acceptodds
Under review as a conference paper at ICLR 2027

MolWorld: Molecule World Models for Actionable Molecular Optimization

Abstract

Molecular optimization in drug discovery aims to discover molecules with improved target properties, but practical lead optimization often requires more than high predicted scores. A useful candidate should also be actionable: it should be reachable from known molecules through a sequence of local structural modifications, providing explicit structural references for interpreting property changes within an evolving chemical series. Existing de novo and single-molecule optimization methods do not explicitly model such reachability, especially when both the target molecules and the intermediate molecules connecting them to known compounds are unknown. In this work, we formulate actionable molecular optimization as iterative expansion of a molecule-transfer graph, where nodes are molecules and edges encode matched molecular pair (MMP) relations representing localized structural differences. We propose MolWorld, a molecule world model-guided framework that uses connected local subgraphs as anchor contexts for graph expansion. Given the current graph and a candidate context, a latent molecule world model predicts the resulting local structural expansion. Its dynamics are shared across property objectives, while a property-specific scorer evaluates predicted expansions to guide context selection. A context-conditioned generator then uses the molecular structures and MMP relations within selected contexts to propose candidates, learning from local graph-completion tasks. Candidates are evaluated, and those connected through verified MMP relations are incorporated into the graph, preserving reachability from the initial set. The updated graph serves as the state for subsequent optimization. Experiments on property optimization and docking-based tasks show that MolWorld discovers high-property molecules while maintaining stronger structural connectivity, supporting actionable and sequential molecular design.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Molecular Optimization, World Models, Generative Modeling, Graph Representation Learning, Matched Molecular Pairs, Conditional Molecular Generation

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_1HNSOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper formulates "actionable" molecular optimization: each generated molecule must be reachable from an initial set via matched-molecular-pair (MMP) paths. MolWorld grows a molecule-transfer graph. A "molecule world model" (JEPA-style graph encoder, Transformer dynamics model predicting the latent embedding of an expansion, property-specific ranking scorer) selects five-molecule anchor contexts. A T5Chem-based set-conditioned generator proposes candidates, and RDKit MMP checks admit them. Evaluation covers PMO-style optimization (GSK3β/QED/JNK3/DRD2; ZINC250 10k calls, PMV21 1k calls) and Vina docking (1k calls), with top-k AUC over reachable molecules and an isolated-node ratio (Iso).

Critical strengths
  • A new, practically motivated formulation with a system that natively exploits analogue-series structure. Ablations show the pieces matter (Fixed Search Graph −9.1/−11.7 AUC points; STONED generator 22.2 vs 55.2 AUC@100; Fig. 4).

  • Unusually transparent evaluation: 9 baselines, 3 seeds with SDs, a random-pool control, encoder/dynamics validation (edge removal raises masked-node loss 3.45×; dynamics error −76.6% vs a mean target), and self-initiated leakage analyses (Tables 8–9) that report unfavourable results.

Critical weaknesses
  1. The comparison is not information- or constraint-matched.
    • (a) The property scorer is trained on oracle labels over ChEMBL-derived expansions (84,490 outcomes) outside the call budget. The generator is also ChEMBL-trained, which likely overlaps the public bioactivity data behind the TDC activity oracles. On ZINC250 JNK3/DRD2, 50%/39% of generated molecules (20/30 and 18/30 top-10 entries) have scaffolds in the scorer's training set but not in the initial graph (Table 9). How much of the margin this explains is not quantified.
    • (b) Seeding uses "cached" property values of the initial graph (up to ~10^5 molecules). It is unclear whether these are budgeted or given to baselines.
    • (c) Reachability is built into MolWorld, while baselines are unconstrained and post-filtered. Fig. 7 (all valid molecules, not reachable-only) shows several baselines approaching MolWorld by 10k calls and everything saturating on QED. Table 1's margin therefore largely reflects the reachability filter. MMP-aware baselines (e.g. the cited MMP-conditioned transformers, or GraphGA/MolLEO with MMP-compatible operators) and unconstrained PMO numbers are missing.
    • (d) In Fig. 4, Random Selection (57.3/51.3 AUC@10/100) already exceeds the best PMV21 baseline means I compute from Table 1 (≈55.3 GenMol, ≈33 MolLEO). The generator and graph design, not the world model, explain most of the gain. The STONED ablation also conflates quality with reachability, since random SELFIES mutations are rarely MMP-compatible.
  2. "Actionability" and the "world model" are weakly validated.
    • Actionability means an RDKit MMP (≤3 cuts, variable fragment ≤33% of heavy atoms) to any node of a graph of hundreds to ~10^5 molecules. That is a loose criterion; the Fig. 6 text notes G4 differs from G1 at several positions.
    • There is no hop-distance distribution, per-edge property deltas, SA, or chemist assessment. The QED case study is descriptor-level.
    • Figs. 3/5/8–10 show counts or trends without baselines, and counts favour whoever emits more reachable molecules. PMV21 GSK3β/JNK3 scores stay ≤~0.5/0.35, so "high-property" is overstated.
    • The world-model-specific gain over a Direct Context Ranker is ~+3 (heuristic pool) to ~+4 (random pool) AUC points. It has no error bars, uses one dataset (PMV21, 1k calls), and the y-axis starts at 20. No code release is mentioned.
Questions for authors
  1. How many oracle-labelled molecules train each scorer? How does performance change with a scorer trained without test-property labels, or only online?
  2. Are initial-graph property values free? Did every baseline receive the same initial set and values?

