acceptodds
Under review as a conference paper at ICLR 2027

RetroTAC: Predicting Route-Derived Synthetic Accessibility for PROTACs

Abstract

Proteolysis-targeting chimeras (PROTACs) enable targeted protein degradation, expanding therapeutic opportunities beyond conventional inhibition. However, assembling their large, multifunctional structures creates synthetic challenges that hinder their discovery. Traditional synthetic-accessibility scores can overestimate this difficulty by penalizing PROTACs' molecular complexity while overlooking modular assembly from accessible building blocks. In contrast, retrosynthesis planning tools provide explicit routes but are computationally expensive, particularly for large molecules. Here, we introduce a human-interpretable, PROTAC-oriented route-derived score that rewards short synthetic depth and balanced joining of molecular fragments, and RetroTAC, a machine-learning surrogate that predicts this score directly from molecular structures. RetroTAC combines complementary models in an ensemble, achieving a root mean squared error of 0.13 and an of 0.65 on held-out molecules. We further evaluate its utility on PROTACs enumerated by recombining warheads, linkers, and E3-ligase ligands, followed by retrosynthetic planning and route analysis. Candidates selected for high predicted scores achieve a mean route-derived synthesizability of 0.73, compared with 0.31 for low-score candidates; 98.75% versus 51.00% are fully resolved under the tested search settings. These results support RetroTAC as a fast prioritization tool for virtual PROTAC libraries, connecting molecular screening with interpretable route characteristics while reserving explicit planning for selected candidates.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Machine Learning, Drug Discovery, AI for Science, Molecular Property Prediction, Retrosynthesis

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_3dZZOpenReview ↗
Rating
3
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

RetroTAC predicts a route-derived synthetic-accessibility score directly from PROTAC structures to prioritize large virtual libraries before running explicit retrosynthesis. The authors obtain labels for 18,076 PROTACs using ShallowTree and eMolecules building blocks, combining normalized route depth with the average heavy-atom balance of reactants at bimolecular steps. Purchasable targets and unsuccessful searches receive fixed endpoint scores. They compare XGBoost, an MLP, and a pretrained Chemprop graph network using warhead-scaffold-grouped cross-validation, then select a 27-member ensemble that achieves a reported held-out RMSE of 0.132 and R2=0.647R^2=0.647. The ensemble also scores approximately 1.30 million enumerated PROTACs, from which 1,600 candidates undergo additional planning after selection by predicted score and model disagreement. High-scoring candidates achieve mean route scores of 0.733 versus 0.313 for low-scoring candidates, with planner-resolution rates of 98.75% versus 51.00%. I see the main contribution in the PROTAC route resource and the practical screening workflow, with further validation needed to connect the score to experimental synthetic effort.

Critical strengths
  • From my experience, PROTAC libraries can grow quickly through recombination of warheads, linkers, and E3 ligands, making explicit planning for every candidate quite expensive. A structure-based surrogate that identifies promising candidates for route inspection addresses a concrete bottleneck in computational degrader design.

  • The additional planning experiment provides meaningful evidence for prioritization. Across 800 high-scoring candidates, the planner resolves 98.75%, compared with 51.00% among 800 low-scoring candidates. The high-scoring groups also yield shorter routes with more balanced fragment joining, showing that the predictor transfers its intended ranking to newly enumerated molecules.

  • The route dataset could support work beyond this particular ensemble. I also appreciate the component-overlap analysis and warhead-scaffold grouping, which address an important source of overly optimistic validation for modular molecules. These choices make the resource and evaluation more useful than a random molecular split would.

Critical weaknesses
  • The methodological contribution is modest, which places more weight on demonstrating practical improvement over existing tools. The regressors and ensemble selection are established methods, and RSPred (Parrot et al., 2023) already learns a continuous score from retrosynthesis outputs. Appendix E includes DeepPSA, but measures agreement between scores rather than comparative screening performance. I would be comfortable with a straightforward model if the authors showed that it selects more useful candidates than existing scores under the same planning budget.

  • Fragment-size balance does not account for coupling difficulty, protecting groups, selectivity, or building-block cost. Appendix C provides useful sensitivity analyses, but agreement among score variants does not establish chemical validity. The labels also inherit the first route returned by one planner, including its stated difficulty with macrocycles. A focused comparison against documented syntheses or blinded chemist assessments would make the practical interpretation much more convincing for me.

