BioXArena-Data-Public-XL

This directory is the public data package for BioXArena-XL. It contains the public inputs, task descriptions, and sample submission templates used by agents during development and inference. Private labels are stored separately in the matching BioXArena-Data-Private-XL package.

Based on the current folder contents, this package contains:

• 6 domains
• 39 task folders
• 39 public/description.md files
• 39 public/sample_submission.csv files
• 39 public/train.csv files
• 39 public/test.csv files

Each task follows the same top-level layout:

BioXArena-Data-Public-XL/
`-- <domain>/
    `-- <task>/
        `-- public/
            |-- description.md
            |-- sample_submission.csv
            |-- train.csv
            |-- test.csv
            `-- modality-specific public assets

Within each task, the public/ directory includes the task description, sample submission template, train/test manifests or tables, and any modality-specific public assets needed for that task. The exact file set varies by task: some tasks are table-based, while others include microscopy images, segmentation inputs, perturbation-response features, cryo-EM volumes, molecular or structure features, sequence metadata, clinical covariates, or multimodal omics features.

Download From Hugging Face

This package is intended to be distributed on Hugging Face at:

• https://huggingface.co/datasets/mbzuai-ai4bio/BioXArena-Data-Public-XL

Users can download and extract it like this:

wget "https://huggingface.co/datasets/mbzuai-ai4bio/BioXArena-Data-Public-XL/resolve/main/BioXArena-Data-Public-XL.tar.gz" -O BioXArena-Data-Public-XL.tar.gz
tar -xzf BioXArena-Data-Public-XL.tar.gz

You can also use huggingface-cli:

huggingface-cli download mbzuai-ai4bio/BioXArena-Data-Public-XL BioXArena-Data-Public-XL.tar.gz --repo-type dataset --local-dir .
tar -xzf BioXArena-Data-Public-XL.tar.gz

Domain Summary

Domain	# Tasks
chemical-biology	7
imaging	7
network-biology	4
sequence-genomics	6
single-cell-perturbation	6
structure	9

What Makes BioXArena-XL Different

BioXArena-XL is a curated larger-scale variant of BioXArena with fewer but heavier and more realistic tasks. The XL signal comes from several dimensions:

• larger cohorts, more samples, or more test items per task;
• broader source coverage, such as multiple upstream datasets, species, cell lines, targets, assay families, tissues, or cancer cohorts;
• harder generalization splits, such as held-out drugs, perturbations, targets, complexes, sources, or OOD chemistry;
• file-manifest submissions where submission.csv references generated arrays, masks, or volumes instead of storing all predictions as scalar CSV fields;
• grouped primary metrics that evaluate performance by source, target, complex, tissue, cancer cohort, cell line, or PDB entry.

Comparison Scope

The comparison below matches domains one-to-one when the domain name is unchanged. The only intentional domain remappings are:

• sequence -> sequence-genomics
• single-cell + perturbation-dynamics -> single-cell-perturbation

Original BioXArena domains without an explicit XL counterpart are not included in these domain-wise comparison tables.

Train/Test Sample-Scale Comparison

The table below compares domains by average train/test rows per task, rather than total rows, because BioXArena and BioXArena-XL do not always have the same number of tasks in each domain. Counts are taken from the current local train.csv and test.csv files. For file-manifest tasks, each row corresponds to one requested prediction item.

Original domain	Original tasks	Original avg train/task	Original avg test/task	XL domain	XL tasks	XL avg train/task	XL avg test/task	XL scale signal
chemical-biology	8	8,721	2,164	chemical-biology	7	71,039	17,542	Larger molecular property and interaction datasets, especially ChemixHub and DTI.
imaging	8	5,213	1,501	imaging	7	15,944	3,986	Larger image and multimodal cohorts, with high-row localization and pathology tasks.
network-biology	8	4,618	1,149	network-biology	4	863,856	233,295	Much larger graph/regulatory-network edge sets, dominated by causal GRN prediction.
sequence	10	169,212	38,469	sequence-genomics	6	52,843	13,109	Broader genomics/oligo-design task mix; original sequence includes very large RNA-binding rows.
single-cell + perturbation-dynamics	18	50,925	11,169	single-cell-perturbation	6	57,755	14,608	Comparable-to-larger per-task scale after consolidating into fewer XL perturbation and atlas tasks.
structure	8	46,267	11,524	structure	9	69,157	13,137	More structures, conformers, decoys, peptide affinity, and cryo-EM file predictions.

