained structure prediction model with one or more input modalities (1H NMR, 13C NMR, or combined), you have generated predicted molecular formulas and connectivity graphs that need to be compared against known ground truth structures.
Use when after computing a histogram of all pairwise mass differences from an MSI dataset, use this skill when you have observed mass difference peaks that may correspond to known adducts (e.g., [M+H]+, [M+Na]+, [M−H2O]+).
Use when you have a trained formula ranking model (such as MIST-CF) and want to measure the specific performance gain from incorporating multiple positive-mode adduct types (e.g., [M+H]+, [M+Na]+, [M+K]+, [M+NH4]+) instead of restricting predictions to [M+H]+ only.
| name | adduct-specific-model-fine-tuning |
| description | Use when when you have access to annotated MS/MS spectra from a specific ionization mode (e.g., negative ESI) or adduct class (e. |
| license | CC-BY-4.0 |
| metadata | {"edam_operation":"http://edamontology.org/operation_3799","edam_topics":["http://edamontology.org/topic_3520","http://edamontology.org/topic_0625","http://edamontology.org/topic_3407"],"tools":["MIST","SCARF","MIST-CF","SIRIUS"],"techniques":["LC-MS"],"license_tier":"open"} |
| derived_from | [{"doi":"10.1021/acs.jcim.3c01082","title":"mistcf"}] |
| evidence_spans | ["an extension of MIST for annotating MS1 precursor masses from MS/MS data","Utilizing sinusoidal formula embeddings as developed in our previous work SCARF"] |
| claims | [] |
| provenance | {"collection":"https://w3id.org/holobiomicslab/asb-skill/collection/metabolomics/v2","assembled_by":"scripts/collect_metabolomics_collection.py","sources":[{"build":"coll_mistcf","doi":"10.1021/acs.jcim.3c01082","title":"mistcf"}],"dedup_kept_from":"coll_mistcf"} |
| schema_version | 0.2.0 |
| attribution | {"generator":"AgenticScienceBuilder","original_doi":"10.1021/acs.jcim.3c01082","all_source_dois":["10.1021/acs.jcim.3c01082"],"zenodo_doi":"10.5281/zenodo.20794027","curators":[],"promoter":"Louis-Félix Nothias","sponsor":"CNRS & Université Côte d'Azur"} |
Fine-tune a formula transformer neural network on MS/MS spectra grouped by ionization mode and adduct type to improve chemical formula ranking accuracy for molecules ionized under specific conditions. This skill adapts a pre-trained model (e.g., MIST-CF) by retraining its adduct tokenization and formula embedding layers on mode-specific data, enabling accurate formula inference for underrepresented ionization modes or novel adduct chemistries.
When you have access to annotated MS/MS spectra from a specific ionization mode (e.g., negative ESI) or adduct class (e.g., [M-H]−, [M+Cl]−) that is either absent or underrepresented in the pre-training data, and you want to improve top-k formula ranking accuracy on that mode without retraining from scratch. This is particularly useful when extending a positive-mode-only model (like the current MIST-CF) to support negative-mode or multi-adduct workflows.
Acquire a curated set of negative-mode or mode-specific MS/MS spectra with ground-truth molecular formulas and adduct annotations from public repositories (e.g., MassIVE, MetaboLights). Preprocess spectra by normalizing intensity to [0, 1] and removing noise below a signal-to-noise threshold (e.g., SNR < 3). Extend the adduct tokenization layer in the pre-trained model to include the target adducts ([M-H]−, [M+Cl]−, [M+FA]−, etc.), preserving the sinusoidal formula embeddings learned by SCARF. Fine-tune the extended model using the mode-specific training set with a held-out validation split (e.g., 80/20), monitoring loss on a validation fold to prevent overfitting. Evaluate the fine-tuned model on a held-out test set by computing top-1, top-5, and top-10 formula ranking accuracy and compile results in a metrics table stratified by adduct type. Use energy-based scoring to rank formula candidates and ensure convergence by checking that validation loss plateaus.
. run_scripts/public_data_train/train_mist_cf.sh