  • There were several implementation details I could not locate. The paper reports 2,232 held-out molecules, then 463 for ensemble selection and 1,854 for evaluation, which sum to 2,317. Appendix E instead reports 378 selection molecules. Figure 3b also shows RMSE 0.128 and R2=0.67R^2=0.67, whereas Table 2 reports 0.132 and 0.647. These may be reporting errors, but they need correction to establish which results come from untouched evaluation data. The uncertainty estimates also need calibration before supporting confidence-based filtering.

Questions for authors
  1. Could you compare the top-ranked candidates from RetroTAC, DeepPSA, PROTAC-adapted FSscore, RAscore, and random selection under an identical planning budget? Please report resolution rate, route depth, and retained chemical diversity. The current high-versus-low comparison demonstrates separation, but does not establish the improvement over available screening methods.

  2. Is the ordering induced by Eq. 1 consistent with chemists’ preferences? A one-step route with no bimolecular reactions receives y=0.75y=0.75 because the fallback fragment balance is B=0.5B=0.5, whereas a two-step route with perfect balance, B=1B=1, receives y=0.95y=0.95. Please examine such cases using documented syntheses or blinded expert rankings, and compare against a depth-only score.

  3. How sensitive are the labels to the first route returned, search budget, and building-block stock? Please repeat planning on a representative subset across seeds and budgets, including macrocycles and zero-labeled molecules, and report changes in resolution and ranking, and clarify how the 22 partial routes in the prospective low-score group contribute to the reported scores.

  4. Was the ensemble-selection subset 463 or 378 molecules? Please confirm that the final evaluation subset influenced neither ensemble weights nor strategy selection.

  5. How does performance vary when the warhead, linker, or E3 ligand is unseen? The warhead-grouped cross-validation is useful, but the final held-out and enumerated sets need corresponding component-overlap statistics. Please also confirm whether prospective deduplication excluded both development and held-out molecules.

  6. Can ensemble disagreement improve selection after controlling for predicted score? Compare errors within narrow score bins and report error as increasingly uncertain predictions are excluded. The current interval spanning two ensemble standard deviations covers only 72.9% of held-out cases.

  7. What are the wall time and hardware requirements for scoring the 1.30 million candidates? Please compare the 27-member ensemble with the single GNN in throughput and top-ranked candidate quality. The RMSE improvement from 0.1378 to 0.1320 is modest, so the additional inference cost does matter for the proposed use case.

  8. How reliable are low predictions across the full library? The selected low-score candidates have predictions below -0.058, despite labels lying in [0,1][0,1], yet 51% still have fully resolved routes. A randomly sampled planning panel would help measure calibration and the fraction of useful candidates discarded at different filtering thresholds. This would clarify how practitioners should use the score.

4 comments

We thank the reviewer for the interesting questions and insights to improve the manuscript. We will respond with several comments.

We agree that practical screening utility and chemical validity require separate evidence. These analyses assess prioritization under a fixed planner and stock; they do not establish experimental synthesis success.

Q1. Comparison under identical planning budgets. Learning a continuous retrosynthesis-derived score has precedent in RSPred. Its underlying Spaya score is proprietary, although experimental code is open source. Our contribution is a PROTAC-focused target, surrogate and evaluation.

We sampled 10,000 enumerated PROTACs uniformly without replacement, selected the top 100 from RetroTAC, DeepPSA, RAscore and generic FSscore, and sampled 100 random controls. Their union has 479 molecules. Each received three ShallowTree runs at 30 minutes (shared candidates reuse planning results). FSscore here is the generic model, not SCScore or the PROTAC-adapted model in the FSscore paper. RetroTAC uses the 27-member RetroTAC ensemble.