Per-Task Train/Test Rows

chemical-biology -> chemical-biology

Dataset	Task	Train rows	Test rows
Original	chemical-biology/bace1-binding-affinity	6,487	1,622
Original	chemical-biology/cell-painting-perturbation	6,087	1,522
Original	chemical-biology/cyp-inhibition-multi-label	1,476	380
Original	chemical-biology/egfr-binding-affinity	8,691	2,127
Original	chemical-biology/gpcr-binding-multi-class	4,520	1,131
Original	chemical-biology/herg-binding-affinity	7,689	1,886
Original	chemical-biology/kinase-selectivity-multi-label	30,158	7,482
Original	chemical-biology/tox21-sr-are	4,660	1,165
XL	chemical-biology/boom-xl-csd	17,549	2,863
XL	chemical-biology/chemcot-xl-molecule-counting	2,487	622
XL	chemical-biology/chemixhub-xl-conductivity	37,199	9,300
XL	chemical-biology/chemixhub-xl-viscosity	253,022	63,256
XL	chemical-biology/solubility-xl	7,986	1,996
XL	chemical-biology/tdc-xl-cyp-inhibition	30,498	7,625
XL	chemical-biology/tdc-xl-dti-binding-affinity	148,531	37,132

imaging -> imaging

Dataset	Task	Train rows	Test rows
Original	imaging/amos-organ-segmentation	288	72
Original	imaging/drug-moa-prediction	944	592
Original	imaging/labelfree-cell-counting	3,727	1,512
Original	imaging/lung-nodule-malignancy	637	140
Original	imaging/mitochondria-counting	5,596	987
Original	imaging/nucleus-type-classification	5,179	2,722
Original	imaging/skin-lesion-diagnosis	8,035	1,980
Original	imaging/virtual-staining	17,295	4,000
XL	imaging/cellsam-xl-segmentation	5,698	1,324
XL	imaging/cryocrab-xl-quality	640	157
XL	imaging/drugclass-xl-if	10,289	2,571
XL	imaging/starc9-xl-crc-tissue	16,000	4,000
XL	imaging/subcell-xl-localization	75,093	18,879
XL	imaging/survpath-xl-multimodal	1,945	487
XL	imaging/survpath-xl-rna	1,945	487

network-biology -> network-biology

Dataset	Task	Train rows	Test rows
Original	network-biology/gene-disease-association	6,001	1,499
Original	network-biology/go-function-multi-label	2,806	702
Original	network-biology/metabolic-network-kegg	4,001	999
Original	network-biology/pathway-membership-reactome	4,801	1,199
Original	network-biology/ppi-prediction-string	7,228	1,772
Original	network-biology/protein-complex-corum	2,103	526
Original	network-biology/synthetic-lethality-prediction	4,801	1,199
Original	network-biology/tf-regulatory-prediction	5,203	1,297
XL	network-biology/causalbench-xl-grn	3,290,506	822,318
XL	network-biology/music-xl-seccomplex	57,359	3,607
XL	network-biology/pring-xl-cross-species-ppi	107,280	107,184
XL	network-biology/scigym-xl-sbml	280	70

