metric-shift/metric-shift-benchmark

Metric Shift Benchmark

A cross-domain benchmark for predicting expensive scientific measurements from cheap surrogates, spanning 6 scientific fields and 134 valid (y1, y2) pairs with a standardized evaluation protocol.

Paper: Metric Shift: A Benchmark for Predicting Expensive Scientific Measurements from Cheap Surrogates (NeurIPS 2026 Evaluations & Datasets Track, under review)

Benchmark Overview

Dataset	Domain	Samples	Feat. dim	Labels	Valid pairs	License
zinc250k	Drug Chemistry	249,455	14	3	6	ZINC academic-use, f...
air_quality	Environmental Science	382,168	7	6	28	CC-BY-4.0 (UCI ML Re...
jarvis_materials	Materials Science	10,800	14	6	30	Public domain / NIST...
protein_fitness_expanded	Protein Biology	61,704	22	24	38	MIT (ProteinGym aggr...
drug_admet	Pharmacology	1,523	14	4	12	CC-BY-4.0 (Polaris H...
climate_stations	Climate Science	28,488	5	5	20	CC-BY-4.0, dual attr...
Total	---	734,138	---	---	134	---

Problem: Metric Shift

Given a shared entity x (molecule, material, protein variant), a cheap source metric y1, and an expensive target metric y2: can we use universally available y1 to improve prediction of the sparsely labeled y2?

Key properties:

• y1 is always available at test time (cheap to measure for any new candidate)
• The input distribution p(x) is fixed; only the prediction target changes
• Unlike domain adaptation (shifts p(x)) or multi-task learning (co-predicts)

Evaluation Protocol

• Split: 60% train / 20% val / 20% test at split_seed=42
• Labeled ratio: 20% of train (main setting); 1% and 5% for ablation
• Seeds: 5 model seeds per pair
• Metrics: R-squared and Spearman rho
• Significance: Paired t-test across seeds + Benjamini-Hochberg FDR at q=0.05
• Aggregation: Macro-median (per-dataset median, then cross-dataset median)
• StandardScaler: fit on labeled train only

Usage

import pandas as pd

# Load one sub-dataset
features = pd.read_csv("zinc250k/features.csv")
labels = pd.read_csv("zinc250k/labels.csv")

# Each (source, target) column pair in labels defines a Metric Shift task
# See metadata.json for the list of valid pairs with Spearman correlations

One-command reproduction of all tables and figures:

pip install metric-shift-benchmark
python -m metric_shift.run_all

Dataset Details

zinc250k — Drug Chemistry

249,455 drug-like molecules, 14 RDKit descriptors, 3 labels (logP, QED, SAS), 6 pairs

• Source: ZINC database (Irwin & Shoichet 2005; Sterling & Irwin 2015)
• License: ZINC academic-use, free redistribution with attribution
• Features (14d): MolWt, HeavyAtomCount, NumHeteroatoms, NumValenceElectrons, TPSA, MolMR, HBA, HBD, NumRotatableBonds, RingCount, NumAromaticRings, FractionCSP3, BalabanJ, BertzCT
• Labels (3col): logP, QED, SAS

air_quality — Environmental Science

382,168 hourly records, 7 meteo features, 6 pollutants, 28 pairs

• Source: Beijing Multi-Site Air-Quality Dataset (Zhang et al. 2017)
• License: CC-BY-4.0 (UCI ML Repository)
• Features (7d): TEMP, PRES, DEWP, RAIN, WSPM, wd_sin, wd_cos
• Labels (6col): PM25, PM10, SO2, NO2, CO, O3

jarvis_materials — Materials Science

10,800 inorganic crystals, 14 composition descriptors, 6 labels, 30 pairs

• Source: JARVIS-DFT 3D (Choudhary et al. 2020)
• License: Public domain / NIST (17 USC §105)
• Features (14d): mean_Z, std_Z, mean_X, std_X, mean_row, std_row, mean_group, std_group, mean_atomic_mass, std_atomic_mass, density, volume_per_atom, n_sites, packing_fraction
• Labels (6col): formation_energy_peratom, optb88vdw_bandgap, bulk_modulus_kv, shear_modulus_gv, n_seebeck, p_seebeck