SelectionResolved, mean %Mean resolved depthDiversityUnique scaffolds
RetroTAC97.32.830.720100
DeepPSA79.33.580.80099
RAscore86.33.110.77698
Generic FSscore81.73.620.75497
Random61.74.110.813100

Resolution includes all 100 candidates per method per seed, and depth averages resolved molecule-seed observations. Diversity is 1−T‾1-\overline{T}, where TT is pairwise Morgan-fingerprint Tanimoto similarity (radius 2, 2,048 bits, chirality), excluding self-pairs. Scaffolds are non-generic whole-molecule Bemis-Murcko scaffolds: atom and bond types, and specified scaffold stereochemistry, are retained.

We can see that RetroTAC improves mean resolution by 35.7 percentage points over random and 11.0 over RAscore, with shallower routes. However, its fingerprint diversity is lower than every method, despite retaining 100 distinct scaffolds. Its picks contain 58 warheads, 81 linkers and 28 E3 ligands, versus 91, 95 and 66 for random selection. Higher planning yield comes with reduced component diversity. To estimate uncertainty, we resample molecules with replacement 2,000 times, keeping each molecule's three seed outcomes together. The 2.5th and 97.5th percentiles give 95% bootstrap intervals: 94-100% for RetroTAC and 52.3-70.7% for random.

Q2. Chemical interpretation. The example is correct: y=(L+B)/2y=(L+B)/2, L=max⁡(0,1−(D−1)/10)L=\max(0,1-(D-1)/10), and fallback B=0.5B=0.5 give 0.75 for the one-step case and 0.95 for the balanced two-step case. Balance does not encode coupling difficulty, protecting groups, selectivity or cost. Against depth alone, Appendix C reports ρ=0.723\rho=0.723 and top-decile Jaccard overlap 0.289 over 16,318 routes.

The exact example occurs infrequently: the development set contains one one-step fallback route and four perfectly balanced two-step routes, while neither case occurs in the held-out set. Nevertheless, this measures changed ranking, and the score can favor longer, more balanced routes, so rarity of the fallback does not establish a broader chemical validity.

Due to the limited time frame, we were unable to perform systematic experts or documented-synthesis validation.

Q3. Seeds, budgets and stock. We sampled 25 molecules per stratum: macrocycles (ring size ≥12\geq12), and low, medium and high predicted scores (tertiles) among non-macrocycles. Each molecule received three ShallowTree runs with different seeds at both 10- and 30-minute budgets:

Stratum30 min: mean score ± mean SD10 min: mean score ± mean SD
Macrocycle0.333 ± 0.0070.300 ± 0.019
Low0.300 ± 0.0060.224 ± 0.038
Medium0.495 ± 0.0110.415 ± 0.035
High0.575 ± 0.0190.562 ± 0.030

These are route scores, not predictions. SD means the mean within-molecule SD across seeds: σˉb=100−1∑iSD(yi,b,1,yi,b,2,yi,b,3)\bar\sigma_b=100^{-1}\sum_i\mathrm{SD}(y_{i,b,1},y_{i,b,2},y_{i,b,3}). It is 0.0107 at 30 minutes and 0.0306 at 10 minutes. At 30 minutes, 70 molecules resolve in all three runs and 30 in none, while pairwise run-ranking correlations are 0.9795-0.9890. At 10 minutes, 58 resolve in every run, 36 in none and six inconsistently.

Averaging across seeds, resolution rises from 61% to 70%: +9.0 percentage points (95% bootstrap interval 4.3-14.7). Mean route score rises by 0.0505 (0.0252-0.0794). Here we resample molecules within strata 5,000 times, preserving both budgets and all seeds per molecule. The correlation between mean scores at the two budgets is ρ=0.916\rho=0.916. Equal stratum sampling means these pooled rates do not estimate library-wide prevalence.

The 22 partial routes receive the same Eq. 1 score, y=(L+B)/2y=(L+B)/2, from their returned reaction trees, even though some terminal building blocks are not in stock. They contribute to the uncertain-low mean of 0.330 but are not counted as resolved. We will replace the claim that all searches failing to reach stock receive zero with: “Unresolved searches with no returned reaction steps receive y=0y=0; nonempty partial routes retain their depth-and-balance score.”

Additional notes:

  • stock sensitivity remains untested;
  • repeated first-route searches assess repeatability, not optimality across alternative routes;
  • macrocycles were excluded from development and held-out sets but remain in the enumerated library.