sequence -> sequence-genomics

Dataset	Task	Train rows	Test rows
Original	sequence/gene-tissue-expression	272,000	68,000
Original	sequence/isoform-expression	179,606	6,555
Original	sequence/multi-tf-binding	38,124	9,532
Original	sequence/protein-protein-interaction	82,744	20,686
Original	sequence/regulatory-element-detection	40,000	10,000
Original	sequence/remote-homology-detection	80,000	20,000
Original	sequence/rna-protein-binding-affinity	39,321	9,831
Original	sequence/rna-protein-binding-signal	930,686	232,672
Original	sequence/rna-reactivity-imputation	5,643	1,411
Original	sequence/variant-effect-pathogenicity	24,000	6,000
XL	sequence-genomics/aso-xl-knockdown-efficacy	20,522	4,095
XL	sequence-genomics/cgbench-xl-variant	7,760	1,940
XL	sequence-genomics/dnalongbench-xl-regulatory	24,425	7,349
XL	sequence-genomics/escape-xl-amp-multilabel	65,870	16,489
XL	sequence-genomics/neurotox-xl-aso	3,369	841
XL	sequence-genomics/sirna-xl-knockdown-efficacy	195,111	47,940

single-cell + perturbation-dynamics -> single-cell-perturbation

Dataset	Task	Train rows	Test rows
Original	single-cell/batch-integration	228,948	72,848
Original	single-cell/cell-type-from-expression	3,252	813
Original	single-cell/chromatin-to-expression	57,614	11,635
Original	single-cell/cite-seq-protein-prediction	76,161	14,100
Original	single-cell/cross-modality-cell-matching	9,323	1,647
Original	single-cell/cross-modality-cell-type	128,727	32,182
Original	single-cell/developmental-stage-prediction	68,676	12,120
Original	single-cell/gene-expression-denoising	3,605	3,605
Original	single-cell/label-projection	30,159	3,347
Original	single-cell/rna-to-protein-prediction	66,175	1,000
Original	perturbation-dynamics/cancer-drug-sensitivity	203,972	35,643
Original	perturbation-dynamics/crispr-perturbation-prediction	8,200	1,280
Original	perturbation-dynamics/drug-transcriptional-response	1,252	1,008
Original	perturbation-dynamics/eccite-multimodal-perturbation	13,758	4,587
Original	perturbation-dynamics/gene-regulatory-network-inference	13	14
Original	perturbation-dynamics/multi-timepoint-perturbation	3,997	698
Original	perturbation-dynamics/rna-velocity-cell-transition	2,594	1,102
Original	perturbation-dynamics/spear-atac-perturbation	10,216	3,406
XL	single-cell-perturbation/cellverse-xl-cta	1,366	342
XL	single-cell-perturbation/perturbench-xl-response	387	97
XL	single-cell-perturbation/scgenescope-xl-treatment	214,783	53,694
XL	single-cell-perturbation/tahoe-xl-drug-response	52,865	13,378
XL	single-cell-perturbation/tahoe-xl-single-cell	74,880	19,577
XL	single-cell-perturbation/xatlas-orion-xl-perturb	2,247	562

structure -> structure

Dataset	Task	Train rows	Test rows
Original	structure/complex-structure-evaluation	8,863	2,216
Original	structure/enzyme-commission-prediction	17,273	1,918
Original	structure/protein-binding-site-detection	34,353	8,595
Original	structure/protein-fold-classification	13,085	3,174
Original	structure/protein-ligand-binding-affinity	2,679	1,239
Original	structure/protein-protein-interface	33,967	8,492
Original	structure/protein-stability-change	258,552	66,215
Original	structure/protein-structure-prediction	1,363	341
XL	structure/cpsea-xl-cyclic-peptide-affinity	70,284	1,583
XL	structure/cpsea-xl-cyclic-peptide-affinity-v2	70,284	1,583
XL	structure/davis-complete-xl-dti	20,332	5,304
XL	structure/decoydb-xl-pose-classify	68,076	17,578
XL	structure/decoydb-xl-pose-rmsd	68,076	17,578
XL	structure/denoise-cryo-xl	232	58
XL	structure/proteinconformers-xl-mqa	310,608	70,945
XL	structure/proteinconformers-xl-mqa-v2	1,400	340
XL	structure/psbench-xl-multimer-mqa	13,121	3,260

Disk Storage Comparison

Disk size is one visible dimension of XL, but it is not the only one. Some XL domains are much larger on disk, while others use compact CSV/NPZ representations of broader upstream datasets. The table below reports a local snapshot measured with du -sh in this workspace.

