Adduct Specific Model Fine Tuning

adduct-specific-model-fine-tuning

Summary

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 to use

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.

When NOT to use

• Positive-mode ESI spectra already well-represented in the pre-training data; fine-tuning adds minimal benefit and risks overfitting.
• Fewer than ~100–200 annotated spectra in the target mode; insufficient training signal for stable fine-tuning of a transformer.
• Pre-training data already includes the target adduct type; use the base model directly rather than fine-tuning.

Inputs

• Pre-trained MIST-CF model weights (formula transformer checkpoint)
• MS/MS spectra in .mgf or standardized format with m/z and intensity arrays
• Ground-truth molecular formula and adduct type annotations (e.g., CSV with [spec_id, formula, adduct])
• List of target adducts for tokenization extension (e.g., [M-H]−, [M+Cl]−, [M+FA]−)

Outputs

• Fine-tuned MIST-CF model checkpoint (weights for adduct tokenization and formula transformer layers)
• Metrics table with top-1, top-5, and top-10 formula ranking accuracy per adduct type
• Validation and test loss curves confirming model convergence
• Ranked formula predictions for test spectra with energy-based scores

How to apply

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.

Related tools

• MIST-CF (Pre-trained formula transformer model to be extended and fine-tuned for negative-mode and multi-adduct support) — https://github.com/samgoldman97/mist-cf
• SCARF (Source of sinusoidal formula embeddings that are preserved and reused during fine-tuning to maintain learned chemical space structure) — https://arxiv.org/abs/2303.06470
• SIRIUS (Used to enumerate candidate molecular formulas for a given precursor mass via dynamic programming (SIRIUS decomp)) — https://bio.informatik.uni-jena.de/software/sirius/

Examples

. run_scripts/public_data_train/train_mist_cf.sh

Evaluation signals

• Top-1 formula ranking accuracy on held-out negative-mode test set is ≥ 5 percentage points higher than zero-shot base model, stratified by adduct type.
• Validation loss converges and does not increase over final 10–20 epochs, indicating stable fine-tuning without overfitting.
• Top-5 and top-10 accuracy improve proportionally with top-1 accuracy across adduct types, showing consistent model improvement.
• Energy-based scores (model outputs) are well-calibrated: higher-ranked correct formulas have lower energy (higher probability) than incorrect candidates in ≥ 80% of test spectra.
• Metrics table shows comparable or improved performance on rare adducts (e.g., [M+FA]−) compared to common adducts (e.g., [M-H]−), confirming that fine-tuning balances representation across adduct classes.

Limitations

• MIST-CF is currently restricted to positive-mode only in the published version; extending to negative mode requires custom implementation of the adduct tokenization layer and retraining from a positive-mode checkpoint.
• Fine-tuning is only effective when the target ionization mode or adduct type is sufficiently distinct from pre-training distribution; modes with very similar fragmentation patterns may not benefit significantly.
• Model performance is data-dependent and may degrade on spectra from instruments or sample types not well-represented in the fine-tuning set (e.g., Orbitrap vs. Q-TOF).
• Energy-based scoring requires careful regularization and validation to avoid mode collapse or trivial solutions during fine-tuning; convergence is not guaranteed on small or imbalanced datasets.

Evidence

• [other] Can the MIST-CF formula transformer architecture be extended to support negative-mode adducts, and what is the resulting top-k formula ranking accuracy on negative-mode MS/MS spectra?: "Can the MIST-CF formula transformer architecture be extended to support negative-mode adducts, and what is the resulting top-k formula ranking accuracy on negative-mode MS/MS spectra?"
• [other] Extend the adduct tokenization layer in MIST-CF to include common negative-mode adducts ([M-H]−, [M+Cl]−, [M+FA]−). Fine-tune the formula transformer neural network using negative-mode spectra, maintaining the existing sinusoidal formula embeddings from SCARF.: "Extend the adduct tokenization layer in MIST-CF to include common negative-mode adducts ([M-H]−, [M+Cl]−, [M+FA]−). Fine-tune the formula transformer neural network using negative-mode spectra,"
• [other] Evaluate the fine-tuned model on a held-out negative-mode test set by computing top-1, top-5, and top-10 formula ranking accuracy. Compile results in a metrics table comparing performance across negative adduct types.: "Evaluate the fine-tuned model on a held-out negative-mode test set by computing top-1, top-5, and top-10 formula ranking accuracy. Compile results in a metrics table comparing performance across"
• [readme] MIST-CF ranks chemical formula and adduct assignments for an unknown mass spectrum using an end-to-end energy based modeling approach, without referencing any spectrum databases: "MIST-CF ranks chemical formula and adduct assignments for an unknown mass spectrum using an end-to-end energy based modeling approach, without referencing any spectrum databases"
• [readme] MIST-CF considers multiple adduct types beyond [M+H]+ (still only positive mode): "MIST-CF considers multiple adduct types beyond [M+H]+ (still only positive mode)"
• [readme] Utilizing sinusoidal formula embeddings as developed in our previous work SCARF: "Utilizing sinusoidal formula embeddings as developed in our previous work SCARF"