acceptodds
Under review as a conference paper at ICLR 2027

Pangaea: Agentic Search over a Unified Construction Space for Automated Molecular Three-Dimensional Descriptor Design

Abstract

Chemically informed molecular descriptors remain valuable for machine-learning based molecular property prediction, especially in data-limited regimes. Among these, three-dimensional (3D) descriptors are particularly appealing due to their ability to encode rich spatial chemical information. However, task-specific 3D descriptor design requires substantial expertise and remains difficult to automate. We abstract 3D descriptor construction as a three-stage process consisting of Formation, Extraction, and Compression, starting from chemically annotated molecular point-cloud sources. This abstraction enables new descriptors to be constructed by combining stage-wise operators and their parameters, thereby reformulating task-specific 3D descriptor design as a search for the combination best suited to each task. Building on this formulation, we propose *Pangaea*, which provides a Unified Construction Space that structures candidate descriptors as executable pipelines, together with an Incumbent Revision Agent that uses a large language model to iteratively refine the best-performing pipeline based on evaluation history, design-space semantics and chemical priors. QM9 experiments under data-limited conditions show *Pangaea* ranks first among descriptor-based methods, achieves a 24% relative reduction in test MAE, and outperforms four deep-learning baselines.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
molecular representation learning, 3D molecular descriptors, automated descriptor design, agentic search

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_8zZWOpenReview ↗
Rating
3(+2)
Confidence
3
Informativeness
1
Presentation
2
Soundness
2(+1)
Read review
Paper summary

This paper proposes Pangaea, a framework for automated task-specific 3D molecular descriptor design. It organizes descriptor construction into a Unified Construction Space and uses an LLM-based Incumbent Revision Agent to search for effective pipelines. The method is evaluated on a low-data QM9 setting and an Fe Mössbauer benchmark, where it reports improvements over hand-crafted descriptors, black-box search methods, and several deep-learning baselines. The overall idea is interesting, but I have some concerns about the experimental validity, especially potential target-information leakage in QM9 and insufficiently strong baselines.

Critical strengths
  1. The decomposition of 3D descriptor construction into Formation, Extraction, and Compression provides a useful structured search space. The LLM-based search over this space is also an interesting direction.
  2. The paper additionally includes comparisons with black-box optimizers and several ablations of the search strategy.
Critical weaknesses
  1. Potential target-information leakage in QM9. Pangaea uses electronic-structure-derived inputs, including DFT-based quantities, while many QM9 targets are themselves quantum-chemical properties. This raises the concern that some prediction targets may already be strongly encoded in the input features. Without carefully ruling this out, it is difficult to attribute the reported gains to the proposed descriptor-search method itself.

  2. The baseline comparison is not sufficiently strong. The paper mainly compares against SOAP, LMBTR, GRIND, and several neural models such as SchNet, ViSNet, MolSpectra, and MABNet. Stronger and more recent molecular representation or foundation-model baselines should be included under matched input-information conditions.

Questions for authors
  1. Can the authors analyze whether any QM9 prediction target can be directly or approximately recovered from the electronic-structure-derived input features? This is important to rule out information leakage.

  2. Can the authors include stronger and more recent molecular property prediction baselines under comparable input conditions?

5 comments

1. Potential target-information leakage (W1 & Q1)

We acknowledge the reviewer’s concern that electronic-structure-derived inputs may already encode QM9 target properties, making it difficult to attribute the reported improvements to the descriptor-search method itself.

To assess the contribution of descriptor search without\color{#8B0000}{\textsf{without}} electronic-structure\color{#8B0000}{\textsf{electronic-structure}} information\color{#8B0000}{\textsf{information}} derived\color{#8B0000}{\textsf{derived}} from\color{#8B0000}{\textsf{from}} DFT\color{#8B0000}{\textsf{DFT}} calculations\color{#8B0000}{\textsf{calculations}}, we conduct experiments under two settings: Pangaea (full-search, struct-only)\color{#8B0000}{\textsf{\textbf{Pangaea (full-search, struct-only)}}} searches\color{#8B0000}{\textsf{searches}} the\color{#8B0000}{\textsf{the}} full\color{#8B0000}{\textsf{full}} source\color{#8B0000}{\textsf{source}} library\color{#8B0000}{\textsf{library}} and\color{#8B0000}{\textsf{and}} then\color{#8B0000}{\textsf{then}} selects\color{#8B0000}{\textsf{selects}} the\color{#8B0000}{\textsf{the}} highest-scoring\color{#8B0000}{\textsf{highest-scoring}} pipeline\color{#8B0000}{\textsf{pipeline}} that\color{#8B0000}{\textsf{that}} uses\color{#8B0000}{\textsf{uses}} only\color{#8B0000}{\textsf{only}} structural\color{#8B0000}{\textsf{structural}} inputs\color{#8B0000}{\textsf{inputs}} and\color{#8B0000}{\textsf{and}} no\color{#8B0000}{\textsf{no}} DFT-derived\color{#8B0000}{\textsf{DFT-derived}} electronic-structure\color{#8B0000}{\textsf{electronic-structure}} information,\color{#8B0000}{\textsf{information,}} whereas\color{#8B0000}{\textsf{whereas}} Pangaea (struct-search)\color{#8B0000}{\textsf{\textbf{Pangaea (struct-search)}}} restricts\color{#8B0000}{\textsf{restricts}} the\color{#8B0000}{\textsf{the}} search\color{#8B0000}{\textsf{search}} to\color{#8B0000}{\textsf{to}} structural\color{#8B0000}{\textsf{structural}} sources,\color{#8B0000}{\textsf{sources,}} excluding\color{#8B0000}{\textsf{excluding}} DFT-derived\color{#8B0000}{\textsf{DFT-derived}} electronic-structure\color{#8B0000}{\textsf{electronic-structure}} information\color{#8B0000}{\textsf{information}} from\color{#8B0000}{\textsf{from}} the\color{#8B0000}{\textsf{the}} outset.\color{#8B0000}{\textsf{outset.}} The two tables below report predictive performance and mean ranks across the 12 tasks, respectively. (LMBTR K2 encodes two-body atomic distances, while K3 additionally captures three-body angular information. GRIND (vdW only) uses only the van der Waals interaction field, likewise excluding explicit electronic-structure information.)