Q4. Counts and evaluation integrity. The corrected split is 2,232=446+1,7862,232=446+1,786: selection and evaluation. The ensemble has 32 members, RMSE 0.12775 and R2=0.67314R^2=0.67314. The older 27-member ensemble has RMSE 0.13202. We will correct the stale counts, composition and metrics.

Caruana weights use only the 446 selection cases. However, the original script selects “best single” and “best backend” using evaluation RMSE and compares strategies there. Strategy choice was therefore not independent. Reselecting baselines on selection cases alone gives evaluation RMSE 0.13355 for a single GNN and 0.14008 for the MLP average. This reanalysis does not provide a fresh untouched test set. We will update Table 2.

The DeepPSA comparison has 2,232−73−432=1,7272,232-73-432=1,727 cases: 14 of the 446 selection cases were already among the 73 training overlaps. Pearson/Spearman/thresholded Kendall correlations are 0.2577/0.3892/0.2555 for route scores and 0.3313/0.4944/0.2478 for RetroTAC. Kendall uses thresholds 0.70 and 0.50 for RetroTAC/route scores and DeepPSA, respectively, and the other correlations use continuous outputs.

Q5. Unseen components. Component-grouped CV (mean ± SD over 25 folds):

GroupingModelRMSER2R^2
WarheadXGB0.177 ± 0.0180.237 ± 0.121
WarheadMLP0.171 ± 0.0160.290 ± 0.114
LinkerXGB0.160 ± 0.0060.384 ± 0.046
LinkerMLP0.150 ± 0.0070.453 ± 0.054
E3 ligandXGB0.174 ± 0.0200.249 ± 0.061
E3 ligandMLP0.160 ± 0.0190.362 ± 0.071

GNN models training is ongoing and its results remain pending. Compared with warhead and E3 ligand grouping, linker grouping performs higher in R2R^2 for both models, possibly suggesting a memorization of repeating PROTAC components to predict the score.

Canonical isomeric SMILES give these exact component overlaps (% of query molecules):

Query → referenceWarheadLinkerE3 ligand
Held-out → development79.5771.0176.79
Enumerated → development52.5167.3156.39
Enumerated → held-out8.6813.1923.07

Whole-molecule overlap is zero for held-out versus development and enumerated versus development. However, three of 1,303,154 enumerated molecules overlap the held-out set.

Q7. Cost and prediction quality. We scored the same 10,000 PROTACs, sampled uniformly without replacement from the enumerated library, three times per model. Jobs used A100 80-GB GPUs, eight EPYC 7742 CPU cores, batch size 256 and chunks of 2,000. Timings include featurization and inference but exclude loading/warm-up.

ModelTime (s), mean ± SDMolecules/sPredicted score, mean ± SDRMS difference vs 32-memberRandom-panel RMSE
Single GNN23.83 ± 0.08419.66 ± 1.340.565 ± 0.1940.1360.330
27-member ensemble384.00 ± 3.5126.04 ± 0.240.531 ± 0.1880.0580.294
32-member ensemble430.76 ± 2.8423.22 ± 0.150.544 ± 0.18200.305

SD is across runs for time and across molecules for scores. RMS difference measures disagreement with the 32-member predictions, not error against reference labels. The ensembles share 72 top-100 picks, versus 25 for GNN and the 32-member ensemble.

For quality, RMSE compares predictions on the random 100-molecule panel against mean route scores from three 30-minute searches. These estimates concern averaged route scores on 100 molecules, not all 1.3M. The ensembles cost 16-18 times more and have lower estimated error, although this small panel leaves their ranking uncertain. Library timing is still unmeasured.

Q8. Filtering low predictions. Negative regression outputs are not probabilities or proof of infeasibility. In the random 100-molecule panel, discarding RetroTAC predictions below 0.5 removes 33% of candidates and loses 15.1% of resolved observations. At 0.7, it removes 84% and loses 77.3% of resolved observations. Estimates average three seeds per molecule. We recommend prioritization rather than treating 0.7 as a safe rejection threshold. This small panel does not establish library-wide calibration.


We will include the additional experiments and results of this rebuttal response to the manuscript. We hope our answers address the reviewer's concerns and that they consider raising their score.