Suggestion:

  • Report hop distance from the initial set for top-10/100 molecules, the fraction reachable only via generated intermediates, and per-edge property deltas.
  • Add MMP-aware baselines, unconstrained PMO results, ablation SDs on ZINC250, diversity/novelty/SA, and a held-out oracle.
2 comments

We thank the reviewer for the careful reading and for the generous assessment of the formulation and of our transparency.

W1. (a) & (b), Q1 & Q2. Each property scorer is trained on labelling that covers roughly 355k molecules. Neither this labelling nor the property values of the initial graph counts against the reported budget, but all baselines were given the same initial set and the same precomputed scores. On docking we do train the scorer online only, since the oracle is expensive there, and the initial graph carries docking scores for just 276 molecules. MolWorld still has the best AUC@10 and AUC@100 on all three targets (Table 2).

How much of the margin the scaffold overlap explains is answered most directly by PMV21, where the initial graph shares no scaffolds with the scorer's label training set and generated-molecule overlap is zero on GSK3β, JNK3 and DRD2, Top-10 and Top-100 included. MolWorld attains the highest AUC@10 and AUC@100 on all three regardless, so the margin survives where there is no overlap to explain it. ZINC250 admits a direct measurement, reported in Table 8. Zeroing the score contributions of overlapping molecules and recomputing moves mean AUC@10 by at most 0.000469 and mean AUC@100 by at most 0.000656.

(c). We have added an opponent that satisfies reachability natively, searching MMP transformation space directly under the same initial graphs, reachability rules and oracle budget. Each round draws a batch of molecules from the current graph, at random or greedily by property, and applies several random valid MMP transformations to each, so every candidate is by construction an MMP of a molecule already in the graph and Iso is exactly zero.

MethodGSK3βQEDJNK3DRD2
MMP-Search (random)0.3016 ± 0.01620.6682 ± 0.03500.1909 ± 0.01290.6007 ± 0.0706
MMP-Search (greedy)0.3965 ± 0.00710.8095 ± 0.01430.2240 ± 0.01830.8368 ± 0.0239
MolWorld0.4225 ± 0.01590.8892 ± 0.01190.2710 ± 0.02860.9018 ± 0.0063

AUC@10 on PMV21. The greedy variant already beats every method in Table 7 yet still falls short of MolWorld, and the gap widens in AUC@100, from 0.6701 to 0.8194 on QED and from 0.6326 to 0.8205 on DRD2. With reachability identical on both sides, what remains can only come from whether the search has direction.

(d). We have added the anchor-only dynamics control the reviewer's Q2 asks for, removing the current-graph representation so that the expansion is predicted from the context alone, everything else held fixed.

MethodAUC@10AUC@100
Anchor-only dynamics0.5885 ± 0.00390.5220 ± 0.0004
MolWorld0.6211 ± 0.00740.5516 ± 0.0059

Averaged over the four properties on PMV21, three seeds. Dropping the graph state costs most of the gain, which locates it in conditioning the expansion prediction on the current state rather than in the extra stage as such. Standard deviations for all ablation variants have been added to Figure 4.

We agree that Random Selection exceeding the baselines points at the framework rather than the world model, and this is what contributions 1 and 3 claim — the reachability formulation, the context-conditioned generator and the iterative graph expansion account for most of the gain, with the world model adding roughly three points on top. The four selection variants are worth reading together in this light. Random at 0.5727, Mean-Property at 0.5809, Direct Ranker at 0.5889 and anchor-only at 0.5885 fall within a band of under two points, while the full model sits clearly outside it, and the randomly sampled pool reproduces the same structure with 0.4545, 0.4897 and 0.4884 against MolWorld's 0.5280.

The reviewer is also right that STONED conflates quality with reachability. We have added a control that changes only the edges, running a single round over a fixed candidate context pool with the context molecules, backbone and training protocol held constant.