\hline \text{Learner} & \text{Setting} & \mu & \alpha & \varepsilon_{\mathrm{HOMO}} & \varepsilon_{\mathrm{LUMO}} & \Delta\varepsilon & R^{2} & \text{ZPVE} & U_{0} & U & H & G & c_{v} \\\\ \hline \text{XGBoost} & \text{Best baseline} & \mathbf{0.500} & 0.983 & 8.88 & \mathbf{9.66} & 12.54 & 20.36 & 0.943 & \mathbf{1613} & \mathbf{1613} & \mathbf{1613} & \mathbf{1604} & 0.403 \\\\ & \textit{Pangaea}\text{ (full-search, struct-only)} & 0.675 & 0.954 & 9.12 & 9.80 & \mathbf{12.36} & 16.23 & 0.981 & 1673 & 1690 & 1670 & 1667 & 0.423 \\\\ & \textit{Pangaea}\text{ (struct-search)} & 0.679 & \mathbf{0.952} & \mathbf{8.83} & 9.99 & 12.53 & \mathbf{15.81} & \mathbf{0.882} & 1685 & 1668 & 1658 & 1649 & \mathbf{0.396} \\\\ \hline \text{Ridge} & \text{Best baseline} & \mathbf{0.534} & 0.679 & 9.05 & 11.85 & 15.37 & 15.01 & 0.409 & 381 & 381 & 381 & 381 & 0.227 \\\\ & \textit{Pangaea}\text{ (full-search, struct-only)} & 0.771 & 0.630 & \mathbf{8.73} & 12.28 & 15.43 & 8.43 & 0.423 & 269 & 307 & 270 & \mathbf{211} & 0.209 \\\\ & \textit{Pangaea}\text{ (struct-search)} & 0.740 & \mathbf{0.590} & 9.17 & \mathbf{11.81} & \mathbf{14.83} & \mathbf{7.71} & \mathbf{0.379} & \mathbf{211} & \mathbf{211} & \mathbf{211} & \mathbf{211} & \mathbf{0.198} \\\\ \hline \text{SVR} & \text{Best baseline} & \mathbf{0.542} & 0.496 & \mathbf{7.79} & 9.33 & 12.35 & 9.91 & 0.480 & 435 & 435 & 435 & 435 & 0.160 \\\\ & \textit{Pangaea}\text{ (full-search, struct-only)} & 0.791 & 0.477 & 8.13 & 9.22 & 12.29 & 6.21 & 0.448 & 240 & 241 & 245 & 241 & \mathbf{0.134} \\\\ & \textit{Pangaea}\text{ (struct-search)} & 0.752 & \mathbf{0.464} & 7.82 & \mathbf{9.15} & \mathbf{11.68} & \mathbf{5.73} & \mathbf{0.420} & \mathbf{199} & \mathbf{199} & \mathbf{199} & \mathbf{200} & 0.150 \\\\ \hline \end{array}$$ $$\begin{array}{lccc} \hline \text{Setting} & \text{SVR} & \text{Ridge} & \text{XGBoost} \\\\ \hline \text{Topo. \\& Global (no-DFT)} & 5.2 & 5.0 & 5.5 \\\\ \quad \text{+SOAP} & 5.4 & 5.2 & 5.4 \\\\ \quad \text{+LMBTR (K2)} & 4.5 & 4.1 & 3.9 \\\\ \quad \text{+LMBTR (K2+K3)} & 4.0 & 3.6 & 2.7 \\\\ \quad \text{+GRIND (vdW only)} & 5.3 & 6.2 & 5.7 \\\\ \quad \text{+}\textit{Pangaea}\text{ (full-search, struct-only)} & 2.4 & 2.3 & 2.9 \\\\ \quad \text{+}\textit{Pangaea}\text{ (struct-search)} & \mathbf{1.3} & \mathbf{1.6} & \mathbf{1.9} \\\\ \hline \end{array}$$

The results show that Pangaea maintains strong predictive performance without electronic-structure inputs. Specifically, Pangaea (struct-search) achieves the best mean rank across all three learners, while Pangaea (full-search, struct-only) ranks second for Ridge and SVR. Moreover, Pangaea (struct-search) achieves an average relative MAE reduction of 23.6%, closely matching that obtained in the full-input setting.

Among the 12 tasks, performance degradation is observed for the μ\mu task. We therefore further investigate this target's sensitivity to electronic inputs. Specifically, we examine its relationship with DFT-derived features under two settings.

  • Geometric features (G): Geometry, composition, and topology descriptors, along with 12 closed-form functionals covering common statistical aggregations and spatial descriptors (including summary statistics, dipole-like and second-order moments, and the Politzer functional), applied to a geometric control point cloud defined by atomic coordinates and atomic numbers.

  • Electronic features (E): The same 12 closed-form functionals applied to six DFT-derived point clouds, including ADCH charges, four ESP extrema sources, and basin attractors.

The following table summarizes the improvements achieved by adding electronic features (E) to geometric features (G) across two machine-learning models. The results show that μ\mu is the only target that benefits substantially from electronic features, consistent with its direct reconstruction from ADCH charges and atomic coordinates. In contrast, cvc_v shows modest, model-dependent improvements, while the remaining ten targets exhibit little to no consistent predictive benefit from electronic features.

\hline \text{Target} & \text{SVR} & \text{XGBoost} \\\\ \hline \mu & +100\\% & +91\\% \\\\ c_{v} & +19\\% & +4\\% \\\\ \text{Other 10 targets} & -52\\%\text{ to }+9\\% & -9\\%\text{ to }+6\\% \\\\ \hline \end{array}$$ For a broader range of molecular property prediction tasks, obtaining target properties often requires additional calculations following DFT geometry optimization. For example, Mössbauer parameters require calculations of the electron density and electric-field gradient at the nucleus [1], whereas NMR chemical shifts require magnetic-shielding calculations [2]. In such cases, DFT-derived electronic-structure information may provide useful molecular descriptors [3]. With this consideration, we include DFT-derived information as an optional descriptor source in *Pangaea*, providing flexibility to explore additional molecular descriptors for a wider range of prediction tasks. [1] Pápai, M.; Vankó, G. On Predicting Mössbauer Parameters of Iron-Containing Molecules with Density-Functional Theory. J. Chem. Theory Comput. 2013, 9 (11), 5004–5020. https://doi.org/10.1021/ct4007585. [2] Willoughby, P. H.; Jansma, M. J.; Hoye, T. R. A Guide to Small-Molecule Structure Assignment through Computation of ¹H and ¹³C NMR Chemical Shifts. Nat. Protoc. 2014, 9 (3), 643–660. https://doi.org/10.1038/nprot.2014.042. [3] Kiryanov, I. I.; Mukminov, F. H.; Tulyabaev, A. R.; Khalilov, L. M. Prediction of ¹³C NMR Chemical Shifts by Artificial Neural Network. I. Partial Charge Model as Atomic Descriptor. Chemom. Intell. Lab. Syst. 2016, 152, 62–68. https://doi.org/10.1016/j.chemolab.2016.01.010.