Original domain(s)	Original public size	Original tasks	XL domain	XL public size	XL tasks	XL storage/source rationale
chemical-biology	59M	8	chemical-biology	195M	7	Larger and broader chemistry sources: CSD OOD molecular properties, ChemixHub mixtures, TDC multi-source CYP/DTI, and molecule-counting tasks.
imaging	37G	8	imaging	30G	7	Larger and more multimodal imaging tasks: TissueNet segmentation, cryo-EM micrographs, IF microscopy, histopathology, subcellular localization, and TCGA imaging/RNA survival tasks.
network-biology	18M	8	network-biology	1.2G	4	Much larger network/omics sources: Perturb-seq causal edges, SEC-MS co-complex evidence, cross-species PPI, and SBML model completion.
sequence	1.8G	10	sequence-genomics	63M	6	More genomics-oriented tasks: ASO/siRNA efficacy, clinical variant interpretation, long-range regulatory links, AMP multilabel activity, and ASO neurotoxicity. Compact public tables hide broader source coverage.
single-cell + perturbation-dynamics	6.1G + 346M	18	single-cell-perturbation	1.2G	6	Consolidated perturbation and single-cell tasks with larger upstream cohorts represented compactly: CellVerse, PerturBench, Tahoe, scGeneScope, and X-Atlas/Orion.
structure	46G	8	structure	115G	9	Strongest disk-scale XL expansion: cyclic peptide affinity, DAVIS-complete variants, DecoyDB poses, DenoiseCryo volumes, ProteinConformers MQA, and PSBench multimer quality.

Overall package size in this local snapshot:

Package	Public size	Domains	Tasks
BioXArena-Data-Public	105G	9	76
BioXArena-Data-Public-XL	147G	6	39

Task Catalog

Chemical Biology

Task Folder	Task Title	Primary Metric	XL Rationale
boom-xl-csd	BOOM-XL-CSD: Out-of-Distribution Molecular Property Prediction (CSD)	Per-source mean Pearson r on OOD rows	Larger OOD chemical-property benchmark with density and heat-of-formation sources.
chemcot-xl-molecule-counting	ChemCoT-XL: Molecular Substructure Counting	Per-source mean Pearson r	Broader molecular reasoning task over functional groups and ring counts.
chemixhub-xl-conductivity	Chemixhub-XL-Conductivity: Mixture Ionic Conductivity Prediction	Per-source mean Pearson r	Mixture-level property prediction with temperature, pressure, composition, and molecular-weight features.
chemixhub-xl-viscosity	Chemixhub-XL-Viscosity: Mixture Log Viscosity Prediction	Per-source mean Pearson r	Larger mixture-property setting with multi-component chemistry and physical conditions.
solubility-xl	Solubility-XL: Aqueous Solubility Prediction	Pearson r	Larger solubility regression set with standardized SMILES-based prediction.
tdc-xl-cyp-inhibition	TDC-XL-CYP-Inhibition: CYP P450 Drug Metabolism Inhibition	Per-source mean ROC-AUC	Broader multi-isozyme CYP inhibition prediction across TDC-style sources.
tdc-xl-dti-binding-affinity	TDC-XL-DTI-Binding-Affinity: Drug-Target Binding Affinity	Per-source mean Pearson r	Broader DTI benchmark spanning multiple affinity sources, drugs, and protein targets.