PropertyReal edgesShuffledRemoved
QED0.7845 ± 0.00000.7299 ± 0.00650.7493 ± 0.0074
JNK30.1177 ± 0.00120.1003 ± 0.00060.1003 ± 0.0042
DRD20.8592 ± 0.00130.8399 ± 0.01450.8382 ± 0.0073
GSK3β0.2653 ± 0.00470.2557 ± 0.00120.2510 ± 0.0131

With the selected molecules identical, altering how the MMP relations are presented is enough to lower Top-10, so the generator is drawing on them.

W2. We have added the analyses the reviewer asks for.

Hop distance and dependence on generated intermediates. Generated molecules were filtered by the batch-wise reachability rule, ranked by property, and their shortest distance to the initial set measured on the final graph. Mean ± s.d. across three seeds:

PropertyTop-10 hopsTop-10 via generated intermediatesTop-100 via generated intermediates
QED2.77 ± 0.2996.67% ± 5.77%94.67% ± 5.86%
JNK32.87 ± 0.81100% ± 0%94.00% ± 6.56%
DRD22.43 ± 0.35100% ± 0%99.67% ± 0.58%
GSK3β3.53 ± 0.57100% ± 0%100% ± 0%

Across all runs, 119 of 120 Top-10 molecules (99.17%) have no MMP edge to the initial set and are reachable only through molecules the search itself produced; for Top-100 the figure is 1165 of 1200 (97.08%). The paths are forward in time rather than completed in hindsight — one Top-10 molecule on GSK3β seed 29 sits six steps out, its nodes entering at rounds 0, 6, 7, 10, 12, 13 and 14.

Per-edge property deltas. We measured the change along each edge in the direction generation takes, from an initial-graph molecule to a new molecule attached to it, over all 1,000 generated molecules per run and across three seeds:

PropertyInitialGeneratedChangeEdges improving
QED0.60940.7785+0.1691 ± 0.003883.98%
JNK30.06490.0982+0.0332 ± 0.023363.48%
DRD20.51600.7818+0.2658 ± 0.036086.61%
GSK3β0.08220.1609+0.0787 ± 0.006673.58%

On every objective the property improves across the majority of these edges, so the local modifications are moving in a consistent direction rather than drifting.

Synthetic accessibility. SA scores for the reachable generated molecules, against the 722 molecules of the initial graph (3.19 ± 0.29):

PropertyAll reachableTop-10Top-100
QED2.88 ± 0.362.70 ± 0.342.73 ± 0.33
JNK33.13 ± 0.253.19 ± 0.333.17 ± 0.30
DRD22.85 ± 0.242.80 ± 0.152.82 ± 0.21
GSK3β3.40 ± 0.283.49 ± 0.293.60 ± 0.26

Generated molecules stay within the synthetic accessibility of the series they come from, on QED and DRD2 somewhat below it, and neither the Top-10 nor the Top-100 sets drift away from it.

We thank the reviewer again for the detailed suggestions, and hope these additions address the concerns raised.

Reviewer_KYP3OpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
3
Soundness
1
Read review
Paper summary

The paper proposes MolWorld, a molecular optimization framework that iteratively expands a matched molecular pair graph while preserving structural reachability from an initial molecule set. Its main contributions are the reachability-constrained formulation, a context-conditioned generator, and a latent world model that predicts expansion outcomes to guide context selection through a property-specific scorer. Experiments cover four property objectives on two datasets and three docking targets under fixed online oracle budgets. The reported results show improved constrained optimization performance and connectivity.

Critical strengths

1.The paper addresses molecular property optimization while explicitly maintaining MMP reachability to the initial molecular set.

2.The framework separates shared structural dynamics from property-specific scoring, allowing the structural model to be reused across optimization objectives.

3.Experiments cover two molecular datasets and three docking targets, showing promising performance against multiple baselines.

Critical weaknesses
  1. The paper introduces structural reachability as a central requirement, but this requirement is ultimately enforced by MMP validation and the acceptance rule. It does not explain what additional decision-relevant information connected contexts provide beyond single-molecule editing or exemplar-based search, or why predicting structural expansions is particularly appropriate for exploiting that information. Consequently, the connection between the stated problem and the core technical design insufficiently justified.

  2. The model regresses averaged latent representations of expansion graphs, while the optimization objective prioritizes the highest-property accepted molecules. The paper neither characterizes which property-relevant distinctions survive this representation nor analyzes how representation-prediction accuracy relates to action-ranking quality. This gap is particularly consequential because the encoder and dynamics model are frozen during property-scorer training, preventing the intermediate representation from adapting to the target objective.

  3. The mathematical formulation contains unresolved gaps between the stated objective and the implemented algorithm. Equation (3) does not justify its additive approximation when generated outputs overlap and the final set utility is non-additive. Equation (8) leaves the local completion criterion under-specified, while Equation (9) incompletely defines the transition from graph-fused representations to token-level decoding. Equations (6) and (12) do not establish the informational complementarity attributed to token decorrelation, and Equations (16)–(17) insufficiently explain how property scales are handled when balancing quality and penalties. Moreover, Equation (18) can induce action preferences inconsistent with Equation (2), while Equation (19) lacks an established statistical consistency relationship with expected action quality under stochastic generation.