2. Strength of baselines (W2 & Q2)

We understand the reviewer’s concerns regarding the strength and fairness of the baseline comparisons. In the original manuscript, we compare our method with four representative molecular representation models. To strengthen the comparison, we include three additional, more recent and advanced models as baselines under comparable input conditions: DGT (Nature Communications, June 2026), EPT (Nature Communications, February 2026), and C-FREE (ICML, July 2026). We evaluate these models in the small-sample QM9 setting adopted in this work, using the same data splits and labeled samples for downstream tasks as in the original manuscript. For C-FREE and EPT, we fine-tune the models using their official pretrained weights. The results are presented below.

\hline \text{Method} & \mu & \alpha & \varepsilon_{\mathrm{HOMO}} & \varepsilon_{\mathrm{LUMO}} & \Delta\varepsilon & R^{2} & \text{ZPVE} & U_{0} & U & H & G & c_{v} & \text{Mean rank} \\\\ \hline \text{SchNet (NIPS 2017)} & \text{0.515} & \text{0.491} & \text{22.51} & \text{20.13} & \text{26.21} & \text{10.09} & \text{15.71} & \text{1050} & \text{1049} & \text{1050} & \text{1050} & \text{0.228} & \text{6.5} \\\\ \text{ViSNet (Nat. Comm. 2024)} & \text{0.397} & \mathbf{0.298} & \text{10.56} & \text{9.977} & \text{13.91} & \text{99.31} & \text{1.375} & \text{891} & \text{888} & \text{887} & \text{882} & \text{0.137} & \text{3.7} \\\\ \text{MolSpectra (ICLR 2025)} & \text{0.222} & \text{0.608} & \text{17.76} & \text{18.51} & \text{15.21} & \text{13.89} & \text{11.49} & \text{9101} & \text{5379} & \text{4098} & \text{4103} & \text{0.192} & \text{6.3} \\\\ \text{MABNet (ICML 2025)} & \text{0.316} & \text{0.370} & \text{9.008} & \text{11.08} & \text{16.80} & \text{12.28} & \text{2.152} & \text{683} & \text{648} & \text{692} & \text{688} & \text{0.158} & \text{3.8} \\\\ \text{DGT (Nat. Comm. 2026)} & \text{0.504} & \text{0.595} & \text{15.46} & \text{17.09} & \text{17.47} & \text{483.20} & \text{13.680} & \text{7.6e4} & \text{7.5e4} & \text{7.5e4} & \text{7.5e4} & \text{0.268} & \text{7.6} \\\\ \text{EPT (Nat. Comm. 2026)} & \mathbf{0.211} & \text{0.364} & \text{18.76} & \text{18.49} & \text{23.23} & \text{7.60} & \text{17.470} & \text{1343} & \text{1339} & \text{1563} & \text{1475} & \text{0.175} & \text{5.6} \\\\ \text{C-FREE (ICML 2026)} & \text{1.030} & \text{8.560} & \text{11.70} & \text{9.35} & \text{14.30} & \text{121.00} & \text{2.000} & \text{4.6e4} & \text{5.9e4} & \text{5.9e4} & \text{5.4e4} & \text{1.990} & \text{6.9} \\\\ \textit{Pangaea}\text{ (SVR, full-search, struct-only)} & \text{0.791} & \text{0.477} & \text{8.13} & \text{9.22} & \text{12.29} & \text{6.21} & \text{0.448} & \text{240} & \text{241} & \text{245} & \text{241} & \mathbf{0.134} & \text{2.7} \\\\ \textit{Pangaea}\text{ (SVR, struct-search)} & \text{0.752} & \text{0.464} & \mathbf{7.82} & \mathbf{9.15} & \mathbf{11.68} & \mathbf{5.73} & \mathbf{0.420} & \mathbf{199} & \mathbf{199} & \mathbf{199} & \mathbf{200} & \text{0.150} & \mathbf{1.9} \\\\ \hline \end{array}$$ As shown in the table, *Pangaea* achieves lower MAEs than all seven deep-learning baselines on 10 of the 12 tasks in the full-search, struct-only setting and on 9 of the 12 tasks in the struct-search setting. These settings yield the second-best and best mean ranks, respectively, demonstrating *Pangaea*’s effectiveness for molecular property prediction in the small-sample setting. **Manuscript revisions.** We add DGT, EPT, and C-FREE as baselines in Table 1 (p. 7).
Reviewer_sNbvOpenReview ↗
Rating
3
Confidence
3
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

The paper studies automated design of molecular 3D descriptors for property prediction in data-limited regimes, where descriptor-based ML remains competitive with end-to-end deep learning. The authors abstract 3D descriptor construction into a three-stage pipeline (Formation, Extraction, Compression) over chemically annotated molecular point clouds, and organize compatible operator combinations into a Unified Construction Space (UCS) with four branches (point, field, expansion, statistics). An Incumbent Revision Agent (IRA), driven by an LLM, iteratively revises the best-so-far pipeline using evaluation history, design-space semantics, and chemical priors, with revisions organized into a four-level hierarchy (branch, representation, carrier, argument). On a 1,498-molecule QM9 subset, Pangaea ranks first among descriptor-based methods for all three ML learners, reduces test MAE by 24% on average over the topology+global baseline, and beats DL baselines on 11/12 tasks; it also improves over expert-crafted coordination descriptors on an Fe Mössbauer dataset. Ablations quantify the value of semantic/chemical guidance and of the hierarchical revision structure, and further analyses cover cross-learner transfer and cross-task reuse.

Critical strengths

1. A unifying, well-grounded abstraction of 3D descriptor construction.

The Formation/Extraction/Compression decomposition, derived from a systematic examination of 20 descriptor families, converts an ad-hoc design problem into a well-defined search space with explicit operator contracts. This is consequential because it is reusable independently of the IRA: any search strategy can operate over the UCS, and the space is extensible to new chemical sources and operators.

