docs/experiments.md

Manuscript experiments and commands

The manuscript contains unfilled scientific-result tables and empirical plot layouts. Executed reference runs are stored separately under results/. Their graph sizes and metrics are listed in validation.md.

Manuscript item	Workflow	Remaining experiment
Figure 1	Editable TikZ and molecule-rendering source in assets/figure1_source/	Final formulation diagram included
Table 1, discrete generation	run_benchmarks.py with toy.yaml, grid.yaml, strings.yaml	Larger held-out graph-instance study, SubTB, and learned-backward TLM comparator
Figure 2, distribution diagnostics	sweep_exact.py, saved training.json, plot_results.py	Full five-seed sweep over all specified graph families
Table 2, molecular construction	run_molecular.py with molecular_public.yaml or enamine_local.yaml; evaluate_candidates.py	Licensed inventory and matched third-party generators
Table 3, molecular properties	fit-properties, score, generation, candidate evaluation	LIT-PCBA docking, full target panel, and external baselines
Figure 3, preferences and scaling	sweep_preferences.py; repeated local-inventory configurations	Full target-specific preference grid and measured catalog scaling
Table 4, ablations	run_benchmarks.py --config configs/ablations.yaml	Matched molecular ablations and larger graph instances
Table 5, scaling	run_molecular.py with parent limits 100, 1000, 10000	Licensed catalog expansion, peak-memory measurements, full timing sweep
Supplement, reaction budgets	molecular_public.yaml runs B = 0, 1, 2	Full target inventory and matched external methods
Supplement, target-specific docking	Common candidate records and receptor/grid protocol	Fifteen prepared target pockets and external generator integrations
Supplement, property prediction	dooable fit-properties	Five split seeds and selected released molecular encoder

Implemented comparisons

uniform assigns equal mass to outgoing actions. reference_tilt reweights that path process by terminal reward through a backward value recursion. tb_uniform learns a forward trajectory-balance sampler against locally uniform backward probabilities. dooable fits the cost-dependent backward prefix values and forward trajectory balance. exact uses both dynamic programs.

zero_cost sets every edge cost to zero for the conditional route law. duplicate_endpoints assigns a separate terminal reward to each incoming stop edge and evaluates the resulting policy after canonical merging. unnormalized uses learned prefix differences directly in the backward path weights during trajectory-balance training. tb_exact uses the exact cost-dependent backward policy during forward learning.

The cost-free and duplicated-endpoint comparisons use exact forward construction to isolate the corresponding distributional change. Their runtimes therefore describe the exact ablation, while learned-policy runtimes include training. Comparisons of optimization speed should group methods by solver mode.

External comparisons

The following original implementations require integration with the selected inventory and scoring protocol. They have not been run in this package.

Method	Verified primary publication	Required integration
SynFlowNet	ICLR 2025, synflownet citation	Parent selection, matched reaction/reagent inventory, candidate and route export
RxnFlow	ICLR 2025, rxnflow citation	Inventory conversion, reward callback, route export
CGFlow	ICML 2025, cgflow citation	Prepared target pockets, pose handling, matched synthesis actions
SyntheMol	Nature Machine Intelligence 2024, synthemol citation	Matched construction inventory, score callback, search-call accounting
SubTB	ICML 2023, subtb citation	Identical graph and reward, original subtrajectory objective
TB with learned backward TLM	ICLR 2025, backward citation	Original backward-policy training objective

These method names refer to their original published implementations. External outputs should provide one canonical structure and its full construction record per candidate, with method, seed, target, budget, requested count, and oracle-call accounting. Adapt native route records to the actions schema in data.md before replay.

Scoring and split protocol

The public example uses measured BACE inhibition and Caco-2 transport to fit separate property predictors. Full target-pocket experiments require target-specific docking. Fix receptor preparation, protonation, ligand preparation, docking grids, Vina version, exhaustiveness, and random seeds before comparing methods. Docking energies use kcal/mol. Record raw energies separately from normalized reward utilities.

Select clipping bounds and hyperparameters on the validation partition. Group parents by Bemis-Murcko scaffold before generating descendants. Resolve products shared across parent partitions into one partition before held-out evaluation. The included small graph experiments optimize one supplied graph per run. Generalization across disjoint graph instances is a separate experiment.

Count every unique oracle evaluation, including scores used during support construction or preprocessing. Store cached scores and reuse them within a declared experiment. For the 168-outcome public graph, the scoring step evaluates all 168 outcomes before training. Sampling counts are recorded separately.

The current public evaluator reports two-dimensional hypervolume over predicted BACE and Caco-2 utilities. The full docking study uses affinity and permeability utilities on the same normalized axes. QED and internal diversity are additional readouts. Report undefined summaries as missing when fewer than the required candidates are available.

Citation verification

bibliography/verified_references.csv and .json contain 72 references, each with authors, publication year, venue, primary publication URL, metadata source, and verification date. Publisher BibTeX, official conference records, journal citation metadata, and publisher-deposited Crossref records were used. Conference versions replace corresponding preprints. Strong Stochastic Flow Maps is identified as an ICML workshop contribution.