  4. The experiments fail to isolate the contribution of the core ideas. The STONED ablation simultaneously changes the generator prior, context size, and relational modeling. Consequently, the gains cannot be reliably attributed to structural prediction and connected-context modeling.

Questions for authors
  1. What decision-relevant information does connected MMP topology contribute beyond the identities of the context molecules?
  2. Under what assumptions do the averaged expansion representation and the objectives in Equations (3), (18), and (19) support the optimization objective in Equation (2)?
  3. How does the end-to-end computational cost scale with graph size and candidate-pool size relative to direct ranking and simpler search strategies?
4 comments

We thank the reviewer for the careful and detailed review.

Several of the concerns below concern the formal basis of the method, and we address them through a new appendix (Formal Framework) added in the revision, which derives the method from the stated objective: the reachability constraint and its round-wise enforcement, the locality argument behind conditioning on a fixed-size context, the relation between the additive surrogate of Eq. (3) and the top-κ\kappa utility, and a bound relating dynamics error to selection regret. We refer to it where relevant below.

W1 & Q1. The problem asks for two things at once — high property and reachability — and the two halves of the framework answer one each.

Reachability is not conferred by the acceptance rule. A rule can only reject, not generate: every method passes the same RDKit MMP check, yet baseline Iso ranges from 0.3 to 1.0 while MolWorld's ranges from 0.02 to 0.11, a gap that can only come from the generative distribution. Formally, reachability is guaranteed by the round-wise update itself (Formal Framework, Prop. 1): distance has optimal substructure, so the global constraint is settled one edge at a time.

To rule out that this advantage merely reflects our having designed for the constraint, we added a baseline that searches MMP transformation space directly, under the same initial graphs, rules and budget. Each round draws a batch of molecules from the current graph (at random, or greedily by property) and applies several random valid MMP transformations to each, so every candidate is by construction an MMP of a molecule already in the graph and Iso is exactly zero. AUC@10 on PMV21:

MethodGSK3βQEDJNK3DRD2
MMP-Search (random)0.3016 ± 0.01620.6682 ± 0.03500.1909 ± 0.01290.6007 ± 0.0706
MMP-Search (greedy)0.3965 ± 0.00710.8095 ± 0.01430.2240 ± 0.01830.8368 ± 0.0239
MolWorld0.4225 ± 0.01590.8892 ± 0.01190.2710 ± 0.02860.9018 ± 0.0063

The greedy variant already beats every method in Table 7, yet still falls short of MolWorld. With reachability identical, the remaining gap can only come from whether the search has direction.

What a connected context adds over existing practice is the relations themselves. Single-molecule editing offers one starting point; exemplar retrieval (RetMol) offers several related molecules without saying how they relate; a connected context states how they transform into one another, each edge giving the scaffold a pair shares and the fragment by which they differ. RetMol's Iso is among the lowest of the baselines (0.40/0.34/0.20/0.49 on ZINC250), which confirms that supplying several related molecules helps in itself. To test the edges directly we ran a control that changes nothing else: a single round over a fixed candidate context pool, with the context molecules, backbone and training protocol held constant.

PropertyReal edgesShuffledRemoved
QED0.7845 ± 0.00000.7299 ± 0.00650.7493 ± 0.0074
JNK30.1177 ± 0.00120.1003 ± 0.00060.1003 ± 0.0042
DRD20.8592 ± 0.00130.8399 ± 0.01450.8382 ± 0.0073
GSK3β0.2653 ± 0.00470.2557 ± 0.00120.2510 ± 0.0131

With the selected molecules identical, altering how these MMP relations are presented is enough to lower Top-10, so the generator is using them rather than the identities of the context molecules alone. As for why we exploit them by predicting the expansion: the property of the molecules already in a context is not the property of what will be generated from it, so the score must apply to the expansion rather than to the context as it stands. The Mean-Property Selection ablation (58.09 against MolWorld's 62.11) bears this out directly.

Q3. The cost has the following structure. The current graph is encoded once per round and the resulting fixed-length representation is reused across the pool, so graph size enters through that single encoding rather than through each candidate. The neural part then grows linearly in the pool size, with an additional O(∣C∣log⁡∣C∣)O(|C| \log |C|) for the ranking itself. Relative to direct ranking, MolWorld performs one extra dynamics forward pass, and that is the whole of the difference at the selection stage.

In practice, over a full DRD3 docking run (1,000 oracle calls, 90 rounds, 3.92 hours of iteration), docking scoring accounts for 80.6% of the time, generation for 16.4%, and ranking and selection for 1.09%. Selection takes 1.2 s per round against 0.99 s for direct ranking.