2. Thorough and honest empirical analysis.

Beyond the headline results (24% MAE reduction, 11/12 tasks beating DL), the paper includes the ablations a reviewer would ask for: guidance ablation (85% vs 75% attainment), hierarchy ablation (99% vs 80%), cross-learner transfer (51/72 beneficial), and cross-task reuse. The cross-task analysis (Fig. 9) is particularly nice: descriptors transfer within thermodynamically related property clusters (U0/U/H/G) but not across clusters, and transfer can be asymmetric (cv to ZPVE vs the reverse), which suggests the search is recovering physically meaningful structure rather than overfitting a single target.

Critical weaknesses

1. Limited methodological novelty over the LLM-agent-for-design paradigm.

The core loop, an LLM agent revising an incumbent solution over a structured search space guided by evaluation history, is close in spirit to recent LLM-for-scientific-discovery systems. The paper's distinctive contribution is the domain-specific UCS rather than the search mechanism itself, which slightly narrows the conceptual advance.

2. The evaluation is confined to small datasets dominated by small, simple molecules.

QM9-sample retains all 1,498-subset molecules with 1 to 4 heavy atoms while capping the 5 to 9 heavy-atom groups at {100, 150, 250, 350, 600}, so the subset is heavily skewed toward very small molecules; the MB dataset contains only 617 Fe sites. This matters for two reasons. First, small molecules have limited conformational and spatial complexity, which is precisely the regime where 3D descriptors add the least over topological descriptors; the 24% MAE reduction may not transfer to larger, conformationally flexible molecules where 3D information should matter most. Second, with 899 training molecules split across five validation folds, search scores are noisy, and it is unclear whether the IRA's improvements reflect robust signal or adaptation to validation noise. The claims are therefore validated only in the easiest corner of the design space the paper aims to address.

Questions for authors
  1. What is the total computational cost of one search run per task (CPU hours for the 2,000 evaluations, plus LLM token/cost for the IRA)? How does this compare to the cost of training the DL baselines?

  2. How sensitive is IRA's performance to the choice of backbone LLM? Would a smaller or open-weight model achieve comparable attainment ratios?

  3. The cross-learner analysis shows XGBoost-selected descriptors transfer poorly (10/24), concentrated on the thermodynamic energy tasks. Do you have a mechanistic explanation beyond the small search gains on those tasks?

  4. QM9-sample is dominated by molecules with 1 to 4 heavy atoms. Do the conclusions hold on subsets restricted to larger molecules (e.g., 7 to 9 heavy atoms only), or on datasets with conformational flexibility where 3D structure is more decisive? Even a single such experiment would substantially strengthen the paper's central claim.

7 comments

1. Novelty of IRA (W1)

We acknowledge the reviewer's concern regarding the methodological novelty of IRA compared with existing LLM-agent-based design paradigms. Although IRA follows a similar proposal–evaluation–feedback loop, relying solely on LLMs to explore the design space may result in inefficient search. To address this limitation, IRA integrates hierarchical revisions guided by design-space semantics and chemical priors.

Specifically, IRA allocates the proposal budget in each round across four revision levels—branch, representation, carrier, and argument—to balance structural exploration and local refinement. The table below presents ablation results for individual revision levels and a comparison with IRA (w/o hierarchy). Without explicit allocation, approximately 70% of revisions occur at the argument level, indicating a strong bias toward local parameter refinement. Figure 7(b) shows the corresponding mean attainment-ratio curves, suggesting that explicitly balancing structural exploration and local refinement may inform the design of LLM-guided search more broadly.

\hline \text{Variant} & \text{IRA} & \text{w/o arg} & \text{w/o carrier} & \text{w/o branch} & \text{w/o repre.} & \text{w/o hierarchy} \\\\ \hline \text{Attainment ratio (\\%)} & \mathbf{98.6} & 91.0 & 84.5 & 82.2 & 82.0 & 80.2 \\\\ \hline \end{array}$$ By incorporating task information and design-space semantics, IRA improves the mean attainment ratio from 74.8% to 84.8% compared to IRA (w/o guidance), demonstrating the effectiveness of the proposed guidance mechanism ([Figure 7(a)](https://anonymous.4open.science/r/test-6BE9/qm9_layer_ablation_aggregate.pdf)). $$\begin{array}{lcccc} \hline \text{Method} & \text{TPE} & \text{TPE-struct} & \text{IRA (w/o guidance)} & \text{IRA} \\\\ \hline \text{Attainment ratio (\\%)} & 47.7 & 60.6 & 74.8 & \mathbf{84.8} \\\\ \hline \end{array}$$

2. Dataset size, molecular complexity, and validation noise (W2 & Q4)

We understand the reviewer’s questions regarding the molecular-size distribution of QM9-sample, as well as the reliability of search scores and the generalizability of improvements achieved through descriptor search. We provide detailed responses to the points raised below.

(1) Molecular-size distribution in QM9-sample. We first clarify the molecular-size distribution of QM9 and the rationale for our sampling strategy. In the original QM9 dataset, molecules with fewer heavy atoms are much less abundant because the number of possible molecular structures increases substantially with molecular size. To construct a benchmark for the limited-data setting, we therefore retain all available molecules with 1–4 heavy atoms and apply sampling caps of {100, 150, 250, 350, 600} to molecules with 5–9 heavy atoms, respectively.

Importantly, retaining all molecules with 1–4 heavy atoms does not mean that they dominate the resulting subset. Of the 1,498 molecules in QM9-sample, only 48 (3.2%) contain 1–4 heavy atoms, while 1,450 (96.8%) contain 5–9 heavy atoms. Moreover, 1,200 molecules (80.1%) contain 7–9 heavy atoms. The detailed molecular-size distribution is shown below:

\hline \text{Number of heavy atoms} & \text{1} & \text{2} & \text{3} & \text{4} & \text{5} & \text{6} & \text{7} & \text{8} & \text{9} \\\\ \hline \text{Number of molecules} & \text{3} & \text{5} & \text{9} & \text{31} & \text{100} & \text{150} & \text{250} & \text{350} & \text{600} \\\\ \hline \end{array}$$ To further address the reviewer’s concern regarding performance on molecules with 7–9 heavy atoms, we separately evaluate this subset using SVR on the original held-out test set and report the relative MAE improvements (%). **As shown in the table below, the improvements for this subset are comparable to or greater than those for the full test set across nearly all tasks, indicating that _Pangaea_’s benefits are not limited to smaller molecules.** $$\begin{array}{ccccccccccccc} \hline \text{Task} & \mu & \alpha & \varepsilon_{\mathrm{HOMO}} & \varepsilon_{\mathrm{LUMO}} & \Delta\varepsilon & R^{2} & \text{ZPVE} & U_{0} & U & H & G & c_{v} \\\\ \hline \text{1–6 heavy atoms} & \text{76.4} & \text{31.8} & \text{−2.1} & \text{6.7} & \text{−3.9} & \text{31.4} & \text{11.8} & \text{40.4} & \text{39.5} & \text{38.6} & \text{39.5} & \text{1.6} \\\\ \text{7–9 heavy atoms} & \text{81.4} & \text{28.8} & \text{−0.4} & \text{9.6} & \text{6.2} & \text{46.8} & \text{31.9} & \text{49.4} & \text{49.5} & \text{48.7} & \text{49.4} & \text{12.0} \\\\ \text{Full test set} & \text{80.4} & \text{29.3} & \text{−0.9} & \text{9.1} & \text{4.5} & \text{44.5} & \text{28.0} & \text{46.4} & \text{46.1} & \text{45.3} & \text{46.1} & \text{9.8} \\\\ \hline \end{array}$$ **(2) Evaluation on a task sensitive to three-dimensional structure.** Regarding the reviewer's concern about conformational flexibility, we note that the Fe Mössbauer (MB) benchmark included in our manuscript provides a complementary evaluation in which three-dimensional structure plays an important role. Its target properties, particularly isomer shift (IS) and quadrupole splitting (QS), are closely related to the local coordination environment of Fe, with QS being especially sensitive to geometric distortions. Notably, _Pangaea_ further improves predictive performance over a baseline that already incorporates spatial information, demonstrating its ability to construct descriptors that capture relevant three-dimensional structural features. **(3) Search-score reliability and generalization.** To mitigate the effects of validation noise, we evaluate candidate pipelines using five-fold cross-validation on the combined training and validation sets. The held-out test set is strictly excluded from pipeline search and selection. The improvements observed on this independent test set provide further evidence that the gains achieved through pipeline search generalize beyond the validation data.

3. Computational cost (Q1)

We thank the reviewer for raising the question of computational cost. To provide a comprehensive comparison, we report the computational time required for a single IRA search per task alongside the training or fine-tuning time of the deep learning baselines, as summarized in the table below. Specifically, each IRA search involves 2,000 evaluations, requiring approximately 10.6 CPU hours, and consumes approximately 472M input tokens (86.7% cached) and 7.6M output tokens. These computational costs reflect the automation of task-specific 3D molecular descriptor design, reducing reliance on expert knowledge and manual descriptor engineering.

\hline \text{Method} & \text{Time Cost} \\\\ \hline \text{SchNet} & \text{0.053 GPU hours} \\\\ \text{ViSNet} & \text{0.191 GPU hours} \\\\ \text{DGT} & \text{0.037 GPU hours} \\\\ \text{MABNet} & \text{0.717 GPU hours} \\\\ \text{EPT} & \text{0.103 GPU hours (fine-tuning from the official checkpoint)} \\\\ \text{C-FREE} & \text{0.087 GPU hours (fine-tuning from the official checkpoint)} \\\\ \text{MolSpectra} & \text{0.267 GPU hours (fine-tuning) + about 120 GPU hours (pretraining)} \\\\ \text{IRA (ours)} & \text{10.6 CPU hours (2000 evaluations)} \\\\ \hline \end{array}$$
Reviewer_Py1mOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper presents Pangaea, a framework for designing molecular 3D descriptors for property prediction with limited labeled data. Its Unified Construction Space (UCS) organizes descriptor construction into Formation, Extraction, and Compression. An LLM-based Incumbent Revision Agent (IRA) iteratively revises the best-performing pipeline using previous evaluations. Experiments on a QM9 subset and an iron-complex Mössbauer spectroscopy dataset compare the resulting descriptors with conventional descriptors, deep learning models, and alternative search strategies.

Critical strengths
  1. The UCS brings operations from different 3D descriptor families into a common construction space. Combining this representation with chemically informed proposals and refinement based on predictive performance offers a useful approach to automated descriptor design in small-data settings.
  2. The evaluation covers multiple target properties and downstream learning algorithms, as well as alternative search strategies. The reported improvements support the usefulness of the resulting descriptors under the evaluated input conditions.
Critical weaknesses
  1. The conventional descriptor configurations need better justification. LMBTR omits three-body angular information, GRIND uses a custom RDKit-based implementation with a different probe combination, and all three baselines use PCA retaining 80% of the variance. Without assessing these choices, it is difficult to determine how much of the reported advantage depends on the baseline configurations.
  2. The experiments rely on geometries optimized with specific DFT methods and on DFT-derived electronic-structure information. The results therefore leave open how well the framework works with other geometry sources or less demanding inputs. This limits the evidence for its practical applicability when DFT calculations are unavailable.
Questions for authors
  1. What motivates the LMBTR and GRIND configurations and the shared PCA threshold? Please justify these choices or provide representative comparisons using three-body LMBTR terms, an established GRIND implementation, and an assessment of the effect of PCA.
  2. How dependent are the results on the chosen DFT geometries and electronic-structure information? A comparison on a representative task using an alternative geometry source, such as GFN2-xTB, without DFT-derived electronic inputs would help clarify this dependence. Please give conventional descriptors the same inputs and state the input conditions under which applicability is claimed.
  3. Please clarify how IRA explores beyond revisions of the current best pipeline and what role hierarchical revisions play. This part of the procedure is difficult to follow, which limits my assessment of the search design.
7 comments

1. Baseline Configuration (W1 & Q1)

We understand the reviewer’s concerns regarding the rationale for the conventional descriptor baseline configurations. Therefore, we include LMBTR with three-body terms and update the GRIND baseline using the encoder from Denmark Lab's ccheminfolib [1]. To address concerns about baseline settings, we also conduct task-specific parameter searches for handcrafted 3D descriptors rather than relying on the default configurations used in our original experiments.

We also tune PCA for each task by fixing the descriptor parameters selected for that task search and evaluating four options: retaining 80%, 90%, or 95% of the variance, or applying no PCA. Both descriptor configurations and PCA settings are selected on a per-task basis using the same search scores as in Pangaea.

