acceptodds
Under review as a conference paper at ICLR 2027

Coupling Synthesis and 3D Geometry for Pocket-Conditioned Molecular Generation

Abstract

Structure-based molecular generation can propose ligands for a target protein, but connecting their synthetic construction with their placement in the binding pocket remains challenging. We present SynTrace, a supervised model that couples synthesis decisions to the evolving three-dimensional state of a molecule in its pocket. The central idea is to learn reaction choices and coordinate evolution from a shared geometric representation, so that each synthesis decision can use the spatial state produced by earlier generation steps. We train this model by aligning synthesis routes with docked product poses, which provide paired action and coordinate targets without reward-based policy optimization. On 100 CrossDocked2020 test pockets, SynTrace generates 10,000 molecules with a median redocking score of −8.45 kcal/mol, a 45.49% AiZynthFinder solve rate using Enamine stock, and diversity of 0.889. Interventions show that synthesis decisions respond to relative ligand–pocket geometry and identify spatial interactions as the main source of pocket sensitivity on reference decision states.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Structure-based drug design, Synthesizable molecular generation, 3D molecular generation, Synthesis route supervision, Pocket-conditioned generation, Flow matching

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_W73uOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes SynTrace, a pocket-conditioned molecular generation framework that jointly models synthesis decisions and 3D molecular geometry. It employs supervised route–pose alignment and a shared geometric network to generate molecules with explicit synthesis routes. Experiments on CrossDocked2020 demonstrate competitive docking performance and synthetic accessibility.

Critical strengths
  • The paper integrates synthesis planning and pocket-conditioned 3D molecular generation into a unified framework.

  • SynTrace achieves a 45.49% retrosynthesis solve rate and competitive molecular diversity compared with existing baselines.

Critical weaknesses
  • Joint synthesis and 3D molecular generation have already been explored in CGFlow/3DSynthFlow and SynCoGen. The paper does not sufficiently establish its methodological advantages over these approaches.

  • SynTrace does not consistently outperform existing methods, particularly in docking and direct-pose quality. The evaluation is limited to 100 pockets from CrossDocked2020, with single-seed experiments. Additional independent benchmarks and stronger SOTA comparisons are needed.

  • Although synthesis decisions respond to spatial information, the experiments do not convincingly demonstrate that geometric conditioning improves pocket-specific chemical quality.

Questions for authors
  1. Can the authors include additional benchmarks and stronger comparisons with 3DSynthFlow and SynCoGen?

  2. Can the authors clarify the methodological novelty over existing joint synthesis–structure generation frameworks?

  3. Can the authors provide stronger evidence that geometric coupling improves pocket-specific molecular quality?

3 comments
AuthorsOpenReview ↗

Response to Reviewer W73u (1/3)

We thank the reviewer for the constructive feedback. For the baselines, we retrained 3DSynthFlow on our building-block library and added both released SynCoGen checkpoints under one protocol, together with receptor-aware PoseBusters and a size-matched redocking comparison. For the evaluation, we added a second benchmark and trained every ablation with three independent seeds. On whether the geometry improves pocket-specific choices, we added two experiments that score the building blocks the policy selects. Changes in the revised PDF are marked in magenta, and the new analyses are in Appendices A.3, B.2, C.4, C.5 and C.6.

W1 / Q2. Methodological advantages over CGFlow / 3DSynthFlow and SynCoGen

3DSynthFlow trains a coordinate model first and freezes it while the construction policy is learned by reward, so the synthesis decision and the 3D placement are never optimized together. SynCoGen does optimize them together, but it conditions on pharmacophore points taken from a reference ligand rather than on the receptor itself. Its coordinates are also generated in their own frame rather than in the pocket, and we could not verify a transformation into the receptor frame, so its output has no pose to check against the protein. SynTrace does both at once: one network learns the synthesis decision and the 3D placement together, conditioned directly on receptor atoms. The route–pose supervision is what makes that affordable: a single docked pose of the final product supervises every step of the route, so no intermediate ever has to be docked and no reward model is needed. One pass therefore returns a molecule, an executable route for it, and a pose that can be evaluated as it stands.

Two sets of results support this. The first removes the coupling inside SynTrace. Separate trains the action and coordinate predictions as two independent networks that share nothing; NoRel keeps the shared network but stops the action head from reading the ligand's position relative to the pocket, at an unchanged parameter count. Both use the same data and training budget as the full model, and all three were trained three times (Table 2; per-run values in Appendix C.6, Table 10):

Score ↓\downarrow (kcal/mol)Min ↓\downarrow (kcal/mol)Dock ↓\downarrow (kcal/mol)PB ligand ↑\uparrow (%)PB complex ↑\uparrow (%)
Full−6.32±0.13\mathbf{-6.32}_{\pm 0.13}−7.35±0.08\mathbf{-7.35}_{\pm 0.08}−8.41±0.03\mathbf{-8.41}_{\pm 0.03}88.10±0.28\mathbf{88.10}_{\pm 0.28}67.74±0.55\mathbf{67.74}_{\pm 0.55}
NoRel−5.77±0.05-5.77_{\pm 0.05}−6.84±0.15-6.84_{\pm 0.15}−8.31±0.08-8.31_{\pm 0.08}85.18±0.3485.18_{\pm 0.34}64.68±0.4864.68_{\pm 0.48}
Separate−5.67±0.04-5.67_{\pm 0.04}−6.90±0.05-6.90_{\pm 0.05}−8.24±0.06-8.24_{\pm 0.06}81.79±1.8181.79_{\pm 1.81}58.20±2.3058.20_{\pm 2.30}

Splitting the two predictions costs 0.650.65 kcal/mol of direct score and 9.59.5 points of receptor-aware PoseBusters, at twice the parameters; keeping one network but hiding the relative geometry still costs 0.550.55 kcal/mol and 3.13.1 points. Both are several times the run-to-run spread, and both appear on the pose the model emits rather than after redocking, where all three lie within 0.170.17 kcal/mol.