I sincerely thank the authors for doing a strong job addressing my concerns, and I understand that not all experiments could be performed in the limited time window. I just wanted to provide enough opportunities to improve the manuscript. I especially appreciate the Q1 response to see how RetroTAC performs against other methods, which despite lower diversity (not the most informative metric anyway) its overall performance is strong compared to othhers. I will maintain my score of 3 (simply because the paper would be better with additional experiments) but will advocate for its acceptance during the AC/Reviewer discussion.

Reviewer_fbnpOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

This paper presents a regression model to access synthesizability of PROTAC molecules whose structures are out-of-distribution of baseline scoring functions. The methods and evaluation are extremely trivial and naive. This method outperforms other baseline methods by a lot on new PROTAC evaluations using naive models (MLP, GNN). This paper is way below the acceptance criteria of major ML conferences.

Critical strengths

None

Critical weaknesses
  • Extremely limited novelty on both data and methods perspective. Comparing with baselines that are not even trained for PROTACs.
  • The training-eval split is problematic. The split should ensure that none similar warheads / e3 ligand are presented in the training set.
  • Paper is not well written. Methods are mostly implementation details and data configs.
Questions for authors
  • How would other models perform if they are finetuned on the same training dataset?
  • How would this method generalize to unseen warheads?
  • Because the oracle is basically a template-based retro-synthesize algorithm, will the model biased towards specific motifs?
2 comments

We thank the reviewer for the questions. Following are our responses.

To start, there are limited methods for assessing synthesizability of PROTACs with cheap surrogate models, hence the novelty of this work and why there is limited previous work we can compare to. For example, FSscore (from Neeser et al., 2024) includes an analysis of a finetuned model on 80 PROTACs, but such specific model is not open source. We do compare with DeepPSA (Zhang et al., 2025) and provide an analysis in Appendix E. We also compare our score with other SOTA retrosynthesis scores in Section 4.1, Figure 2a.

Q1. DeepPSA is a GNN model, so finetuning it would most likely lead to similar performance compared to a pretrained ChemProp model, also based on a GNN architecture. On the other hand, FSscore finetuned on PROTACs is not publicly available and requires experts to label synthesis routes. In the paper, the authors labeled 80 PROTACs, so selecting a diverse pool from our 20K might lead to compareable performance to RetroTAC.

Q2. Our method already generalizes to unseen warheads. We highlight it at lines 188-192 and present a thorough analysis in Appendix B, Table 3.

Q3. Certain chemicals building blocks might indeed appear more frequently, and so be more suitable for synthesis than others. This could lead, for example, to fewer building blocks required to build a PROTAC library.

Across the 15,844 training routes, 3.3% of the building-block vocabulary covers half of all block occurrences, and two thirds of routes contain one of the 50 most frequent blocks. Of the 50, 37 shift the score significantly (p<0.05p < 0.05, Mann–Whitney U, Bonferroni-corrected α=0.001\alpha = 0.001), by a median of 0.046. With ~15,800 routes, significance is reached at shifts as small as 0.006, and only 18 of the 37 exceed 0.05. Large purchasable PROTAC components raise it (e.g., the cereblon (CRBN) ligand brominated thalidomide), small generic reagents lower it (e.g., CO, N, CI, Br). The largest shift is 0.101, or 49% of the score's standard deviation (0.204).

As expected, a route terminating on a pre-made warhead or E3 ligand scores as more synthesizable than one assembled from small reagents. We do acknowledge that a surrogate like RetroTAC might indeed capture part of the signal by recognising known purchasable PROTAC components. This bias clearly depends on the chosen stock and template library for ShallowTree.

However, our cross-validation training is grouped by the warhead's Bemis-Murcko scaffold, so no warhead scaffold spans a train/validation boundary. This split is orthogonal to an E3 ligand scaffold split, and the recurring E3 ligands each appear across roughly 250 of the 3,128 warhead groups. As reported in Appendix B, lines 798-801, having both warheads and E3 ligands in disjoint groups precludes balanced folds.

Hence, we can assume the current warhead-scaffold-split performance as robust to warhead memorisation, but can not claim full robustness to E3-ligand memorisation. We will state this as an additional limitation of our work.

We hope this answer addresses the reviewer's concerns and that they consider raising their score.