Based on these updates, we repeat the comparison. The mean rank is calculated across the 12 tasks for each learner, with lower values indicating better performance. As shown, Pangaea achieves the lowest mean rank with XGBoost, Ridge, and SVR (2.0, 1.4, and 1.3, respectively), demonstrating its consistently superior performance across tasks even after the baselines are updated

\hline \text{Learner} & \text{Method} & \mu & \alpha & \varepsilon_{\mathrm{HOMO}} & \varepsilon_{\mathrm{LUMO}} & \Delta\varepsilon & R^{2} & \text{ZPVE} & U_{0} & U & H & G & c_{v} & \text{Mean rank} \\\\ \hline \text{XGBoost} & \text{Topo. \\& Global} & 0.664 & 1.046 & 8.98 & 10.32 & 12.60 & 20.97 & 1.059 & 1753 & 1723 & 1723 & 1746 & 0.426 & 4.7 \\\\ & \quad \text{+SOAP} & 0.439 & 1.036 & 9.19 & 10.52 & 12.58 & 20.36 & 0.999 & 1789 & 1788 & 1788 & 1790 & \mathbf{0.415} & 4.4 \\\\ & \quad \text{+LMBTR (K2)} & 0.674 & 1.010 & 8.87 & 9.80 & 12.50 & 20.83 & 0.965 & 1736 & 1736 & 1736 & 1733 & 0.421 & 3.3 \\\\ & \quad \text{+LMBTR (K2+K3)} & 0.658 & 1.012 & 8.95 & \mathbf{9.57} & 12.73 & 20.89 & \mathbf{0.944} & 1735 & 1715 & 1715 & 1697 & 0.425 & 3.0 \\\\ & \quad \text{+GRIND} & 0.439 & 1.054 & 8.85 & 10.24 & 12.67 & 21.15 & 1.066 & \mathbf{1710} & \mathbf{1713} & \mathbf{1713} & 1727 & 0.442 & 3.6 \\\\ & \quad \text{+}\textit{Pangaea} & \mathbf{0.124} & \mathbf{0.952} & \mathbf{8.57} & 9.83 & \mathbf{12.21} & \mathbf{16.23} & 0.979 & 1738 & 1735 & 1715 & \mathbf{1671} & 0.419 & \mathbf{2.0} \\\\ \hline \text{Ridge} & \text{Topo. \\& Global} & 0.769 & 0.813 & 9.50 & 13.24 & 16.56 & 17.93 & 0.610 & 391 & 391 & 391 & 391 & 0.194 & 4.5 \\\\ & \quad \text{+SOAP} & 0.447 & 0.721 & 9.13 & 13.18 & 16.60 & 15.01 & 0.518 & 400 & 400 & 400 & 400 & 0.193 & 4.3 \\\\ & \quad \text{+LMBTR (K2)} & 0.733 & 0.730 & 9.21 & 12.59 & \mathbf{15.59} & 17.24 & \mathbf{0.425} & 396 & 396 & 396 & 396 & 0.187 & 3.3 \\\\ & \quad \text{+LMBTR (K2+K3)} & 0.717 & 0.744 & \mathbf{8.95} & 12.88 & 16.03 & 17.20 & 0.485 & 389 & 389 & 389 & 389 & 0.185 & 2.6 \\\\ & \quad \text{+GRIND} & 0.366 & 0.813 & 9.49 & 13.31 & 16.84 & 18.31 & 0.616 & 396 & 396 & 396 & 396 & 0.191 & 4.9 \\\\ & \quad \text{+}\textit{Pangaea} & \mathbf{0.144} & \mathbf{0.667} & 9.29 & \mathbf{12.40} & 15.86 & \mathbf{8.43} & 0.432 & \mathbf{265} & \mathbf{305} & \mathbf{265} & \mathbf{218} & \mathbf{0.175} & \mathbf{1.4} \\\\ \hline \text{SVR} & \text{Topo. \\& Global} & 0.732 & 0.701 & 8.16 & 10.10 & 13.00 & 11.18 & 0.639 & 460 & 460 & 460 & 460 & 0.150 & 4.3 \\\\ & \quad \text{+SOAP} & 0.468 & 0.589 & 8.24 & 10.07 & 13.04 & 9.91 & 0.612 & 468 & 468 & 468 & 468 & 0.147 & 4.7 \\\\ & \quad \text{+LMBTR (K2)} & 0.695 & 0.577 & \mathbf{7.77} & 9.58 & 12.62 & 10.92 & 0.512 & 461 & 461 & 461 & 461 & 0.145 & 3.3 \\\\ & \quad \text{+LMBTR (K2+K3)} & 0.700 & 0.555 & 8.01 & 9.50 & 12.51 & 10.93 & 0.503 & 452 & 452 & 452 & 452 & 0.140 & 2.4 \\\\ & \quad \text{+GRIND} & 0.367 & 0.709 & 8.21 & 10.20 & 13.38 & 11.62 & 0.627 & 464 & 464 & 464 & 464 & 0.151 & 5.1 \\\\ & \quad \text{+}\textit{Pangaea} & \mathbf{0.143} & \mathbf{0.495} & 8.23 & \mathbf{9.18} & \mathbf{12.42} & \mathbf{6.21} & \mathbf{0.460} & \mathbf{247} & \mathbf{248} & \mathbf{252} & \mathbf{248} & \mathbf{0.135} & \mathbf{1.3} \\\\ \hline \end{array}$$ [1] Zahrt, A. F.; Henle, J. J.; Rose, B. T.; Wang, Y.; Darrow, W. T.; Denmark, S. E. Prediction of Higher-Selectivity Catalysts by Computer-Driven Workflow and Machine Learning. Science 2019, 363 (6424), eaau5631. https://doi.org/10.1126/science.aau5631.

The descriptor parameter search spaces are summarized below.

