Release training recipe

The 2.3.0 weights use the final-2.3.0-candidate-v2 recipe identifier and the current data vintage: the curated 2023 affinity snapshot plus the configured mass-spectrometry sources. The identifier is fixed for reproducibility; its name does not describe the software’s release status. The machine-readable settings are in scripts/training/final_230_candidate_v2_recipe.json.

Components

Component	Release configuration
Affinity	35 architectures × 4 folds; minibatch 1024; native PyTorch RMSprop; Glorot initialization; pre-activation LSUV; dropout 0.5; patience 20; terminal weights for selection.
Processing without flanks	128 architectures × 4 folds; minibatch 512; Glorot initialization and Keras-equation Adam; selected ensemble.
Short-flank processing	One kernel-13 architecture × 4 folds; five residues on each side; native RMSprop; Glorot initialization; best inner-validation AP checkpoints.
Boundary processing	One large-ReLU architecture × 4 folds; five flank residues and five peptide residues at each cleavage boundary; Glorot initialization, Keras-equation Adam and terminal checkpoints.
Presentation with flanks	Equal mixture of the four short-flank and four boundary networks, combined with affinity by the fitted logistic model.
Presentation without flanks	Selected no-flank processing ensemble, combined with affinity by its separately fitted logistic model.
Long-flank processing	Separate 15-residue-flank diagnostic grid; it is not the processing ensemble embedded in the presentation predictor.

Training checkpoints and model-selection records remain in the training run. An inference download needs the selected weights, model manifests, allele sequences, fitted presentation coefficients and calibration. Training data and provenance are retained for overlap audits; unused alternative checkpoints are not part of the inference model.

Data and evaluation

Affinity uses measured affinities and the configured mass-spec reassignment. Its synthetic negatives are random amino-acid peptides. Processing training uses protein-derived negatives matched within sample, length and affinity constraints. Presentation training uses protein-derived negatives in its own training table. These training samplers do not define a comparison’s test set.

A model comparison must freeze one positive/negative row set, score every comparator on it, and exclude training overlap identically for every model. Sample/source disjointness and peptide disjointness are separate checks; missing lineage cannot be treated as proof of separation. The maintained release holdout manifests are an input exclusion mechanism, not a guarantee that an arbitrary public comparator has never seen the same data.

Report end-to-end presentation, affinity-only ranking and processing-only ranking separately. Never average their AP values. For multi-patient results, report per-patient metrics and paired patient-bootstrap intervals. See Evaluating trained models for score orientation, common-row joins and output formats.

Calibration and compatibility

Raw affinity is in nM, lower is stronger. Processing and presentation scores range from zero to one, higher is stronger. Percentile ranks range from zero to 100, lower is stronger; they are background ranks, not calibrated biological probabilities.

New calibrations use the shared compact method: 64 knots, increasing to 128 only when independent, label-free background validation justifies it. Affinity uses log(IC50); processing and presentation use logit(score). Existing public histogram calibrations load without refitting or conversion. See Percentile calibration for all predictors for the exact rules and compatibility.

The release workflow calibrates affinity and presentation. Standalone processing calibration is a separate command with an explicit background and flank policy. Evaluating raw scores does not validate a percentile API: check both saved score representations before publishing weights.

Reproduce a run

Use the maintained release entry point and explicit recipe/data settings; see the release workflow and Reproduce the 2.3.0 model training for launch and collection commands. Freeze the source commit, seed, data hashes and complete settings before training. Preserve the resulting release_provenance.json, holdout manifests, training tables, selection records and comparison outputs.

Worker counts and prediction chunks may be autosized. Training minibatches are scientific settings and must not be silently reduced to fit a device. The release path uses eager execution and highest float32 matmul precision.