Dear Reviewer,

We wanted to kindly follow up on our rebuttal, as the discussion period is coming to an end. We would appreciate hearing whether our responses have addressed your concerns, and we would be happy to provide any further clarifications.

Thank you again for your time and consideration.

Reviewer_U9SoOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

This paper introduces RetroTAC, an ML-based framework for predicting route-derived synthetic accessibility scores for PROTACs. The proposed route-derived score combines route depth and reactant-size balance from routes generated by ShallowTree, and is predicted from molecular structures using an ensemble of regression models. Experiments show that RetroTAC predicts the proposed scores on held-out molecules and selects candidates for which ShallowTree more frequently finds complete routes.

Critical strengths
  1. The paper recognizes the modular nature of PROTAC synthesis and highlights that whole-molecule complexity may overestimate synthetic difficulty, which provides a practically relevant motivation for the work.
Critical weaknesses
  1. The paper does not establish that its proposed label meaningfully measures synthetic accessibility. It scores the first complete route found by ShallowTree without examining whether alternative results would substantially change the score. It also provides no independent evidence that joining balanced reactants indicates easier synthesis or should receive equal weight with route depth. Predicting this label accurately therefore does not establish that the model predicts synthesis difficulty. Since the modeling approach largely uses standard methods, the lack of validation for the target undermines the paper’s central contribution.

  2. The prospective experiment does not establish that RetroTAC provides a meaningful screening benefit over random selection or existing scores. The reported planning success rates of 98.75% and 51.00% compare the highest- and lowest-scoring candidates, but without a random-selection baseline or comparisons with other methods, the improvement remains unclear. Approximately 92% of the original labeled dataset is already resolved or purchasable, making it particularly important to report the baseline success rate in the enumerated library. The authors also note that some failures arise from the planner’s inability to decompose macrocycles, but do not examine whether such structures dominate the low-scoring group. The observed separation may therefore reflect the model’s ability to identify limitations of ShallowTree rather than actual synthetic difficulty. The paper neither establishes the added value of the proposed score over directly predicting planning success nor demonstrates better prioritization among molecules for which complete routes can already be found.

Questions for authors
  1. How much does a molecule’s score vary across repeated searches or different complete routes? Is this variation large enough to change the ranking of candidate molecules?

  2. What evidence supports reactant-size balance as an indicator of lower synthetic difficulty? Compared with route depth alone, does including this term produce rankings that better agree with expert assessments or documented synthesis difficulty?

  3. What planning success rates would random selection and existing synthetic accessibility scores achieve in the enumerated library? Could a model trained directly to predict planning success provide comparable screening performance?

  4. Are failures in the low-scoring group concentrated in structures that ShallowTree is known to struggle with, such as macrocycles? After excluding or controlling for these structural features, does the planning success gap between the high- and low-scoring groups remain substantial?

4 comments

We thank the reviewer for their questions. We report our responses as follows.

Q1. To test how stable the route scores are, we split the 1.3M enumerated PROTACs into four strata by RetroTAC score: low, medium, top tertile, and macrocycles (ring size ≥12\geq 12). We then drew 25 PROTACs from each stratum at random, 100 in total. We then ran ShallowTree 3 times per molecule with different random seeds at 30min budget, as in the manuscript.

For each molecule, we computed the mean, SD, and max of its route scores, then averaged them within each stratum. We also calculated the Spearman correlation coefficient across the 3 run pairs (combinations). The results are reported in the table below:

StratumMean ± SDMaxρ\rho(run, run), mean ± SD
Macrocycle0.33 ± 0.0070.340.99 ± 0.005
Low0.30 ± 0.0060.310.99 ± 0.001
Medium0.49 ± 0.0110.510.92 ± 0.012
High0.58 ± 0.0190.590.93 ± 0.030
All0.43 ± 0.0110.440.98 ± 0.005

We can conclude that the rankings are stable across searches, i.e., two runs at different seeds agree with ρ≥0.92\rho \geq 0.92 in every stratum.

Q2. According to the synthetic chemist co-authors, a published route does not necessarily score qualitatively high, as it is often the result of what worked all the way through and yielded the end product, usually deviating from the original planned route. We provide a detailed analysis of the selection of score components in Appendix C, lines 861-888 and Table 5.