\hline \text{Descriptor} & \text{Search space} \\\\ \hline \text{SOAP} & \begin{array}{l} \text{Cutoff radius } r_{\mathrm{cut}}\in\\{3,5\\} \text{ Å;} \\\\ \text{radial expansion order } n_{\max}\in\\{6,8\\} \text{;} \\\\ \text{angular expansion order } l_{\max}\in\\{4,6\\} \text{;} \\\\ \text{Gaussian width } \sigma\in\\{0.5,1\\} \end{array} \\\\ \hline \text{LMBTR} & \begin{array}{l} \textbf{Two-body terms (K2)} \text{:} \\\\ \quad \text{Geometry: distance (} g_2=r \text{) or inverse distance (} g_2=r^{-1} \text{)} \\\\ \quad \quad \text{Gaussian width } \sigma_2\in\\{0.1,1\\} \text{ for } g_2=r \\\\ \quad \quad \text{Gaussian width } \sigma_2\in\\{0.01,0.1\\} \text{ for } g_2=r^{-1} \\\\ \quad \text{Cutoff radius } r_{\mathrm{cut},2}\in\\{6,8\\} \text{ Å;} \\\\ \quad \text{pooling } p_2\in\\{\mathrm{mean},\mathrm{sum}\\} \\\\ \textbf{Three-body terms (K3)} \text{:} \\\\ \quad \text{Geometry: angle (} g_3=\theta \text{) or cosine of the angle (} g_3=\cos\theta \text{)} \\\\ \quad \quad \text{Gaussian width } \sigma_3\in\\{10^\circ,30^\circ\\} \text{ for } g_3=\theta \\\\ \quad \quad \text{Gaussian width } \sigma_3\in\\{0.3,0.5\\} \text{ for } g_3=\cos\theta \\\\ \quad \text{Cutoff radius } r_{\mathrm{cut},3}\in\\{6,8\\} \text{ Å;} \\\\ \quad \text{pooling } p_3\in\\{\mathrm{mean},\mathrm{sum}\\} \end{array} \\\\ \hline \text{GRIND} & \begin{array}{l} \text{Anchor separation } d_{\mathrm{anchor}}\in\\{2,6,12\\} \text{ Å;} \\\\ \text{number of retained points } N_{\mathrm{points}}\in\\{50,100\\} \text{;} \\\\ \text{distance bin width } \Delta d\in\\{0.5,1.5\\} \text{ Å} \end{array} \\\\ \hline \end{array}$$ **Manuscript revisions.** We update the baseline results in Table 1 (p. 7) and provide the searched baseline details in Appendix D.4 (pp. 24–25) table S7.

2. Dependence of Pangaea on DFT-Derived Inputs (W2 & Q2)

We understand the reviewer’s concerns regarding the dependence of our method on DFT-optimized geometries and electronic-structure information, as well as its applicability when DFT-derived inputs are unavailable.

In fact, Pangaea separates descriptor source exploration from the input-availability constraints imposed during final pipeline selection. Specifically, it can search the entire source library and select the highest-scoring pipeline that satisfies the prediction-time input constraints. Alternatively, it can restrict the search to sources available at prediction time. To investigate Pangaea's dependence on DFT-derived electronic information, we conduct experiments using only structural inputs at prediction time. Specifically, we establish baselines by excluding DFT-derived electronic features and evaluate Pangaea under two settings: Pangaea (full-search, struct-only)\color{#8B0000}{\textsf{\textbf{Pangaea (full-search, struct-only)}}} searches\color{#8B0000}{\textsf{searches}} the\color{#8B0000}{\textsf{the}} full\color{#8B0000}{\textsf{full}} source\color{#8B0000}{\textsf{source}} library\color{#8B0000}{\textsf{library}} and\color{#8B0000}{\textsf{and}} then\color{#8B0000}{\textsf{then}} selects\color{#8B0000}{\textsf{selects}} the\color{#8B0000}{\textsf{the}} highest-scoring\color{#8B0000}{\textsf{highest-scoring}} pipeline\color{#8B0000}{\textsf{pipeline}} that\color{#8B0000}{\textsf{that}} uses\color{#8B0000}{\textsf{uses}} only\color{#8B0000}{\textsf{only}} structural\color{#8B0000}{\textsf{structural}} inputs\color{#8B0000}{\textsf{inputs}} and\color{#8B0000}{\textsf{and}} no\color{#8B0000}{\textsf{no}} DFT-derived\color{#8B0000}{\textsf{DFT-derived}} electronic-structure\color{#8B0000}{\textsf{electronic-structure}} information,\color{#8B0000}{\textsf{information,}} whereas\color{#8B0000}{\textsf{whereas}} Pangaea (struct-search)\color{#8B0000}{\textsf{\textbf{Pangaea (struct-search)}}} restricts\color{#8B0000}{\textsf{restricts}} the\color{#8B0000}{\textsf{the}} search\color{#8B0000}{\textsf{search}} to\color{#8B0000}{\textsf{to}} structural\color{#8B0000}{\textsf{structural}} sources,\color{#8B0000}{\textsf{sources,}} excluding\color{#8B0000}{\textsf{excluding}} DFT-derived\color{#8B0000}{\textsf{DFT-derived}} electronic-structure\color{#8B0000}{\textsf{electronic-structure}} information\color{#8B0000}{\textsf{information}} from\color{#8B0000}{\textsf{from}} the\color{#8B0000}{\textsf{the}} outset.\color{#8B0000}{\textsf{outset.}}

The results are presented in the table below, where “Best baseline” denotes the lowest MAE among conventional descriptor settings for each task and learner.

\hline \text{Learner} & \text{Setting} & \mu & \alpha & \varepsilon_{\mathrm{HOMO}} & \varepsilon_{\mathrm{LUMO}} & \Delta\varepsilon & R^{2} & \text{ZPVE} & U_{0} & U & H & G & c_{v} \\\\ \hline \text{XGBoost} & \text{Best baseline} & \mathbf{0.500} & 0.983 & 8.88 & \mathbf{9.66} & 12.54 & 20.36 & 0.943 & \mathbf{1613} & \mathbf{1613} & \mathbf{1613} & \mathbf{1604} & 0.403 \\\\ & \textit{Pangaea}\text{ (full-search, struct-only)} & 0.675 & 0.954 & 9.12 & 9.80 & \mathbf{12.36} & 16.23 & 0.981 & 1673 & 1690 & 1670 & 1667 & 0.423 \\\\ & \textit{Pangaea}\text{ (struct-search)} & 0.679 & \mathbf{0.952} & \mathbf{8.83} & 9.99 & 12.53 & \mathbf{15.81} & \mathbf{0.882} & 1685 & 1668 & 1658 & 1649 & \mathbf{0.396} \\\\ \hline \text{Ridge} & \text{Best baseline} & \mathbf{0.534} & 0.679 & 9.05 & 11.85 & 15.37 & 15.01 & 0.409 & 381 & 381 & 381 & 381 & 0.227 \\\\ & \textit{Pangaea}\text{ (full-search, struct-only)} & 0.771 & 0.630 & \mathbf{8.73} & 12.28 & 15.43 & 8.43 & 0.423 & 269 & 307 & 270 & \mathbf{211} & 0.209 \\\\ & \textit{Pangaea}\text{ (struct-search)} & 0.740 & \mathbf{0.590} & 9.17 & \mathbf{11.81} & \mathbf{14.83} & \mathbf{7.71} & \mathbf{0.379} & \mathbf{211} & \mathbf{211} & \mathbf{211} & \mathbf{211} & \mathbf{0.198} \\\\ \hline \text{SVR} & \text{Best baseline} & \mathbf{0.542} & 0.496 & \mathbf{7.79} & 9.33 & 12.35 & 9.91 & 0.480 & 435 & 435 & 435 & 435 & 0.160 \\\\ & \textit{Pangaea}\text{ (full-search, struct-only)} & 0.791 & 0.477 & 8.13 & 9.22 & 12.29 & 6.21 & 0.448 & 240 & 241 & 245 & 241 & \mathbf{0.134} \\\\ & \textit{Pangaea}\text{ (struct-search)} & 0.752 & \mathbf{0.464} & 7.82 & \mathbf{9.15} & \mathbf{11.68} & \mathbf{5.73} & \mathbf{0.420} & \mathbf{199} & \mathbf{199} & \mathbf{199} & \mathbf{200} & 0.150 \\\\ \hline \end{array}$$