The second set compares complete systems. We retrained 3DSynthFlow on the 274274k Enamine library we share with RxnFlow, and scored it and both SynCoGen checkpoints on the same 100100 pockets with our own pipeline (Table 1; its SynTrace row is run 1 of the three above):

Score ↓\downarrow (kcal/mol)Dock ↓\downarrow (kcal/mol)PB ligand ↑\uparrow (%)PB complex ↑\uparrow (%)Uniq. ↑\uparrow (%)AiZynth ↑\uparrow (%)
SynTrace−6.18\mathbf{-6.18}−8.38\mathbf{-8.38}87.80\mathbf{87.80}67.19\mathbf{67.19}99.4199.4145.49\mathbf{45.49}
3DSynthFlow (retrained)−3.26-3.26−8.36-8.3639.5239.5219.0019.0017.7317.7319.8019.80
SynCoGen (uncond.)—−8.36-8.3684.1384.13—99.53\mathbf{99.53}21.6821.68
SynCoGen (cond.)—−8.24-8.2456.3256.32—99.1499.1419.7719.77

Redocking re-searches the pose, so all four rows land within 0.140.14 kcal/mol of each other there. The gap opens once the pose is taken as the model emits it — 67.19%67.19\% against 19.00%19.00\% on receptor-aware PoseBusters — and again on usable output, where 99.41%99.41\% of our molecules are distinct against 3DSynthFlow's 17.73%17.73\% and AiZynthFinder solves more than twice as many of ours as of any baseline in the table.

AuthorsOpenReview ↗

Response to Reviewer W73u (2/3)

W2 / Q1 (1/3). Direct-pose quality

We added receptor-aware PoseBusters (PB complex), which checks the generated pose against its receptor before any minimization or redocking, for every method that emits receptor-frame coordinates (Table 1):

MethodProduces a routePB complex ↑\uparrow (%)
MolCRAFTno84.5284.52
DecompDiffno76.6276.62
SynTraceyes67.19\mathbf{67.19}
TargetDiffno60.0560.05
3DSynthFlow (retrained)yes19.0019.00
Pocket2Molno6.976.97
RxnFlowyesno coordinates
SynCoGen (uncond.)yesown frame, not verifiable
SynCoGen (cond.)yesown frame, not verifiable

MolCRAFT and DecompDiff are ahead of us, and MolCRAFT also has the better direct Score (−6.25-6.25 against −6.18-6.18). The explanation is the constraint we accept: those models place atoms freely, so only the pocket limits their geometry, whereas every SynTrace molecule must in addition be reachable by an executable two- or three-step route from catalogue building blocks. That constraint restricts the graph and, through it, the geometry — and it is also what produces the solve rate we lead on, 45.49%45.49\%, against 14.00%14.00\% for MolCRAFT and 16.10%16.10\% for DecompDiff. And among the methods that accept the same constraint, SynTrace is the only one whose emitted pose passes a receptor-aware check more often than not.

For redocking, the size effect is a major factor: the larger the molecule, the better the score it tends to receive — a point MolCRAFT's own evaluation makes about Vina Dock (Qu et al., 2024). The synthesis-aware baselines all generate larger molecules than we do, with a median of 3131 heavy atoms for RxnFlow and 2727 for the retrained 3DSynthFlow against 2424 for SynTrace, and in this evaluation the median redocking score falls monotonically across five heavy-atom windows for all three methods built on our library (Appendix B.2, Table 4). Matching the heavy-atom distributions exactly within each pocket removes that confound (Appendix B.2, Table 5):

BaselinePaired advantage to SynTrace (kcal/mol)95% CIPockets favouring SynTrace
RxnFlow+0.186+0.186[0.114,0.270][0.114, 0.270]67/10067/100
3DSynthFlow (retrained)+0.052+0.052[0.001,0.103][0.001, 0.103]58/10058/100
SynCoGen (uncond.)+0.819+0.819[0.743,0.898][0.743, 0.898]100/100100/100

W2 / Q1 (2/3). Independent training seeds

The ablation results quoted above are no longer single-seed: Full, NoRel and Separate were each trained three times, and the table in W1 reports the mean and standard deviation across those runs, with all six metrics for all nine runs in Appendix C.6, Table 10. On every metric the separation between the full model and either ablation is several times the run-to-run spread. On the size of the evaluation, the 100100 generation pockets come from the official CrossDocked2020 test split and span 9393 pocket families, none shared with the training families (Appendix A.3).

W2 / Q1 (3/3). A second benchmark: LIT-PCBA

We evaluated fixed checkpoints on the 1515 LIT-PCBA targets used for pocket-specific optimization by RxnFlow and 3DSynthFlow, with the same receptors, boxes and scoring code for every method, and redocked the benchmark's own known actives under that protocol as a reference point (Appendix C.5, Table 8):

Dock ↓\downarrow (kcal/mol)LE ↓\downarrow (kcal/mol per atom)HAPB ligand ↑\uparrow (%)AiZynth ZINC ↑\uparrow (%)
SynTrace−8.821-8.821−0.375\mathbf{-0.375}24.024.092.67\mathbf{92.67}43.2043.20
RxnFlow−9.396\mathbf{-9.396}−0.302-0.30232.032.0—11.0711.07
3DSynthFlow (retrained)−8.873-8.873−0.315-0.31528.028.040.8040.8017.2017.20
SynCoGen (cond.)−8.850-8.850−0.219-0.21941.041.056.1256.1227.7427.74
Known actives−8.019-8.019−0.332-0.33225.025.0—44.8744.87

The size effect appears again: RxnFlow leads on raw redocking with molecules a third larger. Per heavy atom, SynTrace has the better ligand efficiency on every one of the 1515 targets, and is the only method whose ligand efficiency exceeds that of the known actives (per-target values in Appendix C.5, Table 9). Against the retrained 3DSynthFlow it is +51.9+51.9 points on intramolecular validity ([44.9,59.3][44.9, 59.3]) and +26.0+26.0 points on ZINC-stock solve rate ([22.1,29.9][22.1, 29.9]). Against RxnFlow it has the higher solve rate under both stocks, 46.53%46.53\% against 40.73%40.73\% with Enamine and 43.20%43.20\% against 11.07%11.07\% with ZINC. Matching heavy-atom counts within each target, as on CrossDocked, also favours SynTrace on redocking: by 0.1600.160 kcal/mol over the retrained 3DSynthFlow ([0.014,0.325][0.014, 0.325]) and by 0.4300.430 over conditional SynCoGen ([0.138,0.734][0.138, 0.734]).


