Protein Benchmark · LiteFold
FLIP2
A protein engineering benchmark for evaluating fitness prediction models under realistic distribution shifts.
What FLIP2 is
FLIP2 (Fitness Landscape Inference for Proteins 2) is a benchmark designed to evaluate protein language models and supervised fitness predictors against realistic protein engineering objectives. Created by a collaboration spanning NVIDIA, Microsoft, Caltech, Profluent, and Duke, it succeeds the original FLIP benchmark by expanding beyond early targets—which were limited to thermostability, GB1 binding, and AAV capsid viability—into enzymes, protein-protein interactions, and light-sensitive proteins.
Conventional machine learning evaluation often relies on random data partitioning, which inflates performance estimates compared to real laboratory campaigns. FLIP2 addresses this limitation by formalising distribution shifts that reflect practical protein design challenges, including extrapolation to higher mutational counts, low-data training regimes, sequence similarity thresholds, and held-out binding partners.
Distributed via LiteFold in unified Parquet tables compatible with earlier FLIP evaluation tooling, the benchmark positions itself alongside clinical and deep mutational scanning resources such as ProteinGym as an evaluation standard dedicated specifically to protein engineering.
How it was made
The dataset integrates measurements across seven biological datasets into a standardised tabular format. Each record contains normalised fields including the sequence, sequence length, assay target, score value, and task identification metadata, while preserving original source identifiers, table references, and raw sequence annotations. Accompanying auxiliary tables supply explicit column mappings and source table counts.
Rows are partitioned into 801,710 training examples and 88,646 test examples using a deterministic hash allocation (sha256 of the record identifier modulo 10, reserving bucket 0 for test and buckets 1 to 9 for training). At the benchmark level, tasks are framed with multiple evaluation protocols designed to test out-of-distribution generalisation, paired with standard reference baselines including zero-shot language model likelihoods and ridge regression trained on sequence embeddings.
Key facts
- Format
- Parquet (tabular and text)
- Biological domains
- Enzymes, protein-protein interactions, and light-sensitive proteins
- Preprint release
- February 2026 (bioRxiv)
- Primary data inputs
- Amino acid sequences and experimental fitness scores
What you can use it for
- Protein language model evaluation
- Benchmarking zero-shot likelihoods and embedding representations from pretrained protein models on functional fitness prediction.
- Mutational extrapolation testing
- Assessing how supervised regression models generalize when tested on variants with higher mutational distances than those seen during training.
- Low-data fitness inference
- Measuring model accuracy and sample efficiency when predicting protein properties in data-constrained training regimes.
- Interaction partner generalization
- Evaluating the capacity of machine learning architectures to predict protein-protein binding across held-out interaction targets.
Explore the data
The schema, real rows, full-text search and a SQL filter — read in your browser, nothing downloaded.
Explore
Read the real rows without downloading anything
Reading rows…
LiveRead from LiteFold/FLIP2 at the moment you asked. Nothing is cached or stored — every row above came from that request.
How the columns are distributed
Over 801,710 rows of default/train
assay_name
string_label- trpb/by_position26%
- trpb/two_to_many26%
- trpb/one_to_many26%
dataset_id
string_label- flip2100%
entity_type
string_label- sequence100%
label
string_label- test85%
- train15%
mutation
string_labelNo distribution reported for this column.
100.0% empty
raw_sequence
string_text- min
- 57
- median
- 389
- mean
- 326.91
- max
- 425
raw_target
string_text- min
- 1
- median
- 19
- mean
- 17.90
- max
- 23
record_id
string_text- min
- 64
- median
- 64
- mean
- 64
- max
- 64
score_value
float- min
- -4.84
- median
- 6.8e-3
- mean
- 0.16
- max
- 622
sequence
string_text- min
- 57
- median
- 389
- mean
- 326.91
- max
- 425
sequence_length
int- min
- 57
- median
- 389
- mean
- 326.91
- max
- 425
set
string_label- test85%
- train15%
source_file
string_label- data/unpacked/labeled/flip2/trpb/by_position.csv26%
- data/unpacked/labeled/flip2/trpb/two_to_many.csv26%
- data/unpacked/labeled/flip2/trpb/one_to_many.csv26%
source_row_index
int- min
- 0
- median
- 79,889
- mean
- 91361.15
- max
- 228k
source_table
string_label- tables/data_unpacked_labeled_flip2_trpb_by_position.csv.jsonl26%
- tables/data_unpacked_labeled_flip2_trpb_two_to_many.csv.jsonl26%
- tables/data_unpacked_labeled_flip2_trpb_one_to_many.csv.jsonl26%
split_bucket
int- min
- 1
- median
- 5
- mean
- 5.00
- max
- 9
subtask_name
string_label- two_to_many33%
- one_to_many26%
- by_position26%
table_group
string_label- benchmark100%
target
string_text- min
- 1
- median
- 19
- mean
- 17.90
- max
- 23
task_name
string_label- trpb77%
- hydro14%
- nucb6%
validation
string_label- False97%
- True3%
Strengths and limits
Good at
- Tests realistic protein engineering distribution shifts rather than relying on uniform random splits.
- Broadens benchmark coverage over original FLIP datasets by including enzymes, interaction partners, and light-sensitive targets.
- Includes standardized baseline evaluations such as zero-shot model likelihoods and ridge regression over sequence embeddings.
- Maintains traceability with stable record identifiers, source table pointers, and raw value columns alongside normalized fields.
Watch out for
- The primary repository provides a single global hash-based split rather than baking distinct task-specific shift protocols directly into split files.
- Does not focus on human clinical variant interpretation or non-engineering deep mutational scanning phenotypes.
Questions
Can I look inside FLIP2 without downloading it?
Yes. This page reads its schema, real rows and per-column statistics in the browser, with full-text search and a SQL filter. Nothing is downloaded to your machine.
What licence is FLIP2 under?
Its authors declare other. Read the terms on the original page before reusing the data.
How does FLIP2 differ from the original FLIP benchmark?
The original FLIP was confined to AAV capsid viability, GB1 binding, and Meltome thermostability. FLIP2 broadens this scope with seven new datasets spanning enzymes, protein-protein interactions, and light-sensitive proteins to represent modern engineering targets.
What distribution shifts does FLIP2 test?
It tests generalization across sequence-similarity-based partitions, low-data regimes, extrapolation to higher mutational distances from the wild type, and prediction on held-out interaction partners.
How are records split between train and test sets in this repository?
Rows are split using a deterministic hash calculation, sha256(record_id) % 10. Bucket 0 forms the test split (88,646 rows) and buckets 1 through 9 form the training split (801,710 rows).
What baselines are provided for comparison?
FLIP2 supplies standardized reference baselines, including zero-shot protein language model likelihood estimates and ridge regression models trained on sequence embeddings.
How to cite it
@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}}
Written from FLIP2's own dataset card. Statements citing a figure that does not appear there are removed before publication. Licence and running details come from our registry.