Comments on the paper

2 comments
AuthorsOpenReview ↗

Summary of Responses to Reviewers and Key Contributions [1/2]

We sincerely thank all reviewers for their insightful and constructive feedback. We are grateful that Reviewer 8zZW and Reviewer sNbv have explicitly confirmed that their concerns have been addressed. Although we have received no follow-up feedback from Reviewer Py1m to date, we have provided detailed responses and substantial additional experiments addressing all three of the reviewer's major concerns, some of which overlap with issues raised by the other two reviewers. Therefore, we provide this summary to consolidate our rebuttal responses and highlight the key contributions of our work, aiming to facilitate the final assessment by the AC, SAC, and PCs.

Our work focuses on automated molecular 3D descriptor design for molecular property prediction in data-limited settings. In these settings, improving predictive performance remains challenging. While advances in data and models are important, labeled data are often scarce in specialized chemical tasks, and increasingly sophisticated deep-learning models do not necessarily outperform descriptor-based ML methods. This motivates us to explore another important direction beyond model refinement: the design of more effective molecular descriptors. However, designing task-specific 3D descriptors still relies heavily on chemical expertise and manual effort. By examining 20 representative descriptor families, we identify a shared three-stage process—Formation, Extraction, and Compression—which transforms descriptor design into a structured search problem (Figure 1). Building on this formulation, we propose Pangaea, an automated framework for designing task-specific molecular descriptors tailored to diverse molecular prediction tasks. Pangaea integrates a Unified Construction Space (UCS) for composing executable descriptor pipelines with an Incumbent Revision Agent (IRA) that searches this space guided by evaluation history, design-space semantics, and chemical priors. On QM9-sample, Pangaea achieves an average 24% relative reduction in test MAE and outperforms the best evaluated deep-learning baseline on 11 of 12 tasks.

AuthorsOpenReview ↗

Summary of Responses to Reviewers and Key Contributions [2/2]

During the rebuttal, we conducted additional experiments and analyses to address the reviewers' major concerns:

No.TopicReviewer concernOur responseWhere addressedResolution status
1Conventional baseline fairnessWhether the LMBTR, GRIND, and PCA configurations are sufficiently strong and fairly compared.Added three-body LMBTR, an established GRIND implementation, and task-specific searches over descriptor parameters and PCA settings. Pangaea retains the best mean rank across all three ML learners.Py1m rebuttal-1. In the revised manuscript, we update the baseline results in Table 1 (p. 7) and describe how the descriptor parameters were tuned in Appendix D.4 (pp. 24–25).Py1m: No follow-up received.
2Dependence on DFT-derived inputsWhether Pangaea relies on DFT-derived geometries and electronic information, and whether its advantages persist without them.Structure-only search achieves a 23.6% average relative MAE reduction. It also achieves the best mean rank across the 12 tasks for each of the three ML learners. With GFN2-xTB geometries and matched inputs, Pangaea achieves the lowest MAE on 10/12 tasks across its two structure-only configurations.Py1m rebuttal-2 and 8zZW rebuttal-1. In the revised manuscript, we add Section 6.3 (p. 8) to report structure-only predictions and comparisons using GFN2-xTB geometries.Py1m: No follow-up received. 8zZW: Confirmed concerns addressed, including the overlapping electronic-input concern.
3IRA novelty and exploration mechanismHow IRA differs from conventional LLM-guided search and explores beyond local refinement.Introduced hierarchical revisions guided by design-space semantics and chemical priors, enabling structural exploration and local refinement. Additional ablations demonstrate their effectiveness.Py1m rebuttal-3; sNbv rebuttal-1Py1m: No follow-up received. sNbv: Confirmed concerns addressed, including the overlapping methodological concern.
4Stronger deep-learning baselinesWhether Pangaea remains competitive against stronger and more recent molecular representation models.Added DGT, EPT, and C-FREE under comparable conditions. Pangaea achieves lower MAEs than all seven deep-learning baselines on 10 of the 12 tasks in the full-search, struct-only setting and on 9 of the 12 tasks in the struct-search setting.8zZW rebuttal-28zZW: Confirmed concerns addressed.
5Backbone LLM robustnessWhether IRA's effectiveness depends on a particular backbone LLM.We evaluated three alternative open-weight LLMs. With each backbone, IRA retained a substantial search advantage over black-box optimizers.sNbv rebuttal-4sNbv: Confirmed concerns addressed.
6Experimental reliability and generalizabilityConcerns about molecular size, validation reliability, cross-learner transfer, and computational cost.Clarified the dataset distribution, evaluated larger-molecule subsets, and provided additional validation, cross-learner, and computational-cost analyses.sNbv rebuttal-2; sNbv rebuttal-3; sNbv rebuttal-5sNbv: Confirmed concerns addressed.

Ultimately, our central contribution to cheminformatics is a new formulation of molecular 3D descriptor design as an automated, task-driven process: rather than relying on fixed, expert-crafted representations, Pangaea enables the systematic, composable, and searchable construction of task-specific 3D descriptors. Together with the supporting empirical evidence, this framework provides a foundation for automated molecular representation design across diverse prediction tasks.

We sincerely thank all reviewers, the Area Chair, Senior Area Chairs, and Program Chairs for their time and thoughtful consideration.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.