W3 & Q2. The revision adds a Formal Framework appendix at the end, deriving the method from the stated objective; numbering of the main text and existing appendices is unchanged. The averaged-representation part of Q2 is answered under W2.

Eq. (3), the additive approximation. We agree this is an approximation, and the appendix states its direction and source. Whether a candidate is accepted depends only on its MMP relation to a molecule already in VtV_t (Eq. 4), not on what the rest of the batch produced, so each context is valued on its own and the pool ranked by that value. The value of a batch, however, is not the sum of its members': nearby contexts produce molecules in common, which JJ, a top-κ\kappa utility, counts once where the sum counts twice. The approximation is one-sided, E[J(⋃ZX~t,Z)]≤∑ZQ(Gt,Z)E[J(\bigcup_Z \tilde{X}_{t,Z})] \le \sum_Z Q(G_t, Z), so the ranking optimises an upper bound on what a batch is worth, the gap being exactly the duplicated output. The penalties on repeated use of the same nodes during pool construction (Eqs. 16–17) leave few near-duplicates available to be selected together.

Eq. (8), the local completion criterion. Nlocal(Z)N_{local}(Z) is not a separately imposed criterion but is defined implicitly by the training data: the generator is fitted on pairs (Z,x)(Z, x) where ZZ is an anchor context of the reference MMP graph and xx a molecule adjacent to it, held out from the subgraph ZZ spans (Appendix B.4). Eq. (8) expresses that supervision target rather than a condition to be checked independently.

Eq. (9), the transition to token-level decoding. This is stated imprecisely and we will correct it. Each context molecule is encoded into a token sequence; a masked mean over its valid tokens gives one node vector per molecule; the GNN passes messages over the context edges and returns an updated vector per molecule; this vector is added back to every token of that molecule, followed by layer normalisation; the resulting sequences are concatenated across the context. The decoder cross-attends to that concatenated sequence with padding masked, not to a single pooled vector. Pooling serves only to produce the GNN node inputs and the per-molecule representations of the anchor encoder.

Eqs. (6) and (12), token decorrelation. LdivL_{div} is a regulariser penalising similarity between normalised graph tokens. Complementarity is its motivation, not a property it establishes, and we will adjust the wording accordingly.

Eqs. (16)–(17), property scales. The reviewer is right: on docking we neither normalise the scores nor retune the penalty coefficients, so the penalties are weaker relative to the spread of scores there than on the property tasks. They act only on pool construction — the BB selected contexts are determined by the scorer alone, with no further reranking (Appendix B.3.1) — and MolWorld still attains the best AUC@10 and AUC@100 on all three docking targets (Table 2).

Eq. (18) against Eq. (2). These are aligned. In the appendix J(S)=∑i=1κg(i)(S)J(S) = \sum_{i=1}^{\kappa} g_{(i)}(S) with g(i)(S)=0g_{(i)}(S) = 0 for i>∣S∣i > |S|, that is, zeros are filled in when fewer molecules are accepted — precisely what the fixed denominator of three in Eq. (18) does. The label is therefore a constant multiple of JJ at κ=3\kappa = 3 and does not reorder actions.

Eq. (19), statistical consistency. We do not claim statistical consistency for this ranking loss. The appendix gives a different guarantee: under a Lipschitz condition on the scorer and bounded dynamics and scorer errors, the estimate and the true context value differ uniformly over the pool, and the selection regret is bounded by twice that difference. Selection quality is therefore controlled by the dynamics error, which is measured in Appendix B.2.3. The Lipschitz constant is not estimated, so this describes a dependence rather than a numerical guarantee.

W2. The two-step decomposition rests on a simple fact: the expansion a context induces depends on that context, not on which property is being optimised. The same context yields the same structures whatever the objective; only their worth changes. One dynamics model therefore serves all four objectives and a new property requires retraining only the scorer. Freezing follows from this division of labour rather than being imposed on top of it.

Taking an expectation rather than a best case is a budget decision. Each round selects a batch of contexts and samples several times from each, so what matters is the quality of that batch, not the best single draw within it. Favouring high-variance contexts amounts to a wager — sensible when the budget allows repeated attempts, less so when it is fixed and small. We therefore define the value of a context as the expected utility of its output, regress the dynamics on the expected representation of the expansion, and label the scorer with the expected quality of the molecules it gets accepted.

Prediction accuracy and ranking quality we did keep apart: checkpoints in Appendix B.3.2 are chosen by normalised selection regret, not by regression error. The numbers were simply not reported. On the validation set:

PropertyScorerRandom selection
QED0.08580.3796
JNK30.31490.6482
DRD20.41570.7873
GSK3β0.31030.6597

Regret is well below random selection on all four objectives, so enough property-relevant signal survives encoding, averaging and prediction to rank actions. Proposition 2 in the new appendix states the relation formally: under a Lipschitz condition and bounded errors, selection regret is controlled linearly by the dynamics error.