Qu, Y. et al. MolCRAFT: Structure-Based Drug Design in Continuous Parameter Space. ICML 2024, PMLR 235, 41749–41768.

AuthorsOpenReview ↗

Response to Reviewer W73u (3/3)

W3 / Q3. Evidence that geometric conditioning improves pocket-specific molecular quality

We added two experiments that score the completed products of the blocks the policy selects. The first asks whether the choice follows the pocket, using intermediates that occur in more than one test protein; the second compares the full model with NoRel on intermediates the model generates itself.

We collected every pair where the same intermediate and the same reaction slot appear in two different pockets, A and B — 964964 pairs in all (Appendix C.4). In 86.7%86.7\% of them the model picks one block for A and a different one for B; for NoRel this happens in only 34.1%34.1\%. We then built all four products, both blocks in both pockets, and scored each one with that pocket's own docked reference core held in place. If the model is responding to the pocket, the block it picked for A should beat the other one in A, and the same should hold in B. Adding the two differences, that is what we see in 67.1%67.1\% of the 831831 pairs that returned all four scores, by a key-weighted mean of −1.60-1.60 kcal/mol ([−2.66,−0.57][-2.66, -0.57]). Figure 5 shows one such pair: for COAA, a pantothenate kinase with a polar site, the model picks a basic piperidine; for PTGIS, a P450 with a lipophilic heme-lined channel, an apolar phenyl; and the constrained scores reverse between the two (−10.45-10.45 against −5.30-5.30, and −4.26-4.26 against −7.85-7.85 kcal/mol).

The second experiment runs on intermediates the full model generated itself, rather than on poses taken from docked reference complexes, and compares it with NoRel — the same network with the action head's view of the relative geometry switched off, at an unchanged parameter count (Section 4.5). At each intermediate both models choose from the same legal blocks at the same reaction slot, and we complete the products of the four each one prefers. To score them we dock the intermediate separately and hold its pose fixed, so only the newly added extension can move (Figure 3a–c; protocol in Appendix C.6). Across 9393 equally weighted pocket families the mean Full-minus-NoRel difference is −0.84-0.84 kcal/mol ([−1.41,−0.27][-1.41, -0.27]), with the full model scoring lower in 7171 of them (Figure 3d).

Concluding remarks

We thank the reviewer again for the time and for the detailed requests, which shaped most of the revision: a second benchmark on LIT-PCBA, three independent training runs for every ablation, receptor-aware PoseBusters for every method that emits receptor-frame coordinates, a retrained 3DSynthFlow, and two experiments that score the pocket-specific choices the policy makes. We hope these additions address the concerns raised, and we would be grateful if the reviewer could reconsider the score in light of them. We are of course happy to provide any further details or analyses that would help.

Reviewer_Jf45OpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

SynTrace jointly generates molecules, construction routes, and pocket-conditioned poses. It pairs enumerated reaction routes with docked product poses and uses atom tracking to supervise intermediate states. A shared geometric network predicts synthesis actions and coordinates. Evaluation covers molecular generation, computational route recovery, and the sensitivity of action predictions to spatial conditions.

Critical strengths

The route pose supervision is a useful design. Atom tracking connects changes in the atom set during reaction execution with coordinate targets for atoms retained in the final product. This allows one product pose to provide supervision across several construction stages. Joint training and explicit pocket conditioning provide concrete methodological differences. Both prediction tasks update the same geometric representation, and pocket atoms directly inform the model about the surrounding occupied space. These choices distinguish the approach from the separately trained components in 3DSynthFlow and the pharmacophore conditioning in SynCoGen.

Critical weaknesses

My main reservation is the significance of the methodological advance. The supervision scheme is useful, while the broader construction and coordinate generation framework has a close precedent in CGFlow. The added value of jointly training the two predictions remains less clear. The interventions establish that geometry affects action probabilities, but do not establish that this feedback improves pocket-specific chemical choices. Appendix C.1 explicitly leaves that benefit unresolved.

Questions for authors
  1. For a fixed reaction intermediate, can you explain a concrete case where local pocket geometry leads the policy to favor different building blocks?
  2. What evidence connects that preference to better spatial compatibility of the completed product?
4 comments
AuthorsOpenReview ↗

Response to Reviewer Jf45 (1/2)

We thank the reviewer for the close reading and the precise summary of the method. The questions on what the joint training adds and on pocket-specific choices led us to add an ablation with independent seeds, a head-to-head comparison against a retrained 3DSynthFlow, and two experiments that score the building blocks the policy selects. Changes in the revised PDF are marked in magenta, and the new analyses are in Appendices B.2, C.4 and C.6.

W1. The methodological advance, and what the coupling is worth

In SynTrace a single docked pose of the final product supervises every step of its construction route: route replay supplies the action taken at each step, and atom tracking supplies the coordinate target for the atoms each intermediate keeps. Both losses update one representation, and receptor atoms enter it directly, so the action head chooses the next reaction and building block while reading the ligand's current geometry inside the pocket. As the reviewer observes, this is where we differ from the separately trained components of 3DSynthFlow and from pharmacophore conditioning. How much that coupling is worth is an empirical question, and the revision answers it with two ablations that take it away and with experiments that score the blocks the policy selects (Q1 and Q2). Appendix C.1 could not settle that last point: what was null there is the gap between two candidate rankings (0.1880.188 against 0.1710.171 with a replaced pocket), not the quality of the block it picks.

