PaxBench

This dataset provides the Paxbench data artifacts for cross-species protein abundance prediction.

Dataset Summary

The benchmark contains 88,337 protein abundance observations across seven species. The target is log10_abundance, a log10-transformed whole-organism integrated protein abundance value. Each observation includes a protein sequence and, where available, CDS, mRNA, and cDNA sequence fields.

Species included:

• Arabidopsis thaliana
• Mus musculus
• Homo sapiens
• Drosophila melanogaster
• Saccharomyces cerevisiae
• Schizosaccharomyces pombe
• Escherichia coli

Files

• paxbench_tracka.parquet: main benchmark table in Parquet format.
• paxbench_tracka.csv.gz: compressed CSV version of the main benchmark table.
• paxbench_tracka.pkl: local pickle copy of the main benchmark table.
• metadata/dataset_summary.json: dataset-level summary statistics, species counts, split counts, sequence coverage, and CDS repair summary.
• metadata/schema.tsv: column schema with field names, non-null counts, and dtypes.
• metadata/id_mapping.tsv.gz: mapping from public paxbench_id values to source identifiers and source metadata.
• metadata/cds_terminal_trim_report.tsv: report for CDS terminal trimming repairs.
• results/: compact summary outputs for embedding, LoRA, and zero-shot experiments.
• demo/paxbench_demo.tsv: tiny 30 train / 3 validation / 3 test demo table used by the companion code repository.

Main Table Columns

• paxbench_id: public row identifier used for benchmark joins and prediction files.
• species: species name for the observation.
• protein_sequence: amino-acid sequence.
• cds_sequence: coding DNA sequence, when available.
• mrna_sequence: mRNA sequence, when available.
• cdna_sequence: cDNA transcript sequence, when available.
• log10_abundance: regression target, expressed as log10-transformed integrated protein abundance.
• split_species: species holdout split.
• split_homology: homology-aware split.
• split_highlow: abundance-stratified high/low split.
• split_random: random row split.
• split_fungi2fungi: fungi-to-fungi split; non-fungal species are missing.
• split_fungirandom: random split restricted to fungal species; non-fungal species are missing.
• split_cluster: sequence-cluster split.
• split_species2: alternate species holdout split.
• split_species3: alternate species holdout split.

Split columns use train, validation, and test labels when a row belongs to that split. Some specialized split columns contain missing values for rows outside the relevant species subset.

Metadata

The metadata/ directory makes the benchmark auditable without duplicating the main sequence table.

• dataset_summary.json records dataset size, column names, species counts, split counts, sequence coverage, and the number of CDS terminal trim repairs.
• schema.tsv records each main-table field, non-null count, and dtype.
• id_mapping.tsv.gz links paxbench_id to source-level identifiers and source metadata.
• cds_terminal_trim_report.tsv lists the 40 CDS records where one or two terminal nucleotides were removed so the CDS length is divisible by three and remains consistent with the protein sequence.

Results

The results/ directory contains compact summary score tables:

• results/embedding/embedding_score.csv: embedding extraction plus ML regressor results.
• results/lora/lora_score.csv: LoRA finetuning results.
• results/zeroshot/zeroshot_score.csv: zero-shot scoring results.

These files summarize experiments and are not required to load the benchmark.

Intended Use

This dataset is intended for benchmarking sequence-derived and multimodal predictors of protein abundance, including cross-species generalization using protein, CDS, mRNA, or cDNA sequence inputs.

Limitations

The abundance labels are integrated whole-organism estimates and may reflect biological condition, assay protocol, source dataset, and integration effects. CDS and mRNA coverage is incomplete, and missingness is not random. Cross-species evaluation may reflect both biological differences and dataset-source effects.

Responsible Use

This benchmark should not be interpreted as condition-specific abundance ground truth. It is not intended for clinical, regulatory, or direct drug-discovery decision-making without further validation.

Loading

import pandas as pd

df = pd.read_parquet("paxbench_tracka.parquet")
print(df.head())

For environments without Parquet support:

import pandas as pd

df = pd.read_csv("paxbench_tracka.csv.gz")
print(df.head())