What freezing buys is reuse. One encoder and one dynamics model support four lightweight scorers, each of which reaches the best or near-best result on its objective (Table 1) — an outcome that would be hard to obtain four times over if the intermediate representation had lost the distinctions that matter.

W4. STONED does vary several things at once, so we have run two further controls, each isolating one claim. For connected-context modelling, we hold the context molecules, the backbone and the training protocol fixed and change only the MMP relations inside the context (the table under W1). The same molecules reach the generator in every setting; shuffling or removing the relations between them is enough to lower Top-10. For structural prediction, we leave generation untouched and vary only how contexts are picked out of the pool. Direct Context Ranker skips the expansion prediction and ranks straight from the graph and context representations; anchor-only still predicts the expansion but cannot see the current graph. Averaged over the four properties on PMV21, three seeds, with standard deviations now in Figure 4:

MethodAUC@10AUC@100
Random Selection0.5727 ± 0.01420.5132 ± 0.0107
Mean-Property Selection0.5809 ± 0.01010.5205 ± 0.0076
Direct Context Ranker0.5889 ± 0.01740.5225 ± 0.0160
Anchor-only dynamics0.5885 ± 0.00390.5220 ± 0.0004
MolWorld0.6211 ± 0.00740.5516 ± 0.0059

All four controls sit within a band of 0.57 to 0.59; the full model sits well outside it. Anchor-only falls back into the band, which places the gain in predicting the expansion against the current state of the search.

We thank the reviewer for pressing on the formal side of the method, and hope the appendix and the controls above resolve the concerns raised.

Thank you for the detailed response. The following concerns remain unresolved, so I keep my score.

  1. Appendix C.4.1 fixes the pool of 300 candidate contexts, but each edge configuration selects its own 20 contexts for generation. Changing the edges therefore affects both context selection and the generator’s input. The results show a benefit for the overall pipeline, but leave open whether it comes from better context ranking, better generation from the same contexts, or both.
  2. Optimizing expected Top-3 utility and predicting a mean latent representation are distinct design choices. Different expansion distributions can share the same mean representation while having different expected Top-3 utilities. Proposition 2 assumes that the scorer can estimate true context value from this mean with a uniform error bound across the candidate pool. This is precisely the assumption I asked the authors to justify. The reported results still do not explain how latent prediction accuracy affects action ranking.
  3. I agree that, with zero padding and k=3, the expected label in Eq. (18) is proportional to single-context utility in expectation. It remains unclear how training Eq. (19) on finite-sample labels supports batch selection and improves the utility of the accumulated candidate set in Eq. (2) under the oracle budget.
  4. The DRD3 timing shows that ranking accounts for a small fraction of one docking run. The reported per-round selection times leave open how end-to-end cost scales with graph and pool size, including pool construction and MMP validation, and how it compares with direct MMP search.
Reviewer_JrQ5OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper introduces MolWorld, a framework for molecular optimization that seeks candidates with favorable target properties while preserving reachability from an initial set of molecules through matched molecular pair (MMP) relations. Optimization proceeds by iteratively expanding a molecule-transfer graph. A latent world model predicts representations of local graph expansions from the current graph and candidate anchor contexts, with structural dynamics shared across property objectives. A property-specific scorer uses these predictions to select contexts. A separate generator, trained on local graph-completion tasks, uses the selected contexts’ molecular structures and MMP relations to propose candidates. Those with verified MMP connections to the current graph are added and can be reused in subsequent iterations. Experiments cover four property objectives on ZINC250 and PMV21 and three protein–ligand docking tasks under fixed online oracle budgets. The authors report better running top-k AUC results and lower isolated-node ratios than the evaluated baselines under the MMP-reachability criterion.

Critical strengths
  1. Iterative MMP-graph expansion preserves structural traceability and allows newly generated molecules to guide subsequent optimization.
  2. Using expansion predictions to select contexts is well motivated and improves over direct context ranking in the reported ablations.
  3. Property and docking experiments report better AUC results and lower isolated-node ratios than the evaluated baselines under the stated online budgets and reachability criterion.
Critical weaknesses
  1. The paper does not explain why each context contains exactly five molecules or report a context-size ablation. It is therefore unclear how this choice affects optimization performance, structural connectivity, and computational cost.
  2. A baseline that directly searches through MMP transformations under the same reachability rules and online budget is missing. Such a comparison would help establish the benefit over direct MMP-based search.
  3. Despite the direct-ranker comparison, missing anchor-only dynamics controls and ablation uncertainty estimates limit attribution of the world model’s gains.
  4. The STONED ablation replaces the learned generator with single-molecule mutation. It therefore does not isolate the benefit of MMP edges within the generator.
  5. The paper explicitly separates offline training costs from the online budget, but does not quantify the cost of obtaining target-property labels for scorer training. This makes the method’s total cost-effectiveness difficult to assess.
  6. The main text would benefit from a brief description of the dynamics architecture, which is detailed only in Appendix. The correspondence between the molecular example and graph nodes in Figure 1, and the prediction–selection–generation flow in Figure 2, are not clear to follow.
Questions for authors
  1. Why is each context restricted to five molecules? How sensitive are AUC, Iso, and computational cost to context size?
  2. How does the full dynamics model compare with an anchor-only version?
  3. How does MolWorld compare with a simple MMP-transformation search baseline using the same initial graphs, reachability rules, and online oracle budgets?
  4. What happens when the MMP edges supplied to the generator are removed or shuffled, while keeping the input molecules, backbone, and training protocol unchanged?
  5. What are the offline labeling and training costs for each property scorer, and how are these costs amortized across repeated searches?
2 comments

