docs/data_dictionary.md

Data dictionary

data/processed/omniflic_nkp46.csv and omniclic_nkp46.csv share the same 30-column schema. Each row represents a unique VH/VL sequence combination within one platform.

Field	Meaning
sample_id	Platform plus a hash of the VH/VL sequences; key used to align data and embeddings
clone_id	Selected original clone identifier; identical OmniFlic sequences prefer Rat_, while OmniClic excludes _v2 and the 50 / 0050_v2 group
VH	Complete heavy-chain variable-domain amino acid sequence
VL	Fixed light-chain variable-domain sequence for the platform
KD_M	Original SPR equilibrium dissociation constant of the selected record, in M
pKD	Regression label, -log10(KD_M)
platform	OmniFlic or OmniClic
measurement_count	Number of original measurements used for the label; one for every released sample
pKD_min, pKD_max	Range of retained pKD measurements; both equal pKD in the current release
source_table	Patent table containing the source KD
vh_seq_id, vl_seq_id	SEQ IDs in the filed ST.26 XML; multiple IDs, when present, are separated by `
CDRH1, CDRH2, CDRH3	VH CDR amino acid sequences under IMGT numbering
V_gene, J_gene	Closest amino acid reference gene assignments; tied candidates separated by ;
V_allele_candidates, J_allele_candidates	AntPack's tied allele candidates in the original output format
V_identity, J_identity	Amino acid identity fractions for the inferred V/J matches
numbering_identity	Identity fraction for AntPack's numbering alignment
annotation_warning	Numbering-tool warning; empty when no warning is present
numbering_imgt	JSON array mapping each VH residue to its IMGT position, including insertion codes
annotation_source	Annotation tool/version, numbering scheme, and reference species
numbering_start, numbering_end	IMGT positions of the first and last VH residues; 1 and 128 in the main datasets
family_id	Connected-component ID under the configured V/J, CDRH3-length, and edit-distance rules
fold_id	Fixed outer test-fold ID, 0-4; a family belongs to one fold. The remaining four folds directly serve as inner validation folds

Hugging Face metadata explicitly stores clone_id and vh_seq_id as strings. When reading local CSVs with pandas, use dtype={"clone_id": str, "vh_seq_id": str} to preserve identifiers as text.

Source and exclusion tables

data/raw/measurements.csv retains all 415 original clone records. clone_to_sample.csv adds the corresponding sample_id, original pKD, the included flag, and exclusion_reason. An excluded record can share sample_id with a selected record while retaining included=false. Both tables preserve patent sequence IDs, PDF pages, OCR coordinates, and original KD text.

excluded_measurements.csv contains eight OmniClic _v2 records (excluded_clone_id_suffix) and 31 OmniFlic numeric-ID measurements replaced by corresponding Rat_ records (superseded_by_preferred_clone_id_prefix). The latter cover 30 sequence groups; 18464 and 9468 are both replaced by Rat_108.

The released OmniFlic task contains 138 Rat_ records and 75 numeric-ID records. OmniClic contains 161 non-_v2 records. Every label uses its selected record's original KD.

excluded.csv contains two excluded deduplicated sequences and their annotations: Gen_1, whose VH lacks the IMGT framework terminus (incomplete_VH_IMGT_terminal_positions), and OmniClic 50 (excluded_sequence_group). The latter records the removal of the complete 50 / 0050_v2 group; _v2 is also recorded in excluded_measurements.csv. Sequence-group exclusions are applied after partition assignment so the remaining samples retain their family and fold IDs.

Sequence and embedding files

FASTA headers match the CSV sample_id values. A custom embedding NPZ contains a one-dimensional string array sample_ids and an N x D numeric matrix embeddings; the evaluator aligns rows by sample_id.