# Additive ridge assay regression v1 ## Intended use Rank already proposed amino-acid substitution variants of one parent protein for follow-up measurements in one unchanged assay. The first research use case is small protein-engineering groups prioritizing binding-reagent variants. This model is a transparent starting point for a pilot, not a validated general protein-function predictor. It is not a disease, therapeutic, safety, or clinical model. It does not generate proteins. ## Implementation and license `deltafold/variants.py`, model ID `additive-ridge-v1`, is original DeltaFold code under this repository's MIT license. There are no pretrained model weights, downloaded inference dependencies, or inference-time network requests. Each fit uses only the project's measured sequences and values. No customer dataset is automatically reused for another customer or for the public benchmark. The objective is sum of squared residuals plus `10 * sum(beta**2)`, with an unpenalized intercept. Each training-observed non-parent amino acid at a position is an indicator feature. Cyclic coordinate descent runs for at most 200 sweeps with tolerance `1e-7 * (max(y)-min(y))` for nonconstant outcomes. Features are learned from the training split; target values are not transformed. The public benchmark's alpha, samples, and splits were fixed before its first execution. The result retains coefficients, intercept, observed residues, training mutation range, model version, parameters, convergence, source description, assay metadata, training IDs, and a SHA-256 digest of the normalized input snapshot. Saved job snapshots supply the corresponding original inputs. This is sufficient to inspect and reconstruct a fit without opaque weights. ## Inputs and constraints - One parent sequence of 4–2,000 canonical amino acids; 1–500 uniquely identified variant sequences of the same length, with at most 20 variable positions. - Strict FASTA with unique IDs. Duplicate sequences under different IDs are rejected; record repeated assays through replicate rows instead. - CSV columns `variant_id,value,replicate,round`; optional `endpoint,unit,condition` must exactly match the project's immutable metadata. Import errors reject the entire input rather than silently dropping records. - No missing, non-finite, Boolean, or censored (`