We measured this with two ablation models. Separate trains the action and coordinate predictions as two independent networks that share nothing. NoRel keeps the shared network but stops the action head from reading the ligand's position relative to the pocket, at an unchanged parameter count. Both use the same data and training budget as the full model, and all three were trained three times (Table 2; per-run values in Appendix C.6, Table 10):

Score ↓\downarrow (kcal/mol)Min ↓\downarrow (kcal/mol)Dock ↓\downarrow (kcal/mol)PB ligand ↑\uparrow (%)PB complex ↑\uparrow (%)
Full−6.32±0.13\mathbf{-6.32}_{\pm 0.13}−7.35±0.08\mathbf{-7.35}_{\pm 0.08}−8.41±0.03\mathbf{-8.41}_{\pm 0.03}88.10±0.28\mathbf{88.10}_{\pm 0.28}67.74±0.55\mathbf{67.74}_{\pm 0.55}
NoRel−5.77±0.05-5.77_{\pm 0.05}−6.84±0.15-6.84_{\pm 0.15}−8.31±0.08-8.31_{\pm 0.08}85.18±0.3485.18_{\pm 0.34}64.68±0.4864.68_{\pm 0.48}
Separate−5.67±0.04-5.67_{\pm 0.04}−6.90±0.05-6.90_{\pm 0.05}−8.24±0.06-8.24_{\pm 0.06}81.79±1.8181.79_{\pm 1.81}58.20±2.3058.20_{\pm 2.30}

Giving up the shared representation costs 0.650.65 kcal/mol of direct score and 9.59.5 points of receptor-aware PoseBusters, at twice the parameters; keeping the shared network but cutting only the action head's view of relative geometry still costs 0.550.55 kcal/mol and 3.13.1 points. Both effects are several times the run-to-run spread. The gap is widest on the pose the model emits and narrowest after redocking, where all three lie within 0.170.17 kcal/mol. Redocking re-searches the pose and erases those placement differences, so what the coupling buys is the placement itself.

3DSynthFlow trains its coordinate model first and freezes it during reward-based policy learning, which makes it the natural comparison. We retrained it on the 274274k Enamine library shared with SynTrace and RxnFlow and scored it under our protocol on the same 100100 pockets (Table 1; its SynTrace row is run 1 of the three above):

Score ↓\downarrow (kcal/mol)Dock ↓\downarrow (kcal/mol)PB ligand ↑\uparrow (%)PB complex ↑\uparrow (%)Uniq. ↑\uparrow (%)AiZynth ↑\uparrow (%)
SynTrace−6.18\mathbf{-6.18}−8.38\mathbf{-8.38}87.80\mathbf{87.80}67.19\mathbf{67.19}99.41\mathbf{99.41}45.49\mathbf{45.49}
3DSynthFlow (retrained)−3.26-3.26−8.36-8.3639.5239.5219.0019.0017.7317.7319.8019.80

Once a pose search is allowed the two reach the same redocking score, and the size effect works in 3DSynthFlow's favour even there: its molecules average 28.0728.07 heavy atoms against our 24.4024.40, and matched within pockets SynTrace is ahead in 5858 of 100100 by 0.0520.052 kcal/mol [0.001,0.103][0.001, 0.103] (Appendix B.2, Table 5). The comparison separates once the pose is taken as the model emits it: 67.19%67.19\% against 19.00%19.00\% on receptor-aware PoseBusters, and −6.18-6.18 against −3.26-3.26 kcal/mol on the direct score. It separates again on usable output: 99.41%99.41\% of our molecules are distinct against 17.73%17.73\%, and AiZynthFinder solves 45.49%45.49\% against 19.80%19.80\% under Enamine stock.

AuthorsOpenReview ↗

Response to Reviewer Jf45 (2/2)

Q1. A concrete case where local pocket geometry favours different building blocks

Figure 5 shows one intermediate and one reaction slot evaluated in COAA and in PTGIS. The two receptors are chemically very different. COAA is a pantothenate kinase: it phosphorylates pantothenate in the first committed step of coenzyme A biosynthesis, and its active site binds that polar substrate together with ATP. PTGIS is prostacyclin synthase, a cytochrome P450 that isomerises prostaglandin H2\mathrm{H}_2 to prostacyclin through a heme-lined channel shaped for a lipophilic C20\mathrm{C}_{20} substrate.

Both candidates are 1010-heavy-atom hydrazides sharing the same linker and differing only in the extension: a nitrogen-containing piperidine ring versus a phenyl ring (Figure 5a). The model prefers the basic piperidine in COAA and the apolar phenyl in PTGIS, following the polarity contrast between the two sites. Within each pocket we fix the core shared by the two products at the position supplied by that pocket's docked reference product, then score the completed products (Figure 5b):

Completed productCOAA ↓\downarrow (kcal/mol)PTGIS ↓\downarrow (kcal/mol)
Piperidinyl extension−10.45\mathbf{-10.45}−5.30-5.30
Phenyl extension−4.26-4.26−7.85\mathbf{-7.85}

The model's choice is in bold in each column, and the ordering reverses between them: the phenyl product scores 2.552.55 kcal/mol lower in PTGIS, the piperidine product 6.196.19 kcal/mol lower in COAA, with the contacts behind each shown in Figure 5c. For this fixed intermediate, then, which block is better is decided entirely by the local environment, and the policy picks the better one in each of the two pockets.