protein_fitness_expanded — Protein Biology

61,704 variants, 22-d mutation features, 24 DMS assays, 38 within-protein pairs

• Source: ProteinGym substitution benchmark (Notin et al. 2023)
• License: MIT (ProteinGym aggregation)
• Features (22d): protein_id, n_mutations, AA_A_diff, AA_C_diff, AA_D_diff, AA_E_diff, AA_F_diff, AA_G_diff, AA_H_diff, AA_I_diff, AA_K_diff, AA_L_diff, AA_M_diff, AA_N_diff, AA_P_diff, AA_Q_diff, AA_R_diff, AA_S_diff, AA_T_diff, AA_V_diff, AA_W_diff, AA_Y_diff
• Labels (24col): p53_null_etoposide, p53_null_nutlin, p53_wt_nutlin, blat_deng_2012, blat_firnberg_2014, blat_jacquier_2013, blat_stiffler_2015, pten_matreyek_2021, pten_mighell_2018, cp2c9_amorosi_abundance_2021, cp2c9_amorosi_activity_2021, hsp82_flynn_2019, hsp82_mishra_2016, spike_starr_bind_2020, spike_starr_expr_2020, a0a2z5u3z0_doud_2016, a0a2z5u3z0_wu_2014, rl401_mavor_2016, rl401_roscoe_2013, rl401_roscoe_2014, ccdb_adkar_2012, ccdb_tripathi_2016, vkor1_chiasson_abundance_2020, vkor1_chiasson_activity_2020

drug_admet — Pharmacology

1,523 compounds, 14 RDKit descriptors, 4 ADME endpoints, 12 pairs

• Source: Biogen ADME-Fang v1 (Fang et al. 2023)
• License: CC-BY-4.0 (Polaris Hub)
• Features (14d): MolWt, HeavyAtomCount, NumHBD, NumHBA, TPSA, MolLogP, NumRotatableBonds, RingCount, NumAromaticRings, FractionCSP3, MolMR, BertzCT, BalabanJ, NumHeteroatoms
• Labels (4col): LOG_HLM_CLint, LOG_RLM_CLint, LOG_SOLUBILITY, LOG_MDR1-MDCK_ER

climate_stations — Climate Science

28,488 daily records, 5 context features, 5 climate variables, 20 pairs

• Source: Open-Meteo Historical Weather API / ERA5 reanalysis
• License: CC-BY-4.0, dual attribution to Open-Meteo and Copernicus C3S/ERA5
• Features (5d): lat, lon, day_sin, day_cos, year_norm
• Labels (5col): temp_max, temp_min, precip, windspeed, solar_radiation

Responsible AI

• Personal / sensitive data: None. All datasets contain scientific measurements on molecules, materials, proteins, pollutants, or climate variables. No human subjects, no personally identifiable information.
• Intended use: Benchmarking ML methods for the Metric Shift problem. Not intended for direct clinical, regulatory, or safety-critical deployment.
• Known limitations: (1) All six datasets are re-curations of existing public sources; our contribution is pair construction, validity filter, and protocol. (2) Domain coverage spans chemistry, biology, materials, environment, and climate --- not yet high-energy physics, astronomy, or social science. (3) Feature spaces are intentionally low-dimensional (5--22d) to isolate the contribution of y1; higher-dimensional encoders may change relative method rankings.
• Potential misuse: drug_admet contains ADME measurements that could theoretically inform adverse drug design; however, the 1,523-compound dataset is far too small and coarse for such purposes, and all data is already public.

Maintenance

The authors commit to maintaining this repository for at least 2 years post-publication, with semantic versioning (v1.0, v1.1, ...) and a CHANGELOG for every split, filter, or protocol change.

Citation

@inproceedings{metric_shift_2026,
  title={Metric Shift: A Benchmark for Predicting Expensive Scientific Measurements from Cheap Surrogates},
  author={Anonymous},
  booktitle={NeurIPS 2026 Evaluations and Datasets Track},
  year={2026},
  note={Under review}
}