Jerrychen229/PhenoAIR_MOA

PhenoAIR MoA Benchmark Data

This directory contains three mechanism-of-action benchmark tracks:

Track	Reference compounds	Test compounds	Task
moa_verified	63	200	MoA prediction in the verified label space
moa_extended	533	600	MoA prediction in the extended label space
moa_novel	63	100	Joint known-MoA prediction and novel-MoA detection

The release contains dataset records and phenotype features only. It does not contain model predictions, agent trajectories, prompts, controller rules, or experiment logs.

Layout

Each track is self-contained:

<track>/
├── reference.parquet
├── test.parquet
├── labels.json
└── features/
    ├── cellprofiler_features.npy
    ├── cellprofiler_index.parquet
    ├── cellprofiler_feature_names.json
    ├── cellclip_features.npy
    ├── cellclip_index.parquet
    └── feature_manifest.json

dataset_manifest.json describes the release-level schema and row counts. checksums.sha256 provides SHA-256 checksums for all release files.

Compound Tables

reference.parquet and test.parquet use the following columns:

Column	Type	Description
sample_id	string	Stable row identifier within a benchmark track
split	string	reference or test
compound_id	string	JUMP-CP compound identifier
smiles	string	Compound SMILES
inchi	string	Compound InChI
inchikey	string	Compound InChIKey
primary_moa	string	Primary biological MoA annotation
valid_moas	list[string]	All valid biological MoA annotations
benchmark_label	string	Label used for exact-match evaluation
benchmark_valid_labels	list[string]	Labels accepted for any-match evaluation
is_novel	bool	Whether the compound is outside the known reference label space

Compound names and target annotations are omitted from the benchmark tables to reduce direct knowledge leakage.

For moa_novel, benchmark_label is novel moa when is_novel is true. primary_moa and valid_moas retain the biological annotations for downstream error analysis.

Phenotype Features

Every *_features.npy file is a two-dimensional float32 array. Row i in a matrix corresponds exactly to row i in its *_index.parquet file, as recorded by the row_index column.

CellProfiler features are well-level, harmonized JUMP-CP profiles with variance, MAD, and intensity feature filtering. Their column names are stored in cellprofiler_feature_names.json.

CellCLIP features are 768-dimensional well-level embeddings. They were plate-normalized using negative-control median/MAD statistics and globally whitened using negative-control profiles.

The feature index tables contain:

• row_index
• profile_id
• compound_id
• split
• source
• plate
• well
• CellCLIP-only acquisition metadata when available

Load a matrix and its aligned metadata with:

import numpy as np
import pandas as pd

features = np.load("moa_verified/features/cellclip_features.npy", mmap_mode="r")
index = pd.read_parquet("moa_verified/features/cellclip_index.parquet")
assert features.shape[0] == len(index)

Labels

Each labels.json distinguishes the primary reference-label space from the larger set of all valid annotations. Label strings are case-sensitive. The special label for out-of-reference compounds is exactly novel moa.

Citation

If you use this dataset, please cite:

@inproceedings{chen2026phenoair,
  title     = {From Retrieval to Reasoning: Agentic Mechanism Prediction from Cell Painting Profiles},
  author    = {Chen, Jiayuan and Yu, Botao and Liu, Tianyu and Pham, Thai-Hoang and Wu, Meng and Zhang, Ping},
  booktitle = {Advances in Neural Information Processing Systems},
  year      = {2026}
}