Q2. Evidence connecting that preference to spatial compatibility of the completed product

The pair in Q1 is one of many. Some intermediates occur in more than one test protein, and we collected every pair where the same intermediate and the same reaction slot appear in two different pockets, A and B — 964964 pairs in all (selection rule in Appendix C.4). In 86.7%86.7\% of them the model picks one block for A and a different one for B; for NoRel this happens in only 34.1%34.1\%. For every pair we repeated what Q1 shows for one: build all four products, both blocks in both pockets, and score each with that pocket's own reference core held in place. If the model is responding to the pocket, the block it picked for A should beat the other one in A, and the same should hold in B. Adding the two differences, that is what we see in 67.1%67.1\% of the 831831 pairs that returned all four scores, by a key-weighted mean of −1.60-1.60 kcal/mol (95% CI [−2.66,−0.57][-2.66, -0.57]).

Those pairs take their poses from docked reference complexes, so we ran the comparison again on intermediates the full model generated itself, this time against NoRel. At each intermediate both models choose from the same legal blocks at the same reaction slot, and we complete the products of the four each one prefers. To score them we dock the intermediate separately and hold its pose fixed, so only the newly added extension can move (Figure 3a–c; protocol in Appendix C.6). In the 2pc8 case of Figure 3b the four products chosen with the geometry average −10.97-10.97 kcal/mol against −10.44-10.44 without it. Across 9393 equally weighted pocket families the mean Full-minus-NoRel difference is −0.84-0.84 kcal/mol (95% CI [−1.41,−0.27][-1.41, -0.27]), with the full model scoring lower in 7171 of them (Figure 3d).

Taken together, these connect the preference in Q1 to the completed product: the blocks the policy prefers are the ones whose finished molecules fit their pocket better, and that advantage disappears once the policy can no longer see the geometry.

Concluding remarks

We thank the reviewer again for the close reading of the method and for the constructive questions. The revision prices each half of the coupling with ablations trained on independent seeds, puts SynTrace and a retrained 3DSynthFlow under one protocol, and adds two experiments that score the building blocks the policy selects, including the pocket-specific cases the reviewer asked to see. We hope these additions address the concerns raised, and we would be grateful if the reviewer could reconsider the score in light of them. We are of course happy to provide any further details or analyses that would help.

Thanks for the new experiments. The Separate/NoRel ablations with multiple seeds are solid, and the COAA vs PTGIS case does show the policy picking different blocks based on pocket polarity.

My concern is more about how much this matters in practice. The block preference differences average under 0.5 kcal/mol, which is small given Vina’s own noise. And once you redock, the gap between the full model and the ablations mostly disappears. Same story with 3DSynthFlow. Since anyone using these molecules will redock them anyway, the coupling’s main benefit gets erased right where it counts.

Also, solve rates for Separate and NoRel are missing, so I can’t tell if coupling helps synthesizability or just pose quality.

The new experiments improve the paper but the practical gain over the CGFlow framework still looks thin to me. I’ll keep my score.

Thank you for reading the revision and for the detailed response. We address each point below.

1. Practical value extends beyond redocking scores.

The reviewer states that "anyone using these molecules will redock them anyway." However, direct-pose quality is also important in practice. There are at least two established workflows where it has independent value beyond redocking:

  1. Rapid filtering. A common pipeline generates a large candidate pool, then filters by direct pose score or geometric checks before sending survivors to expensive computation (redocking, MD, FEP). If the generated pose is poor, promising molecules are discarded before ever reaching docking.

  2. Pose-dependent property prediction. Downstream ML models for binding free energy, selectivity, and ADMET increasingly take the 3D complex as input. A geometrically invalid pose degrades those predictions regardless of any Dock score, because these models consume the generated pose, not a redocked one.

In both settings, the model that generates the more accurate pose delivers a practical advantage that no amount of downstream redocking can recover. Redocking therefore does not erase pose-quality differences between models, and direct-pose metrics carry practical significance.

2. SynTrace and 3DSynthFlow: same paradigm, different architecture, SynTrace leads.

The reviewer notes similarities with CGFlow. The two share a generation paradigm, but the architectures differ and so do the results. We retrained 3DSynthFlow on our building-block library (274k Enamine, 71 templates) and scored both systems under our protocol on the same 100 test pockets (Table 1):

Score ↓\downarrowMin ↓\downarrowDock ↓\downarrowPB ligand ↑\uparrowPB complex ↑\uparrowUniq. ↑\uparrowAiZynth ↑\uparrow
SynTrace−6.18\mathbf{-6.18}−7.26\mathbf{-7.26}−8.38\mathbf{-8.38}87.80\mathbf{87.80}67.19\mathbf{67.19}99.41\mathbf{99.41}45.49\mathbf{45.49}
3DSynthFlow (retrained)−3.26-3.26−5.60-5.60−8.36-8.3639.5239.5219.0019.0017.7317.7319.8019.80

The two reach nearly the same Dock (−8.38-8.38 vs −8.36-8.36), yet SynTrace leads by 2.922.92 kcal/mol on Score, 4848 percentage points on PB complex, and 8282 percentage points on uniqueness. Redocking re-searches the pose and can mask large coordinate differences — as the pocket-conditioning ablation shows: Dock −8.17-8.17 while 98%98\% of outputs have positive direct scores (Table 2).

The two systems also differ substantially in uniqueness (99.41%99.41\% vs 17.73%17.73\%). Both draw from the same 274k building blocks, so the gap is not about vocabulary; SynTrace selects a different building block for different pockets 86.7%86.7\% of the time (Section 4.5), whereas 3DSynthFlow converges on repeated routes across pockets. The solve rates also differ (45.49%45.49\% vs 19.80%19.80\%). The gap between SynTrace and 3DSynthFlow spans every metric except Dock — the one metric that re-searches the pose and masks the difference.