We thank the reviewer for the careful review and the constructive questions.

W1 & Q1. The choice of kk is set by a trade-off between the cost of a context and the structural neighbourhood it can cover. Conditioning on the full graph is not viable: with the conditioning set entered as node-level representations, each decoding step attends over every molecule, so a single generation costs Ω(∣V∣⋅T)\Omega(|V| \cdot T) for a decoding length TT and a current node set VV, which grows monotonically with the iterations, leaving this cost unbounded. We therefore condition on a fixed-size connected subgraph ZZ with kk nodes, giving a cost of O(k⋅T)O(k \cdot T) per generation, independent of ∣V∣|V|.

On the lower side, the locality of the MMP relation makes a fixed-size context tenable, but ZZ is the entirety of what the generator conditions on, so the smaller kk is, the narrower the local structure it can cover. In practice, analogue series are delimited at this scale: [1] defines a matched molecular series as a congeneric series with a minimum of five compounds that are all MMPs of one another. We therefore set k=5k=5, as the smallest context that carries the structural information of a conventional analogue series.

W2 & Q3. We added a baseline that searches MMP transformation space directly, using the same initial graphs, reachability rules and oracle budget. A transformation library is built from the matched molecular pairs of the provided molecules. Each round draws a batch of molecules from the current graph and applies several random valid MMP transformations to each, so every candidate is by construction an MMP of a molecule already in the graph and the isolated-node ratio is exactly zero. The two variants differ only in how the batch is drawn — at random, or greedily by current property value. Results on PMV21 (mean ± s.d. over three seeds):

AUC@10

MethodGSK3βQEDJNK3DRD2
MMP-Search (random)0.3016 ± 0.01620.6682 ± 0.03500.1909 ± 0.01290.6007 ± 0.0706
MMP-Search (greedy)0.3965 ± 0.00710.8095 ± 0.01430.2240 ± 0.01830.8368 ± 0.0239
MolWorld0.4225 ± 0.01590.8892 ± 0.01190.2710 ± 0.02860.9018 ± 0.0063

AUC@100

MethodGSK3βQEDJNK3DRD2
MMP-Search (random)0.2052 ± 0.00940.4775 ± 0.02060.1178 ± 0.00620.2813 ± 0.0530
MMP-Search (greedy)0.3160 ± 0.00150.6701 ± 0.01000.2025 ± 0.00370.6326 ± 0.0199
MolWorld0.3490 ± 0.02200.8194 ± 0.00980.2175 ± 0.02240.8205 ± 0.0047

The greedy variant is a strong baseline — on AUC@10 it beats every baseline method in Table 7 on GSK3β, QED and DRD2. Since both variants satisfy reachability exactly, the remaining gap cannot come from the acceptance rule. Rule-based enumeration guarantees a candidate is a valid MMP, but says nothing about which molecules and which transformations are worth an oracle call. The random-to-greedy gap shows what property-directed selection of starting molecules is worth; the further gap to MolWorld widens from AUC@10 to AUC@100 (0.080 to 0.149 on QED, 0.065 to 0.188 on DRD2), suggesting greedy hill-climbing can reach a few high-scoring molecules but not a high-quality top-100 set.

W3 & Q2. Standard deviations are now shown in Figure 4. We have also trained an anchor-only dynamics model, which sees the context but not the current graph. Averaged over the four properties on PMV21, three seeds:

MethodAUC@10AUC@100
Direct Context Ranker0.5889 ± 0.01740.5225 ± 0.0160
Anchor-only dynamics0.5885 ± 0.00390.5220 ± 0.0004
MolWorld0.6211 ± 0.00740.5516 ± 0.0059

Without the graph state the model does no better than direct ranking. The gain comes from predicting the expansion in the light of where the search currently stands, not from the extra stage.

