AI for Bio · Benchmark 19

PMO (Practical Molecular Optimization)

Compares sample efficiency of molecular optimization within 10,000 oracle calls

bench-19Type DEvidence L2Small molecules◐ Low-frequency Self-reported中文

Overview

NamePMO (Practical Molecular Optimization) · aliases: PMO, Practical Molecular Optimization, mol_opt, PMO-1K
Full namePractical Molecular Optimization
Ambiguous names (not auto-linked)PMO-Dock
FamilyNot a family entry. Single benchmark; PMO-1K and PMO-Dock are derivative extensions, not family members.
DomainSmall molecules (also: Molecule generation)
InclusionOriginal 30
TypeD · Computational-oracle benchmark with a fixed call budget
Evidence levelL2 · Computational oracles only, no wet-lab evidence
PublisherGeorgia Institute of Technology, Massachusetts Institute of Technology (Academic lab)
First release · paper v1
Basis: arXiv 2206.12411 v1 (v2 2022-10-09) source
Other milestones
  • Preprint v1: · arXiv 2206.12411 v1 source
  • Journal online: † · NeurIPS 2022 Datasets and Benchmarks track; year only source
Latest update† partly verified · new version
Basis: PMO-Dock preprint arXiv 2610.19011 v2, a docking-oracle extension
PMO-Dock (ICLR 2026), a docking-oracle extension, arXiv 2610.19011 v2; the mol_opt repository's last commit is 2024-05-23.
Note: Task count inconsistent: the ICLR page says 23, a secondary index says 25. source
The most recent sourced event, not a claim that nothing changed since.
Next expectedNo next version or round announced by the organizers.
Status◐ Low-frequency · Paper final; repository lightly maintained; no official online leaderboard
Scale23 oracles, a 10,000-call budget, 5 seeds, 25 algorithms compared in the paper.
  • Oracles: 23 (paper (2022))
  • Call budget: 10000 oracle calls (paper (2022))
  • Seeds: 5 (paper (2022))
  • Algorithms compared: 25 (paper (2022))
Wet-lab refreshStatic data · No wet lab; all ground truth comes from computational oracle functions.
Own wet lab needed to evaluate a new modelNo, run it as a download
Models reported2 counted (self 1, modified protocol 1, cited only 1)
Unverified fields† latest_update, ranking.entries[4], ranking.snapshot_date, release_milestones.journal_online
As of2026-11-16

What it measures

  • Task: Given black-box scalar oracles and a fixed call budget, measure how efficiently an optimizer finds high-scoring molecules.
  • Truth: Computational oracle scores: QED, DRD2, GSK3B, JNK3 and 19 objectives taken from GuacaMol, served through TDC.
  • Primary metric: AUC top-10, summed over 23 tasks (max 23)
  • Secondary metrics: Per-task AUC top-10; Final top-10 mean score
  • Version traps: State the budget (10,000 / 1,000 / 100 calls), seed count, whether surrogate training counts as calls, and whether the original suite or PMO-Dock was used.

Current ranking

Self-reported

MetricAUC top-10, summed over 23 oracles (max 23), 10,000-call budget
Eval setPMO 23 oracles (QED, DRD2, GSK3B, JNK3 plus 19 GuacaMol-derived), oracle functions via TDC
Snapshot† partly verified
BoardNo official board
Retrieved2026-11-16

Caveat. No official board; each paper's own table serves as the ranking, with inconsistent budgets (10k/1k/100 calls), no significance tests, and 1-7% gaps at seed-noise level.

Shown in the overview table:

Sum 18.362 (Table 3, 'Goal-directed hit generation results', 23 PMO tasks, AUC top-10)
AUC top-10, summed over 23 oracles (max 23), 10,000-call budget · PMO 23 oracles (QED, DRD2, GSK3B, JNK3 plus 19 GuacaMol-derived), oracle functions via TDC
Self-reported · stale
not independently reproduced

