StarLiu714/MOFinder_PU

Dataset Preparation for PN/PU Comparative Experiments

Overview Statistics

Inputs:

• datasets/csv_before_assembly/processed_positive.csv: 15340 P entries
• datasets/test/PN_holdout.jsonl: 1320 P + 1275 N, used for holdout alignment and training-side exclusion

Outputs:

Output file	Pool recipe	Labels and sampling	Number of rows
datasets/pools/U_native_{1x,10x,50x}.jsonl	U_r	Standard U pool	{11293,104932,573594}
datasets/train/E1234_PU_train_{1x,10x,50x}.jsonl	P_train + U_r	Shared input for Elkan–Noto, nnPU, Bagging PU, and uPU; U remains U	{22586,116225,584887}

Here, P_train=11293. The standard U_r pools contain {1,9.2918,50.7920} * len(P_train) rows, respectively.

Construction Pipeline

Step 1: CSV dedup and preprocessing

Input:
processed_positive.csv: 15340 P entries

Process:

• Filtering based on the 8 required fields
  • Required fields: 'metal_precursor', 'organic_linker', 'modulator', 'solvent', and 'time_h'; discard a row if any of these is missing
  • At least two of M_L_ratio, metel_concnertation, and temperature_c must be nonempty and parseable
• Exact deduplication
  • After trimming the 8 fields, some rows may share the same 'condition_key' while differing in fields outside these 8 fields. Remove these duplicates by retaining the first P row encountered
  • 13306 P remain after deduplication
• Balance trim: drop 18 rows to align the count; default trim_seed=123

Output:
P pool: 13288 P

Step 2: Assembly and Development-Set Isolation

Align with the 2595-row holdout, split off the development set, and exclude the development portion from the P and N pools.

Step 2.1: Holdout Exclusion

From a P-only perspective, this step follows the split logic in https://github.com/zzhenglab/MOFinder/blob/main/src/mofinder/training/prepare.py to ensure consistent holdout evaluation across all comparisons. Directly exclude entries appearing in the PN holdout from the P pool.

Input:
13288 P datasets/test/PN_holdout.jsonl: 1320 P + 1275 N

Process:
Read datasets/test/PN_holdout.jsonl and exclude its 1320 P entries from the P pool.

Output:
P pool: 11968 P

Step 2.2: Development-Set Split

Input:
11968 P

Process:
Split seed default to 666.

Output:

• PU dev: 598 P
• P pool: 11370 P

Step 2.3: Exclude P-Pool Records That Are Too Similar to the Development Pool

Input:
11370 P

Process:

• Remove training-side records whose conditions/families overlap with the holdout;
• Deduplicate by 'cluster_key'

Output:
P pool: 11293 P

Step 3: Construct Shared Candidate Pools

At the end of Step 2, the P pool and development pool have been determined. However, constructing the PU training dataset also requires U and C pools.

3a: U Pool Construction

• Construct a joint metal-linker pool from the P pool:
  • First exclude metal_precursor entries with frequency <3, unparseable entries, or placeholders; retain 718 'organic_linker' and 321 'metal_precursor' entries (follows Park2022 JCIM)
  • The overlap between this 718x321 Cartesian combination pool and the metal-linker combinations observed in the 11293-row P pool covers 6589 P (out of 11293), comprising 2836 joint metal-linker combinations
  • Set the numbers of L and M selected through stratified sampling according to the target U size
  • 1x: the intersection of top100_metal x top400_linker with the 2836 combinations gives 1573 joint metal-linker combinations (covering 4773 out of 11293 P-pool entries); 10x: use all 2836 joint metal-linker combinations; 50x: the union of top100_metal x top400_linker with the 2836 combinations gives 40971 joint metal-linker combinations;
  • First holdout U group (10692 entries): top10_metal x top40_linker; second U group (2508 entries): extract observed, complete metal-linker pairs from all P/N records in datasets/test/PN_holdout.jsonl, deduplicate, and sample uniformly without separating metal from linker
• Construct a joint 6-field pool from the P pool:
  • Overall counts: 1173 'modulator', 216 'solvent', 439 'metal_concentration_mM', 568 'M_L_ratio', 74 'temperature_C', and 158 'time_h'
  • Define the Cartesian Space: trim the long tail of each field, retaining values corresponding to 95% coverage for each individual field, namely: 793 'modulator', 39 'solvent', 193 'metal_concentration_mM', 273 'M_L_ratio', 23 'temperature_C', and 35 'time_h', giving a combination space of $1.31*10^{12}$ (1311758M)
  • Define the 6-field joint space: among the 6-field combinations constructed for the standard joint pool, require 'modulator', 'solvent', 'metal_concentration_mM', 'M_L_ratio', 'temperature_C', and 'time_h' to come together as a complete set from the same P record; there are 7611 such joint 6-field sets
  • Training: Set the sampling counts from the Cartesian 6-field space according to the target U size: {1x: 7.2, 10x:37, 50x:14}
  • Holdout: All PU experiments share datasets/test/PU_holdout.jsonl, with a fixed 13200 U. The first group of 10692 U comes from candidates constructed using the Cartesian 6-field space above; the second group of 2508 U comes from PN joint-construction candidates. The current set is a fixed, selected subset: it does not sample a fixed 330 entries per metal-linker pair and does not change with the training-set P-U ratio
  • Deduplicate the two U groups using the normalized 8 fields and keep them mutually exclusive; exclude complete conditions appearing in known P/N, dev, and the previously trained PU inputs checked in this round, drawing replacements when conflicts occur. Retain the original 1320 P from datasets/test/PN_holdout.jsonl. The combined holdout contains 1320 P + 13200 U = 14520 entries and is shared by all PU experiments
  • QC filtering: solvent must not be missing; concentration (0,10000] mM; M:L (0,100]; temperature [-80,300] °C; time (0,2160] h; modulator may be null (meaning unknown/not reported)
  • Multiple reports do not increase the sampling weight;
• 8-field U combination pool
  • Sample uniformly from the joint 6-field pool
  • Sample without replacement within the same metal-linker combination; reuse is allowed across different combinations
  • Prioritize sampling majority metal-linker tuples; if a blacklisted combination is drawn, continue drawing replacements

U pool global random seed default to 1234.

3b: C Pool Construction

Summary

At the end of Step 3, the shared pools are as follows:

Shared pool file	Number of rows / label	Starting point and construction method	Downstream users
P_train	11293 / P	E1-E4, M1-M2	Shared P set for all training datasets; P source for global/local candidates
U	~{1,9.2918,50.7920} *11293 / U	from P parent to 1–4 field rejection sampling	E1-E4
C_local	To be implemented in Step 3.2 / U	Not generated by the current workflow	M1-M2
dev	598 P	The split preserves order and labels	Held out from training for checkpoint selection
holdout	1320 P + 13200 U = 14520	Read the P portion of datasets/test/PN_holdout.jsonl for training-side exclusion; all PU experiments share the same fixed 13200 U, regardless of the training-set PU ratio	align with PN holdout

The size of the constructed $U_{train}$ pool varies with the P-U ratio:

P-U ratio	P size	U size	metal-linker sampling	6-field sampling
1x	11293	11293	top100_metal × top400_linker $\cap$ M-L joint = 1573	standard 6-field space, per-pair sampling
10x	11293	104932	M-L joint = 2836	per-pair sampling
50x	11293	573594	top100_metal × top400_linker $\cup$ M-L joint = 40971	max U_master, whitelist hit then resample