3. The ablation gap is substantial on Score, Min, and PB complex.

The reviewer cites a gap "averaging under 0.50.5 kcal/mol." That figure comes from the Dock column, where Full–NoRel is 0.100.10 kcal/mol. This is expected: Dock applies a full pose search that re-optimises atomic positions, erasing exactly the differences in how each model placed the atoms. On the metrics that evaluate the pose as the model emits it, the gaps are larger:

Score ↓\downarrowMin ↓\downarrowPB complex ↑\uparrow
Full−6.32\mathbf{-6.32}−7.35\mathbf{-7.35}67.74\mathbf{67.74}
NoRel−5.77-5.77−6.84-6.8464.6864.68
Separate−5.67-5.67−6.90-6.9058.2058.20

Full–NoRel: 0.550.55 on Score, 0.510.51 on Min, 3.13.1 percentage points on PB complex. Full–Separate: 0.650.65, 0.450.45, 9.59.5 percentage points. All exceed the run-to-run standard deviation by several times (Appendix C.5). PoseBusters complex checks whether the generated pose is geometrically valid against the receptor — atom clashes, bond-angle distortions — and is not revisited by redocking. A model whose poses pass PB complex more often is placing atoms more accurately in 3D, and that advantage persists regardless of any subsequent Dock score.

4. Solve rates for the ablations.

AiZynth Enamine ↑\uparrow (%)AiZynth ZINC ↑\uparrow (%)
Full45.4745.4742.5442.54
NoRel44.2344.2342.1942.19
Separate45.4145.4142.3042.30

The three are close, which is expected: whether a synthesis route can be found depends on which building blocks and templates the library contains, not on which block the policy prefers at a given step. The close solve rates confirm that geometric coupling acts on pose quality and pocket-specificity rather than on route feasibility — exactly the claim the ablation is designed to test.

We hope these clarifications address the remaining concerns and we remain available for any further questions.

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

This paper proposes SynTrace, a pocket-conditioned molecular generation framework that jointly models synthesis actions and 3D molecular coordinates. By aligning reaction routes with docked poses, the method enables supervised training of a shared geometric representation without reward-based policy optimization. Experiments on CrossDocked2020 demonstrate competitive docking performance, improved retrosynthesis solve rates, and high molecular diversity.

Critical strengths
  1. The paper addresses an important problem of integrating synthetic accessibility with structure-based molecular generation. The proposed route–pose alignment provides a reasonable and practical learning framework.

  2. The experimental evaluation covers multiple aspects, including docking scores, molecular properties, retrosynthetic accessibility, and geometric ablations. The reported 45.49% retrosynthesis solve rate and 0.889 diversity are encouraging.

  3. The geometric intervention experiments provide useful evidence that synthesis decisions respond to relative ligand–pocket geometry.

Critical weaknesses
  1. Insufficient evidence for the practical benefit of geometric coupling. While the intervention experiments demonstrate that synthesis decisions are sensitive to pocket geometry, they do not establish that such sensitivity leads to better pocket-specific chemical choices. In Appendix C.1, the improvement in candidate-ranking correlation under the original pocket condition is not statistically significant.

  2. The comparison with existing methods is not fully convincing. Although SynTrace achieves a higher retrosynthesis solve rate, it does not consistently outperform existing methods in docking performance. For example, RxnFlow achieves a better median redocking score. Additionally, the results for 3DSynthFlow are taken from the original paper under a different evaluation protocol, making direct comparisons difficult.

Questions for authors

See the weakness

2 comments
AuthorsOpenReview ↗

Response to Reviewer VxVV (1/2)

We thank the reviewer for the careful reading and constructive feedback. Both weaknesses the reviewer raised ask for stronger evidence, and they led us to add two experiments that score the completed products the policy selects, an exact within-pocket size-matched comparison against the synthesis-aware baselines, and a 3DSynthFlow retrained on our own building-block library and evaluated under our protocol. Changes in the revised PDF are marked in magenta, and the new analyses are in Appendices B.2, C.4 and C.6.

W1. Evidence that geometric coupling yields better pocket-specific choices

We agree that the interventions by themselves do not settle this, so we added two experiments that score the completed products of the blocks the policy selects. The first asks whether the choice follows the pocket, using intermediates that occur in more than one test protein; the second compares the full model with a control on intermediates the model generates itself.

We collected every pair in which the same intermediate and the same reaction slot appear in two different pockets, A and B — 964964 pairs in all (Appendix C.4). In 86.7%86.7\% of them the model picks one block for A and a different one for B. We then built all four products, both blocks in both pockets, and scored each one with that pocket's own docked reference core held in place. If the model is responding to the pocket, the block it picked for A should beat the other one in A, and the same should hold in B. Adding the two differences, that is what we see in 67.1%67.1\% of the 831831 pairs that returned all four scores, by a key-weighted mean of −1.60-1.60 kcal/mol ([−2.66,−0.57][-2.66, -0.57]). Figure 5 shows one such pair in COAA and PTGIS, where the model swaps a basic piperidine for an apolar phenyl and the scores reverse between the two pockets.

Those pairs take their poses from docked reference complexes, so we ran the comparison again on intermediates the full model generated itself, this time against a control. The control, NoRel, is identical to the full model except that its action head cannot see where the ligand sits relative to the pocket; the parameter count is the same, so this access is the only difference (Section 4.5). At each intermediate both models choose from the same legal blocks at the same reaction slot, and we complete the products of the four each one prefers. To score them we dock the intermediate separately and hold its pose fixed, so only the newly added extension can move (Figure 3a–c; protocol in Appendix C.6). Across 9393 equally weighted pocket families the mean Full-minus-NoRel difference is −0.84-0.84 kcal/mol ([−1.41,−0.27][-1.41, -0.27]), with the full model scoring lower in 7171 of them (Figure 3d).

Together these replace the candidate-ranking comparison of Appendix C.1, which the revision reports unchanged: there the preference–score correlation was 0.1880.188 with the original pocket against 0.1710.171 with a replaced one, a difference we agree is not significant. That test asked whether seeing the pocket improves how the model orders a set of alternatives. Our claim is about the alternative it ends up choosing, and that is what both experiments score.