Imaging

Task Folder	Task Title	Primary Metric	XL Rationale
cellsam-xl-segmentation	CellSAM-XL: Cell + Nucleus Binary Segmentation	Per-source mean Dice	Larger tissue segmentation benchmark with one output mask file per test row.
cryocrab-xl-quality	CryoCRAB-XL: Cryo-EM Micrograph Quality Regression	Pearson r	Cryo-EM micrograph quality prediction with microscope metadata and image paths.
drugclass-xl-if	DrugClass-XL: Drug Treatment Classification from IF Microscopy	Macro-F1	Larger IF microscopy classification over drug-treated fields of view.
starc9-xl-crc-tissue	STARC-9-XL: Colorectal Cancer Tissue Classification	Macro-F1	Histopathology tissue classification across 9 colorectal cancer tissue classes.
subcell-xl-localization	SubCell-XL: Subcellular Localization Multi-Label Classification	0.5 macro-F1 + 0.5 mAP	More labels and multilabel image predictions over subcellular localization categories.
survpath-xl-multimodal	SurvPath-XL-Multimodal: TCGA Survival from WSI + RNA + Clinical	Per-cancer mean C-index	Multimodal survival prediction combining histology, RNA, and clinical features.
survpath-xl-rna	SurvPath-XL-RNA: TCGA Survival from Bulk RNA + Clinical	Per-cancer mean C-index	Larger cohort survival task with cancer-specific grouped C-index.

Network Biology

Task Folder	Task Title	Primary Metric	XL Rationale
causalbench-xl-grn	CausalBench-XL: Gene-Regulatory Edge Classification	Per-source mean ROC-AUC	Perturb-seq-derived causal edge prediction across cell-line sources.
music-xl-seccomplex	MuSIC-XL-SECComplex: Protein Co-Complex Prediction from SEC-MS	ROC-AUC	Large SEC-MS protein-pair evidence matrix for co-complex classification.
pring-xl-cross-species-ppi	PRING-XL: Cross-Species PPI Prediction	Per-source mean ROC-AUC	Broader PPI prediction across multiple species.
scigym-xl-sbml	SciGym-XL: SBML Model Completion	Pearson r on delta_reactions	Systems-biology model completion using SBML structural counts.

Sequence Genomics

Task Folder	Task Title	Primary Metric	XL Rationale
aso-xl-knockdown-efficacy	ASO-XL: Antisense Oligonucleotide Knockdown Efficacy	Per-source mean Pearson r	Larger oligonucleotide efficacy prediction with source-aware evaluation.
cgbench-xl-variant	CGBench-XL: Clinical Variant Pathogenicity	Macro-F1	Clinical variant interpretation moved into genomics with a 5-class imbalanced label space.
dnalongbench-xl-regulatory	DNALongBench-XL: Enhancer-Target-Gene + eQTL Classification	Per-source mean ROC-AUC	Broader long-range regulatory benchmark across 12 sources.
escape-xl-amp-multilabel	ESCAPE-XL: Multilabel Antimicrobial Peptide Classification	0.5 macro-F1 + 0.5 mAP	Multilabel peptide activity task over antibacterial, antifungal, antiviral, antiparasitic, and antimicrobial labels.
neurotox-xl-aso	Neurotox-XL: ASO Neurotoxicity	Per-source mean Pearson r	ASO safety-oriented regression with source-aware evaluation.
sirna-xl-knockdown-efficacy	siRNA-XL: RNA Interference Knockdown Efficacy	Per-source mean Pearson r	Broader RNAi knockdown efficacy prediction.

Single-Cell Perturbation

Task Folder	Task Title	Primary Metric	XL Rationale
cellverse-xl-cta	CellVerse-XL: Single-Cell Annotation & Drug Response from Ranked Gene Lists	Macro-F1	Consolidates annotation and response classification from ranked gene-list inputs.
perturbench-xl-response	PerturBench-XL: Perturbation-Response Prediction	Per-source mean delta-Pearson	File-manifest task predicting full mean-expression response vectors.
scgenescope-xl-treatment	scGeneScope-XL: Single-Cell Treatment Classification	Macro-F1	Larger treatment classification over 29 treatment classes and single-cell embeddings.
tahoe-xl-drug-response	Tahoe-XL-DrugResponse: Multimodal Drug Response Prediction	Per-cell-line mean Pearson r	Held-out-drug response prediction with one generated gene-response vector per test combo.
tahoe-xl-single-cell	Tahoe-XL-SingleCell: Per-Cell Multimodal MoA Prediction	Macro-F1	Per-cell mechanism-of-action prediction with drug and cell-line context.
xatlas-orion-xl-perturb	X-Atlas-Orion-XL: Genome-Wide Perturb-seq Response Prediction	Per-source mean delta-Pearson	Genome-wide perturbation-response task over large upstream X-Atlas/Orion sources.

