FLIP2

FLIP2 is the second-generation Fitness Landscape Inference for Proteins benchmark, released as a bioRxiv preprint in February 2026 by a group spanning NVIDIA, Microsoft, Caltech, Profluent, and Duke. It picks up where the original FLIP (Dallago et al., NeurIPS 2021) left off, that benchmark was confined to thermostability (Meltome), GB1 binding, and AAV capsid viability, and the field had moved past it. FLIP2 expands the scope to seven new datasets covering enzymes, protein-protein interactions, and light-sensitive proteins, so that protein language model and supervised fitness predictor evaluations cover the kinds of targets people actually engineer. The central design idea is the same as FLIP1 but pushed harder: provide standardized train, validation, and test splits that explicitly probe generalization under realistic protein engineering distribution shifts (sequence-similarity-based splits, low-data regimes, extrapolation to higher mutational distance, and held-out interaction partners) rather than random splits. Each dataset ships with multiple split protocols, sequences, fitness labels, and reference baselines (best zero-shot pLM likelihood per split, ridge regression on pLM embeddings, etc.), all in a uniform format compatible with the existing FLIP tooling at benchmark.protein.properties. The intent is that FLIP2 sits alongside ProteinGym (which leans on DMS landscapes and clinical labels) as the protein-engineering-oriented half of the standard evaluation stack.

Splits

• train: 801,710 rows
• test: 88,646 rows

Rows are assigned with a deterministic hash split: sha256(record_id) % 10, where bucket 0 is test and buckets 1-9 are train.

Columns

The table includes common browsing and modeling columns:

• record_id: stable SHA-256 row identifier
• dataset_id, source_file, source_table, source_row_index: source provenance
• task_name, subtask_name, assay_name: FLIP2 task metadata
• sequence, sequence_length, target, score_value, label: normalized convenience fields
• split_bucket: deterministic split bucket

The original FLIP2 fields are preserved as snake_case string columns: raw_sequence, set, raw_target, and validation. See metadata/column_mapping.parquet for the original field-name mapping and metadata/source_tables.parquet for per-table row counts.

Usage

from datasets import load_dataset

ds = load_dataset("LiteFold/FLIP2")
print(ds)
print(ds["train"][0])

Load only the common columns:

from datasets import load_dataset

cols = ["record_id", "task_name", "subtask_name", "sequence", "score_value", "label"]
train = load_dataset("LiteFold/FLIP2", split="train", columns=cols)

Filter to a task:

from datasets import load_dataset

train = load_dataset("LiteFold/FLIP2", split="train")
trpb = train.filter(lambda row: row["task_name"] == "trpb")

Metadata tables can be loaded directly:

from datasets import load_dataset

source_tables = load_dataset(
    "parquet",
    data_files="hf://datasets/LiteFold/FLIP2/metadata/source_tables.parquet",
    split="train",
)
column_mapping = load_dataset(
    "parquet",
    data_files="hf://datasets/LiteFold/FLIP2/metadata/column_mapping.parquet",
    split="train",
)

Citation

@article{didi2026flip2,
  title     = {{FLIP2}: Expanding Protein Fitness Landscape Benchmarks for Real-World Machine Learning Applications},
  author    = {Didi, Kieran and Alamdari, Sarah and Lu, Alex X. and Wittmann, Bruce and Johnston, Kadina E. and Amini, Ava P. and Madani, Ali and Czeneszew, Maya and Dallago, Christian and Yang, Kevin K.},
  journal   = {bioRxiv},
  year      = {2026},
  publisher = {Cold Spring Harbor Laboratory},
  doi       = {10.64898/2026.02.23.707496},
  url       = {https://www.biorxiv.org/content/10.64898/2026.02.23.707496v1}
}