acceptodds
Under review as a conference paper at ICLR 2027

AFIMol: Adaptive Fragmentation and Fragment-Pocket Interaction Learning for 3D Molecular Generation

14 views

Abstract

Generating 3D ligands conditioned on protein pockets remains a central challenge in structure-based drug design. However, fragment-based generators often rely on fixed chemical rules for molecular decomposition and lack explicit supervision for evaluating candidate fragments according to their residue-level interaction preferences. Consequently, predefined fragment boundaries may fail to capture molecule-specific structural organization, while fragment selection may not fully exploit localized interaction evidence from the protein pocket. To address these limitations, we propose AFIMol, an autoregressive framework for 3D molecular generation that integrates adaptive fragmentation with fragment-pocket interaction learning. Rather than determining fragment boundaries solely through fixed cutting rules, AFIMol learns molecule-specific boundary scores from atom-level structural responses while restricting cuts to chemically admissible bonds. At each generation step, AFIMol combines the current generation state with each candidate fragment to predict its interaction preferences over pocket residues. The interaction predictor is supervised using contact annotations from protein-ligand complexes, and the predicted residue-level interactions are aggregated into compatibility scores that adjust candidate logits during inference.This mechanism guides the generator toward fragments that better match the local pocket environment. Experiments on CrossDocked2020 show favorable synthetic accessibility and molecular diversity while maintaining competitive docking performance. Ablation studies validate the contributions of adaptive fragmentation and fragment-pocket interaction learning, while structural case studies illustrate the localized interaction patterns of the generated ligands.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Molecular Generation, Adaptive Fragmentation, Fragment-Pocket Interaction

Reviews

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

This paper introuced a pocket-based molecule generation method targeting on modeling fragment-pocket interaction.

Critical strengths
  1. The benchmark is well-established with very comprehensive literature/baseline review as well as reasonable metrics. Multiple views were evaluated during the baseline comparison.
  2. The method was clearly presented.
Critical weaknesses
  1. Although the authors showed that using adaptive fragment-based generation can improve SA and diversity, the method is still not novel and insightful enough. There should be more insights and conclusions, for example, proposing a way to improve energy and affinity.
  2. The motivation for proposing such a method/these modules is not strong or supportive enough based on the experimental results.
  3. The overall performance is still not perfect. There should be more improvements to the method.
  4. More pocket-based generation methods should be compared in the benchmark table.
Questions for authors

See weakness.

5 comments
AuthorsOpenReview ↗

Methodological Novelty and Research Insights

Thank you for this comment. We agree that binding affinity is an important objective, but argue that pocket-conditioned molecular generation should not be judged by docking score alone. A useful candidate must also be synthesizable, drug-like, and structurally diverse.

Table 1 illustrates this trade-off. AMDiff obtains a more favorable Vina score than AFIMol (-7.466 vs. -7.241), but has lower QED (0.479 vs. 0.596), SA (0.681 vs. 0.788), and Diversity (0.729 vs. 0.771). PMDM also reports better Vina (-7.572) and High Affinity (0.628 vs. 0.589), but substantially lower SA (0.611) and Diversity (0.709). Thus, better docking scores do not necessarily translate into better overall molecular-generation quality.

AFIMol’s methodological insight is to model two complementary decisions within fragment-autoregressive generation: SAF learns molecule-dependent priorities among chemically admissible fragmentation boundaries, while FIPRA evaluates each candidate fragment through residue-level, candidate-conditioned pocket interaction reasoning and adjusts its generation logit. This is more specific than simply combining fragments with global pocket conditioning.

We agree that explicit energy- or affinity-guided optimization is a valuable direction. However, adding such an objective does not automatically improve overall candidate quality and may shift the balance among binding, synthesizability, and drug-likeness. AFIMol therefore focuses on improving this balance rather than claiming to minimize docking energy universally. (Weakness 1)

AuthorsOpenReview ↗

Motivation and Experimental Support

Thank you for raising this concern. The motivation for AFIMol is supported both by the problem each module addresses and by controlled experiments. It was also recognized in the other reviews: Reviewer 1 described the combination of adaptive fragmentation and residue-level interaction guidance as “well motivated”; Reviewer 2 considered adaptive fragmentation and pocket-conditioned fragment scoring “plausible ways” to represent molecules while retaining protein-context information.

SAF addresses a limitation of fixed fragmentation rules: they identify admissible cuts but do not rank alternative boundaries according to each molecule’s local context. Figure 2 shows that SAF produces fewer, larger fragments. The non-zero LM-only normalized gradient (0.012) further confirms that the generation objective supplies a learning signal to its boundary scorer. Consistently, removing SAF reduces SA from 0.788 to 0.765, Diversity from 0.771 to 0.753, and High Affinity from 0.589 to 0.523.

FIPRA addresses the need to distinguish candidate fragments by their compatibility with pocket residues. Removing it changes Vina from -7.241 to -7.082 and High Affinity from 0.589 to 0.520. Figure 3 provides complementary outcome-level evidence: median nearby-residue density changes by +0.060 and carbon-associated proximity by +0.047. These are geometric analyses, not contact-prediction accuracy measurements.

Finally, we examined synthetic feasibility beyond the SA proxy. For four representative molecules generated for pocket 4XLI, AiZynthFinder found top-ranked routes of 2–4 steps terminating in stock-listed precursors; the corresponding solved-route fractions were 28.0%, 32.3%, 37.9%, and 80.0%. These computational results support the relevance of the design objective, while experimental synthesis and activity remain to be validated. (Weakness2)

AuthorsOpenReview ↗

Overall Performance

Thank you for raising this concern about the overall performance of AFIMol. We agree that AFIMol does not lead on every metric, and do not claim otherwise. Its contribution is a competitive multi-objective performance profile.