RankModelScore (verbatim)Metric · eval set / protocolSource type · dateSource and note
1GenMolSum 18.362 (Table 3, 'Goal-directed hit generation results', 23 PMO tasks, AUC top-10)AUC top-10, summed over 23 oracles (max 23), 10,000-call budget
PMO 23 oracles (QED, DRD2, GSK3B, JNK3 plus 19 GuacaMol-derived), oracle functions via TDC
Self-reported
source Re-verified from the primary table (23 tasks, mean of 3 runs). Same table: f-RAG 16.928, Genetic GFN 16.213, REINVENT 14.196. The repository README lists a V2 score of 17.943 (site note, not re-verified). Self-reported.
2f-RAGSum 16.928AUC top-10, summed over 23 oracles (max 23), 10,000-call budget
PMO 23 oracles (QED, DRD2, GSK3B, JNK3 plus 19 GuacaMol-derived), oracle functions via TDC
Third-party
source Not a site model; re-verified as a baseline row.
3Genetic GFNSum 16.213AUC top-10, summed over 23 oracles (max 23), 10,000-call budget
PMO 23 oracles (QED, DRD2, GSK3B, JNK3 plus 19 GuacaMol-derived), oracle functions via TDC
Third-party
source Not a site model; 16.213 is self-reported in its own paper and re-tabulated here.
5REINVENT (prior generation; the site page covers REINVENT 4)Sum 14.196 (top of the original PMO paper's 25-method comparison; Graph GA 13.751, SELFIES-REINVENT 13.471, GP BO 13.156, lowest MolDQN 5.620)AUC top-10, summed over 23 oracles (max 23), 10,000-call budget
PMO 23 oracles (QED, DRD2, GSK3B, JNK3 plus 19 GuacaMol-derived), oracle functions via TDC
Third-party
source The 14.196 value is re-verified in GenMol's Table 3; the other totals from the PMO paper come from the site note and are not re-verified. REINVENT 4 itself reports no PMO score, so the pairing points at the prior-generation REINVENT.
—ReaSyn v2Withheld: Incomplete provenance (no url, no date), so the honesty rule forbids showing a number.AUC top-10, summed over 23 oracles (max 23), 10,000-call budget
PMO 23 oracles (QED, DRD2, GSK3B, JNK3 plus 19 GuacaMol-derived), oracle functions via TDC
Self-reported
No sourced date†
no source link No 23-task total was located this round, so the score is null.

Models reported

Grouped by relation type; each line shows its source type. Counted relations add up to 2. A pairing whose evidence rests only on this survey's own page is shown as unverified.

Developer report, official protocol (1)

  • GenMol Self-reported · PMO 23 tasks, original protocol · evidence: link

Developer report, modified protocol (1)

  • ReaSyn v2 Self-reported · Uses 15 TDC oracle tasks rather than the full 23 †partly verified · evidence: link

Cited as comparison only (1, not counted)

  • REINVENT 4 Third-party · The PMO paper used the prior-generation REINVENT as a baseline; REINVENT 4 itself reports no PMO score. †partly verified · evidence: link

Caveats

  1. All oracles are cheap computational functions and most come from GuacaMol, so the two are not independent evidence; no docking, no assays.
  2. Self-reported gains of 1-7% sit at the noise level of 5 seeds, and most papers run no significance test.
  3. Summing 23 tasks is an arbitrary weighting; task difficulty varies and a single task can dominate the total.
  4. Budgets are mixed (10,000 / 1,000 / 100 calls), so numbers across papers are often not comparable.
  5. No evidence links PMO scores to measured hit rates; rediscovery tasks can be solved from pretraining memory.

How to use it

Clone mol_opt, build per-method environments, pull oracles through TDC and run 5 seeds for AUC top-10; no wet lab needed. Report the budget, the seed count and whether surrogate training counts as calls, and re-run baselines instead of copying other papers' tables; claims about real value need synthetic-accessibility, QED and diversity filters plus a docking or free-energy oracle.

Pairs well with: GuacaMol (+MOSES) · CACHE · TDC (ADMET + DTI) · CrossDocked2020

Sources

  1. PMO paper, arXiv 2206.12411 (preprint) ·
  2. mol_opt repository (repository)
  3. GenMol paper with the 23-task PMO table (preprint) ·
  4. PMO-Dock preprint (preprint) · †
  5. PMO-Dock page at ICLR 2026 (official site)
  6. Third-party analysis of docking-oracle exploitation (third-party paper)

Corrections

3 corrections
FieldWasNowBasisFound
latest_update20262026-11-09 (PMO-Dock preprint arXiv 2610.19011 v2; v1 2026-10-13)source2026-11-16
ranking.entriesReaSyn v2 shown with a PMO scorescore withheld: no url and no date for the figuresource2026-11-16
models_reportedREINVENT 4 counted as reporting PMOcited-as-comparison-only: only the prior-generation REINVENT appears, as a baseline in the PMO papersource2026-11-16

Part of the AI for Bio benchmarks survey · all 44 benchmarks · open in the explorer. Record as of 2026-11-16.