AuthorsOpenReview ↗

Response to Reviewer VxVV (2/2)

W2. Docking performance: RxnFlow

We agree on the median the reviewer quotes from the submitted Table 1: over all generated products RxnFlow reaches −8.89-8.89 against our −8.45-8.45 kcal/mol. Its median molecule carries 3131 heavy atoms against our 2424, so the two medians are not taken over comparable molecules. Redocking scores improve with molecular size — a point MolCRAFT's own evaluation makes about Vina Dock (Qu et al., 2024) — and in this evaluation the effect is monotonic for all three methods built on our library (Appendix B.2, Table 4):

Heavy atomsSynTraceRxnFlow3DSynthFlow
[18,22)[18,22)−8.006\mathbf{-8.006}−7.686-7.686−7.915-7.915
[22,26)[22,26)−8.600\mathbf{-8.600}−7.854-7.854−8.339-8.339
[26,30)[26,30)−8.838\mathbf{-8.838}−8.586-8.586−8.542-8.542
[30,34)[30,34)−9.076-9.076−8.923-8.923−9.368\mathbf{-9.368}
[34,∞)[34,\infty)−9.383-9.383−9.084-9.084−9.384\mathbf{-9.384}
Median heavy atoms24.024.031.031.027.027.0

The values are median redocking scores in kcal/mol. SynTrace leads in the three windows where most of its molecules fall, and RxnFlow's overall median benefits from the two largest windows, where it places 3131 heavy atoms against our 2424.

We therefore matched the heavy-atom distributions exactly within each pocket, keeping k=min⁡(nours,nother)k=\min(n_{\text{ours}}, n_{\text{other}}) molecules per heavy-atom count and discarding the surplus in ascending index order, independently of score. Matching retains 3,5373{,}537 molecules per method (≈35%\approx 35\% of each sample) and covers all 100100 pockets. On that shared size range SynTrace has the lower median in 6767 of 100100 pockets, with a per-pocket paired advantage of 0.1860.186 kcal/mol (95% CI [0.114,0.270][0.114, 0.270]). The same holds against the other synthesis-aware baselines (Appendix B.2, Table 5):

BaselinePaired advantage to SynTrace (kcal/mol)95% CIPockets favouring SynTrace
RxnFlow+0.186+0.186[0.114,0.270][0.114, 0.270]67/10067/100
3DSynthFlow (retrained)+0.052+0.052[0.001,0.103][0.001, 0.103]58/10058/100
SynCoGen (uncond.)+0.819+0.819[0.743,0.898][0.743, 0.898]100/100100/100

All five heavy-atom windows agree in direction (Appendix B.2, Table 4), and replacing the index-based discard rule with random retention over ten seeds gives +0.165+0.165 to +0.213+0.213 against RxnFlow. Ligand efficiency gives the same picture, −0.343-0.343 against −0.265-0.265. As a supplementary point, Table 1 now reports Dock as a mean throughout, and on that aggregation SynTrace is the lowest in the table (−8.38-8.38 against −8.28-8.28).

W2. Baseline comparability: 3DSynthFlow

Table 1 now reports 3DSynthFlow as we measured it ourselves, in place of the published numbers. Its released checkpoint needs a building-block vocabulary that we could not reconstruct from public assets, so we trained the public implementation on the 274274k Enamine library that SynTrace and RxnFlow use, and evaluated its 100,000100{,}000-update checkpoint on the same 100100 test pockets with our own scoring pipeline. The published high-β\beta numbers are kept separately as reproduction context in Appendix B.2 (median Vina −9.61-9.61, 36.1%36.1\% solve rate).

Score ↓\downarrow (kcal/mol)Dock ↓\downarrow (kcal/mol)PB ligand ↑\uparrow (%)PB complex ↑\uparrow (%)Uniq. ↑\uparrow (%)AiZynth ↑\uparrow (%)
SynTrace−6.18\mathbf{-6.18}−8.38\mathbf{-8.38}87.80\mathbf{87.80}67.19\mathbf{67.19}99.41\mathbf{99.41}45.49\mathbf{45.49}
3DSynthFlow (retrained)−3.26-3.26−8.36-8.3639.5239.5219.0019.0017.7317.7319.8019.80

Redocking re-searches the pose, so the two land within 0.020.02 kcal/mol of each other there. Even that parity is on 3DSynthFlow's side of the size effect, since its molecules average 28.0728.07 heavy atoms against our 24.4024.40; matched within pockets, SynTrace is ahead in 5858 of 100100 (Appendix B.2, Table 5). The comparison separates once the pose is taken as the model emits it: 67.19%67.19\% against 19.00%19.00\% on receptor-aware PoseBusters, and −6.18-6.18 against −3.26-3.26 kcal/mol on the direct score. It separates again on usable output, where 99.41%99.41\% of our molecules are distinct against 17.73%17.73\%, and AiZynthFinder solves 45.49%45.49\% of them against 19.80%19.80\%.

Concluding remarks

We thank the reviewer again for the time and the careful reading. The revision answers both weaknesses with evidence at the level of the finished molecule: products scored where the policy chooses, size matching within every pocket against each synthesis-aware baseline, and a 3DSynthFlow retrained and evaluated under our own protocol. We hope these address the concerns raised, and we would be grateful if the reviewer could reconsider the score in light of them. We are of course happy to provide any further details or analyses that would help.


Qu, Y. et al. MolCRAFT: Structure-Based Drug Design in Continuous Parameter Space. ICML 2024, PMLR 235, 41749–41768.