Q3. We address the question by carrying out the following experimental steps:

  1. We subsampled 10K data from the enumerated 1.3M PROTACs (uniformly without replacement).
  2. We then ran RetroTAC, DeepPSA, FSscore (non-finetuned), RAscore on those 10K samples.
  3. We isolated the top-100 scoring PROTACs for each method.
  4. We subsampled without replacement 100 of those 10K as a control.
  5. We ran ShallowTree N=3N = 3 times with different random seeds and used the original 30min time budget.

We can see in the table below that seed-to-seed agreement is very high. In fact, the gap between "all 3" and "≥ 1" (at least one) is 1 molecule for every scoring method and 3 for the control, i.e., ShallowTree is almost deterministic, and partial-resolution cases are rare. Additionally, every method beats the random control substantially, and RetroTAC separates from the next-best (RAscore) by 11 points on all-3 resolution and cuts the complete-failure rate from 13% to 2%.

MethodAll 3 seeds resolved≥ 1 seed resolvedNo seed resolved
RetroTAC97982
RAscore868713
FSscore (non-finetuned)818218
DeepPSA798020
Random (control)606337

Regarding predicting planning success, given the budget specified in the Methods Section, only 7.5% of the development set and 12.2% of the held-out set were not resolved by ShallowTree. Because of that, predicting synthesis success as a binary classification task would be extremely unbalanced to learn. Moreover, predicting a continuous score would be beneficial for RL generative pipelines.

Q4. ShallowTree failures are not concentrated in macrocycles. Within the low-scoring group, macrocycles (largest ring ≥12 atoms) resolve more often than non-macrocycles: 56.8% (179/315) versus 47.2% (229/485).

To investigate ShallowTree failure modes, we further screened 14 descriptors spanning classes CASP tools might struggle with, i.e., macrocyclicity, largest ring, ring count, aromatic and aliphatic rings, bridgehead atoms, stereocenters, quaternary carbons, gem-dimethyl groups, Fsp3, heavy atoms, MW, TPSA, rotatable bonds. Each individually showed weak discrimination between resolved and failed low-scoring molecules (AUCs 0.425–0.562).

To assess whether the planning-success gap persists after structural restrictions, we compared success rates after excluding macrocycles and after restricting molecules to the high group’s observed minimum-to-maximum ranges for all 14 descriptors simultaneously. The gap, defined as the high-group success rate minus the low-group success rate, remains substantial in both comparisons:

-\frac{\text{low successes}}{\text{low molecules}}\right)$$ |Comparison|High success rate|Low success rate|Gap| |---|---:|---:|---:| |All molecules|98.8% (790/800)|51.0% (408/800)|47.8 pp| |Macrocycles excluded|98.8% (790/800)|47.2% (229/485)|51.5 pp| |Within all 14 descriptor ranges|98.8% (790/800)|42.1% (8/19)|56.6 pp| Applying each descriptor restriction individually also leaves substantial gaps, ranging from 41.2 to 68.6 percentage points. However, the constrained low-scoring subset (remaining 19 elements) and the absence of shared Bemis-Murcko scaffolds limit our ability to rule out other structural confounding. We hope this answer addresses the reviewer's concerns and that they consider raising their score.

Dear Reviewer,

We wanted to kindly follow up on our rebuttal, as the discussion period is coming to an end. We would appreciate hearing whether our responses have addressed your concerns, and we would be happy to provide any further clarifications.

Thank you again for your time and consideration.

Thank you for the additional experiments. Please correct me if I am wrong, but my understanding is that the main contribution is a ShallowTree-based SA score intended to better assess synthetic accessibility for PROTACs than traditional SA scores. The new results support its usefulness for predicting ShallowTree outcomes, but I still do not see clear evidence that the proposed score provides a more meaningful assessment of synthetic accessibility for PROTACs in general. Since the predictive methods are largely standard, and the practical value of the proposed score remains my main concern about the contribution, I maintain my score of 1.

I really appreciate the author's effort in addressing my answers. However, this work should be considered more of a data-specific application paper, which might not be suitable for an ML conference presentation. Therefore, I will maintain my rating.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.