ImmunocanAI/CLC-mini-bench

CLC mini benchmark

Version 0.6.0. Two sequence-only NKp46 binding-affinity regression tasks for evaluating frozen common-light-chain antibody language model representations. The datasets, fixed folds, and evaluation code are prepared; model evaluations have not been run.

Code and evaluation instructions | Protocol | Data dictionary

Dataset overview

Each row is a unique (platform, VH, VL) combination with an experimentally measured SPR KD. Each platform uses one fixed VL. Evaluate OmniFlic and OmniClic separately.

Configuration	Source clone records	Benchmark samples	Families	Fixed VL	Test samples in folds 0 / 1 / 2 / 3 / 4
omniflic	245	213	27	SEQ ID 2214	63 / 34 / 33 / 43 / 40
omniclic	170	161	32	SEQ ID 2217	27 / 28 / 43 / 28 / 35

The regression label is pKD = -log10(KD_M), with KD in molar units. Higher pKD means stronger binding. In nM, the OmniFlic KD range is 2.6-1000 with median 36.9; the OmniClic range is 0.0739-1220 with median 14.4.

Load the data

Install datasets and huggingface_hub. If authentication is required, run hf auth login with an account that has repository access.

from datasets import load_dataset

omniflic = load_dataset("ImmunocanAI/CLC-mini-bench", "omniflic", split="data")
omniclic = load_dataset("ImmunocanAI/CLC-mini-bench", "omniclic", split="data")

print(len(omniflic), len(omniclic))  # 213, 161

# Outer fold 0: use the other four folds as the training pool.
training_pool = omniflic.filter(lambda row: row["fold_id"] != 0)
test = omniflic.filter(lambda row: row["fold_id"] == 0)

# One inner split: existing fold 1 is validation; folds 2, 3, 4 are train.
inner_train = training_pool.filter(lambda row: row["fold_id"] != 1)
inner_val = training_pool.filter(lambda row: row["fold_id"] == 1)

Both configurations expose a single data split containing all samples. The provided fold_id defines evaluation membership. Pass revision=<Hub commit> to pin a release.

Download the complete dataset and matching reproducibility files:

from huggingface_hub import snapshot_download

snapshot_download(
    "ImmunocanAI/CLC-mini-bench",
    repo_type="dataset",
    local_dir="CLC-mini-bench",
)

Record selection and labels

• OmniFlic: for all 30 groups with identical VH and VL, use the original Rat_ measurement. This supersedes 31 numeric-ID measurements, including two numeric IDs in one group. The final task has 138 Rat_ samples and 75 numeric-ID samples.
• OmniClic: remove all eight _v2 records and the complete 50 / 0050_v2 sequence group. Its original KD values are 4.17 and 0.327 nM, a 12.75-fold difference. The 161 released samples use their original non-_v2 measurements.
• Completeness: remove Gen_1, whose published VH ends at IMGT position 121 and lacks positions 122-128.

Each released label uses one original measurement, so measurement_count is one and pKD_min = pKD_max = pKD. The raw table retains all 415 source records, with inclusion decisions available in clone_to_sample.csv.

Evaluation protocol

1. Freeze the language model and mean-pool final-layer VH residue embeddings, excluding padding and special tokens. Follow each model's native sequence input convention.
2. Use the five supplied outer folds. In each round, hold out one fold as test.
3. Directly reuse the remaining four outer folds for inner cross-validation: one validation fold and three training folds per inner round.
4. Fit StandardScaler -> Ridge(solver="svd"). Select alpha from [0.001, 0.01, 0.1, 1, 10, 100, 1000, 10000] by the unweighted mean of the four validation MAEs in pKD units. Fit the scaler separately inside each training split.
5. Refit the selected pipeline on the complete outer training pool and predict the held-out test fold.
6. Repeat across all five outer folds. Report pooled out-of-fold Spearman correlation and pKD MAE, together with per-fold results.

The prepared comparators are CLC-LM, ESM-2, and a 21-feature baseline consisting of VH amino acid frequencies and CDRH3 length. They use the same folds and Ridge protocol, with independently fitted predictors for each representation. A CLC-LM checkpoint is connected through the code's adapter or NPZ embedding interface when evaluation starts.

Families are connected components under shared human V/J gene candidates, identical CDRH3 length, and CDRH3 edit distance at most two. Named sequence-group exclusions are applied after family and fold assignment, preserving the remaining samples' existing family IDs and fold memberships. AntPack 0.3.8 supplies IMGT numbering and amino-acid-based V/J assignments. Whole families are assigned to folds using seed 42, balancing sample counts without affinity labels. Family membership is disjoint between train, validation, and test within every split.

Files and fields

The two main CSV files share an explicitly typed 30-column schema. Core fields are:

Field	Meaning
sample_id	Stable key derived from platform and the VH/VL sequences
clone_id	Selected original clone identifier, stored as a string
VH, VL	Complete heavy-chain variable sequence and the platform's fixed light-chain variable sequence
KD_M, pKD	Original SPR KD in M and the transformed regression label
platform	OmniFlic or OmniClic
V_gene, J_gene, CDRH3	Inferred gene assignments and IMGT CDRH3 sequence
family_id, fold_id	Sequence-family identifier and fixed outer test-fold ID, 0-4
source_table, vh_seq_id, vl_seq_id	Patent table and filed sequence-listing identifiers

The full data dictionary describes all fields. FASTA files use sample_id as the sequence identifier. excluded_measurements.csv records the eight excluded _v2 records and 31 superseded numeric-ID records; excluded.csv records the incomplete Gen_1 VH and the excluded OmniClic 50 sequence group. manifest.json contains processing rules, annotation metadata, counts, and file SHA-256 values. RELEASE.json identifies the matching Git source revision.

Provenance and preparation checks

The source is OmniAb patent WO2025137555A1 and the filed WIPO ST.26 XML sequence listing. Complete VH and KD are mapped from tables 4/5 for OmniFlic and 7/8 for OmniClic. Source files, OCR records, manual corrections, and verification records are included in data/raw/.

All 415 clone-to-VH mappings were verified, and KD extraction was checked at two OCR resolutions. Preparation passes 17 tests and an independent rebuild of all nine processed CSV/FASTA files. All 40 inner splits reuse the original fold members and have pairwise-disjoint train/validation/test family sets. See validation records.

Version 0.6.0 removes the 50 / 0050_v2 OmniClic sequence group. This removes one additional benchmark sample relative to version 0.5.0, reducing OmniClic from 162 to 161 samples. OmniFlic and all remaining OmniClic rows retain their previous labels, family IDs, and fold assignments. The protocol continues to use five outer folds and four inner folds. Evaluation entry points and output formats are documented in the GitHub README.