Comments on the paper

2 comments
AuthorsOpenReview ↗

General Response: Summary of Revisions

We thank all reviewers for their careful and constructive feedback. We are glad the reviewers found the route–pose supervision “a useful design”, one that “allows one product pose to provide supervision across several construction stages” (Jf45). We are also glad that the route–pose alignment was seen as “a reasonable and practical learning framework”, with intervention experiments that “provide useful evidence that synthesis decisions respond to relative ligand–pocket geometry” (VxVV), and that the paper “integrates synthesis planning and pocket-conditioned 3D molecular generation into a unified framework” (W73u). In the revised PDF, changes are marked in magenta and new analyses are collected in Appendices A.3, B.2, C.4, C.5 and C.6. Below we summarize the main points raised and how the revision addresses them; details are in the individual replies.

1. Whether geometric coupling improves pocket-specific chemical choices (VxVV, Jf45, W73u)

  • Across 964964 test pairs where the same intermediate and reaction slot occur in two different proteins, the model picks a different block for the two pockets in 86.7%86.7\% of them. Scoring all four products in both pockets with each pocket's reference core fixed, the blocks chosen for a pocket score better there in 67.1%67.1\% of the 831831 scored pairs, by a key-weighted mean of −1.60-1.60 kcal/mol [−2.66,−0.57][-2.66, -0.57] (Appendix C.4).
  • On states the model generates itself, against a parameter-matched control whose action head cannot read the relative geometry, the mean Full-minus-control difference is −0.84-0.84 kcal/mol [−1.41,−0.27][-1.41, -0.27] across 9393 equally weighted pocket families, with the full model scoring lower in 7171 of them (Section 4.5, Figure 3; Appendix C.6).
  • Figure 5 illustrates one pair: a basic piperidine is chosen for COAA, a kinase with a polar site, and an apolar phenyl for PTGIS, a P450 with a lipophilic channel, and the scores reverse between the two.

2. What the joint training contributes (Jf45, W73u)

  • Separate, which trains the action and coordinate predictions as two networks sharing nothing, costs 0.650.65 kcal/mol of direct score and 9.59.5 points of receptor-aware PoseBusters at twice the parameters. NoRel, which keeps one network but hides the relative geometry from the action head at an unchanged parameter count, still costs 0.550.55 kcal/mol and 3.13.1 points (Section 4.3, Table 2).
  • Both gaps appear on the pose the model emits. After redocking all three variants lie within 0.170.17 kcal/mol, since redocking re-searches the pose and erases the differences in how each model placed the atoms.

3. Baseline comparability (VxVV, W73u)

  • Table 1 now reports the synthesis-aware baselines as we measured them, with no published numbers carried over: 3DSynthFlow is retrained on our shared 274274k Enamine library and rescored under our protocol, and both released SynCoGen checkpoints are added (Appendix B.2).
  • Receptor-aware PoseBusters (PB complex) is now reported for every method that emits receptor-frame coordinates. SynTrace passes at 67.19%67.19\%: behind MolCRAFT (84.52%84.52\%) and DecompDiff (76.62%76.62\%), and first among the methods that also produce a synthesis route, the next being retrained 3DSynthFlow at 19.00%19.00\% (Table 1).
  • Redocking scores improve with molecular size in this evaluation (Appendix B.2, Table 4), so we match heavy-atom distributions exactly within each pocket: SynTrace then leads in 67/10067/100 pockets against RxnFlow, 58/10058/100 against retrained 3DSynthFlow and 100/100100/100 against unconditional SynCoGen (Appendix B.2, Table 5).

4. A second benchmark and independent seeds (W73u)

  • On the 1515 LIT-PCBA targets, under common receptors, boxes and scoring code, SynTrace has better ligand efficiency on every target than all three generative baselines, and is the only method exceeding the benchmark's known actives (Appendix C.5, Tables 8 and 9).
  • Full, NoRel and Separate were each trained three times. On every metric the separation between the full model and either ablation is several times the run-to-run spread (Table 2; Appendix C.6, Table 10).

5. Presentation

The opening of the Introduction, the captions of Figures 1 and 4, and the DYRK2 example in Section 4.6 were condensed. These edits change no claim and no number, so they are not marked in magenta.

AuthorsOpenReview ↗

Summary of Key Revisions

Dear reviewers,

Thank you again for the reviews and for the constructive suggestions, which have been very helpful to our paper. We would very much like to address any remaining concerns while the discussion is still open.

Whether pocket geometry improves the building-block choice. The revision scores the completed products built from the blocks the policy selects. When the same intermediate appears in two different pockets, the model picks a different block for each in 86.7%86.7\% of 964964 test pairs (against 34.1%34.1\% for the ablation that cannot read the relative geometry), and the block it picks for a pocket scores better there in 67.1%67.1\% of 831831 scored pairs, by −1.60-1.60 kcal/mol on average. Comparing the full model with that ablation head to head, the full model's choices score −0.84-0.84 kcal/mol better across 9393 pocket families (Section 4.5; Appendices C.4 and C.6). These results show that the geometric coupling does lead the policy to pick building blocks whose completed products fit their target pocket better.

Comparability of the baselines. We retrained 3DSynthFlow on our own building-block library and added both released SynCoGen checkpoints, so every synthesis-aware baseline is now measured under one protocol. We also added receptor-aware PoseBusters for every method that emits receptor-frame coordinates, and matched the heavy-atom distributions exactly within each pocket before comparing redocking scores (Table 1; Appendix B.2). Under these controlled conditions, SynTrace leads on direct-pose quality among all methods that also produce a synthesis route, and leads on redocking after size matching in the majority of pockets.

Scope of the evaluation. We added a second benchmark on the 1515 LIT-PCBA targets and trained the full model and both ablations three times with independent seeds (Appendices C.5 and C.6). On LIT-PCBA, SynTrace has better ligand efficiency than every generative baseline on all 1515 targets, and the ablation gaps remain consistent across all three runs.

We have carefully responded to each reviewer's comments individually. We sincerely appreciate the time and effort each reviewer has devoted to our work, and we look forward to any further feedback you may have.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.