Table 4 makes the complementary point and may have been easy to miss in the appendix. Holding the graph inputs and targets fixed and cyclically shifting the anchor contexts within each validation batch raises the smooth-L1 error from 0.0488 to 0.2148, worse than the 0.2083 of a mean-target baseline. Accurate prediction thus needs both the context and the state.

[1] Keefer CE, Chang G. The use of matched molecular series networks for cross target structure activity relationship translation and potency prediction. Medchemcomm. 2017 Oct 11;8(11):2067-2078. doi: 10.1039/c7md00465f. PMID: 30108724; PMCID: PMC6072506.

W4 & Q4. We agree that STONED changes the generator, the context size and the edges at once. We have added a control that changes only the edges.

To avoid confounding from graph evolution across rounds, we run a single round over a fixed candidate context pool of 300 contexts, shared across the three seeds of a given property. Each setting — real edges, randomly rewired edges, no edges — uses those edges consistently through the round: the generator conditions on them when producing 16 candidates from each of the top 20 contexts, and because the anchor encoder is taken from the generator and fuses over the same edges, the world model's scoring of the pool is affected by them as well. The context molecules, the backbone and the training protocol are unchanged, so the edges are the only variable.

Top-10

PropertyRealShuffledRemoved
QED0.7845 ± 0.00000.7299 ± 0.00650.7493 ± 0.0074
JNK30.1177 ± 0.00120.1003 ± 0.00060.1003 ± 0.0042
DRD20.8592 ± 0.00130.8399 ± 0.01450.8382 ± 0.0073
GSK3β0.2653 ± 0.00470.2557 ± 0.00120.2510 ± 0.0131

Iso (%)

PropertyRealShuffledRemoved
QED0.75 ± 0.350.94 ± 0.660.96 ± 0.33
JNK33.64 ± 0.734.36 ± 1.213.92 ± 0.62
DRD24.78 ± 1.356.62 ± 1.255.59 ± 2.09
GSK3β1.16 ± 0.541.32 ± 0.681.27 ± 1.13

Real edges give the highest Top-10 and the lowest Iso on all four properties, though the Iso margins are narrow. This is to be expected: the pool is itself built from real MMP edges, so the same molecules reach the generator in all three settings and are already related to one another. Shuffling or removing the edges changes only whether that relation is stated explicitly, not the structural proximity the set implies — the design understates what the edges contribute. Even so, altering how the relations are presented is enough to lower Top-10, which indicates the generator uses them rather than the identities of the context molecules alone.

W5 & Q5. The two task families differ here, so we report them separately.

Docking. The docking oracle is expensive, so the scorer is trained entirely online with no offline property labelling at all. MolWorld still attains the best AUC@10 and AUC@100 on all three targets (Table 2), so the method does not depend on offline labels.

Property. For the four property objectives the scorer is pretrained offline, with labelling covering roughly 355k molecules per property. These oracles are TDC's pretrained predictors, so a single call is computationally negligible — a different regime from docking. Fitting one scorer takes 613.4 ± 79.6 seconds (mean ± s.d. over 12 runs) on a single NVIDIA RTX A6000, including training and per-epoch validation but excluding data preparation, checkpoint I/O, final evaluation and failed attempts. This is a one-off cost per property: the scorer is reused unchanged when the same property is searched from a different initial graph or dataset, with no relabelling.

This is in line with current practice in the field. As noted by InVirtuoGen [4], both GenMol [2] and f-RAG [3] initialise their populations by screening the whole of ZINC250k, amounting to roughly 250k oracle calls before optimization begins. Their own papers describe the same thing: GenMol ranks its initial fragment vocabulary by the mean target-property value of the molecules containing each fragment, computed before optimization, and f-RAG likewise builds a target property-specific fragment vocabulary and seeds its population with the top-N molecules under the target property.

W6. Thank you for pointing this out. In the revision we will add a brief description of the dynamics architecture to Section 3.3, redraw Figure 1 to make the correspondence between the molecular example and the graph nodes explicit, and redraw Figure 2 so that the prediction–selection–generation flow is easier to follow.

We are grateful for the reviewer's questions, which prompted several of the experiments above, and hope our answers are satisfactory.

[2] Lee, Seul, et al. "Genmol: A drug discovery generalist with discrete diffusion." arXiv preprint arXiv:2501.06158 (2025).

[3] Lee, Seul, et al. "Molecule generation with fragment retrieval augmentation." Advances in Neural Information Processing Systems 37 (2024): 132463-132490.

[4] Kaech, Benno, et al. "Refine drugs, don’t complete them: Uniform-source discrete flows for fragment-based drug discovery." International Conference on Learning Representations. Vol. 2026. 2026.

Comments on the paper

2 comments
~Priyansh_Singhal1OpenReview ↗

Request to make the code repo public

Could the authors please release the code repository publicly?

Thanks

Thank you for your interest in MolWorld. We will release the code once the paper is published at a venue.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.