Structure

Task Folder	Task Title	Primary Metric	XL Rationale
cpsea-xl-cyclic-peptide-affinity	CPSea-XL: Cyclic Peptide-Protein Binding Affinity	Pearson r on rosetta_dG	Larger cyclic peptide-protein affinity setting with engineered structural and docking features.
cpsea-xl-cyclic-peptide-affinity-v2	CPSea-XL-CyclicPeptideAffinity v2: Sequence-Only	Pearson r on rosetta_dG	Sequence-only variant for harder generalization without full engineered features.
davis-complete-xl-dti	DAVIS-Complete-XL: Drug-Kinase Binding with Variants	Per-kinase-base mean Pearson r	DAVIS-complete expansion with mutation/phosphorylation-aware kinase variants.
decoydb-xl-pose-classify	DecoyDB-XL-PoseClassify: Native-vs-Decoy Pose Classification	Per-complex top-k accuracy	Pose-ranking style classification over native and decoy ligand poses.
decoydb-xl-pose-rmsd	DecoyDB-XL-PoseRMSD: Protein-Ligand Pose RMSD Regression	Per-complex mean Spearman rho	Per-complex pose-quality ranking through RMSD regression.
denoise-cryo-xl	DenoiseCryo-XL: 3D Cryo-EM Atom-Type Classification & Map Denoising	Per-PDB macro-Dice	Very large 3D cryo-EM volume task with file-manifest voxel predictions.
proteinconformers-xl-mqa	ProteinConformers-XL: Model Quality Assessment	Per-target mean Pearson r	Larger conformer MQA task emphasizing within-target ranking.
proteinconformers-xl-mqa-v2	ProteinConformers-XL-MQA v2: Model Quality Assessment with 3D Structure	Per-target mean Pearson r	Adds 3D structure paths/features for harder conformer quality prediction.
psbench-xl-multimer-mqa	PSBench-XL: Protein-Complex Quality Assessment	Per-source/per-target mean Pearson r	Multimer model quality task across CASP-style sources and targets.

File-Manifest Submission Tasks

Most tasks require scalar or label predictions directly in submission.csv. Five tasks use submission.csv as a manifest pointing to generated prediction files in the output directory:

• imaging/cellsam-xl-segmentation: mask_file
• single-cell-perturbation/perturbench-xl-response: response_file
• single-cell-perturbation/tahoe-xl-drug-response: response_file
• single-cell-perturbation/xatlas-orion-xl-perturb: response_file
• structure/denoise-cryo-xl: type_label_file, bin_label_file, reg_label_file

For these tasks, agents should write the referenced files under their task output directory, typically under predictions/, and put relative paths in submission.csv.

How To Use

1. Choose a task under a domain.
2. Open that task's public/ directory.
3. Read public/description.md first.
4. Load the task-specific public inputs from the same public/ directory.
5. Generate predictions following public/sample_submission.csv.
6. Save outputs to the corresponding task output directory used by the runner.

Notes

• Public artifacts are heterogeneous across tasks. Depending on the task, public/ may contain tables, images, compressed arrays, control vectors, MRC inputs, structure features, or other modality-specific assets.
• The exact input and output expectations are task-specific, so description.md is the authoritative entry point for each task.
• sample_submission.csv defines the required submission columns, row count, and first-column ID order.
• Private labels are not included in this package. Evaluation uses the matching BioXArena-Data-Private-XL package.