AFIMol achieves Vina of -7.241, High Affinity of 0.589, QED of 0.596, SA of 0.788, LogP of 3.003, Lipinski-related score of 4.959, and Diversity of 0.771. In Table 1, AFIMol achieves the best reported SA and Diversity, while its High Affinity ranks second among methods reporting this metric (PMDM: 0.628). We acknowledge that several methods obtain more favorable mean docking scores.

Ablation experiments further demonstrate the complementary contributions of SAF and FIPRA:

Without SAF: SA decreases from 0.788 to 0.765, Diversity from 0.771 to 0.753, and High Affinity from 0.589 to 0.523.

Without FIPRA: Vina worsens from -7.241 to -7.082 and High Affinity decreases from 0.589 to 0.520, although QED and Diversity increase to 0.625 and 0.782, respectively.

Without both modules: The model obtains its least favorable Vina (-6.789), SA (0.671), and Diversity (0.711).

These non-monotonic changes demonstrate that the two modules contribute to different, partially competing objectives rather than uniformly improving every metric.

Evaluation scale and consistency. We evaluate 100 test pockets, performing 1,000 sampling attempts per pocket and retaining 100 valid, unique molecules. This yields 10,000 evaluated molecules. The full model and all ablations follow identical sampling, filtering, docking, and evaluation procedures.

Overall, these results demonstrate the complementary contributions of SAF and FIPRA and AFIMol's favorable balance across multiple molecular-generation objectives, rather than universal superiority on every individual metric. (Weakness3)

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

This paper presents AFIMol, a 3D molecular generation method that combines Structure-Aware Fragmentation (SAF) with Fragment–Pocket Interaction Reasoning and Aggregation (FIPRA). SAF scores chemically admissible bonds, selects bonds to cut, and represents the resulting connected components as fragment tokens. An autoregressive GPT-style generator then predicts fragments conditioned on a protein pocket. FIPRA scores candidate fragments using learned fragment and residue features, predicts residue-level interaction beliefs, aggregates them into compatibility scores, and adjusts the candidate logits.

The experiments use the CrossDocked benchmark and compare AFIMol with several existing molecular-generation systems. AFIMol reports strong synthetic-accessibility and diversity results, while its docking score is competitive but not uniformly best. Ablations remove SAF, FIPRA, or both. The paper’s central proposal is promising, but several links between learned fragmentation, the fragment vocabulary, the generator, and the resulting 3D molecules are not specified clearly enough to assess or reproduce the method.

Critical strengths
  1. The paper addresses a relevant limitation of fragment-based 3D generation. Adaptive fragmentation and pocket-conditioned fragment scoring are plausible ways to represent molecules at a useful level of granularity while retaining protein-context information.

  2. The reported results show potentially useful gains on selected objectives. AFIMol reports strong synthetic accessibility and diversity, and the ablations indicate that SAF and FIPRA affect the measured outcomes. These results motivate further study of the proposed components, although they do not resolve the methodological questions below.

Critical weaknesses
  1. The SAF-to-generator training and representation pipeline is not specified sufficiently to evaluate the central method. The paper describes a bond-scoring objective and mentions a straight-through estimator, but does not make clear how the language-model loss supplies learning signal to individual boundary decisions through top-BB selection, connected-component construction, and discrete fragment tokenization. The paper also does not fully define how fragments map to token IDs, how the vocabulary and geometric tokens are constructed, or how SAF and the GPT-style generator are trained together. These details determine what model is actually learned and how the claimed adaptive fragmentation affects generation.

  2. The reported atom-retention statistics are difficult to reconcile with the stated fragmentation definition. Equation (3), F=CC⁡(V,E∖S)F=\operatorname{CC}(V,E\setminus S), removes selected edges and takes connected components over the original vertex set, which by itself partitions the atoms without dropping them. Appendix D.3 nevertheless reports substantially lower atom retention for several elements (for example, 57.2057.20\\% for oxygen on the test set). The paper does not explain what filtering, vocabulary mapping, or other processing accounts for the missing atoms. This leaves unclear what molecular representation is used as the generator target and whether the reported fragment sequence represents the complete ligand.

Questions for authors
  1. Underlying generator and end-to-end pipeline: What is the underlying GPT generator (architecture and source), how is the pocket provided as conditioning input, and how are fragment identity and geometric tokens represented? Please describe how a generated fragment sequence is assembled into a complete 3D molecule.

  2. SAF learning signal and gradient: Equation (4) regularizes the mean boundary response and its binarity, while Equation (9) also includes the language-model loss. What exact gradient or surrogate-gradient path allows the language-model objective to train the individual bond scores through top-BB selection, connected components, and discrete tokenization? Please specify the straight-through estimator used.

  3. Vocabulary and tokenization: How are fragment token IDs and geometric tokens created when SAF changes the fragment boundaries? Is the vocabulary fixed before GPT training, updated during training, or rebuilt for each input? Are SAF and the generator trained jointly or in separate stages?

  4. SAF during inference: Is SAF used only to construct training fragments and the fragment vocabulary, or does it also run during generation? Please provide the full inference sequence from pocket input to generated ligand.

  5. OOV fragments and coverage: Figure 2 reports a 25.325.3\\% test OOV token rate for SAF. How are these out-of-vocabulary fragments represented or handled, and what fraction of test ligands can be represented completely under the reported vocabulary? Please clarify whether the OOV statistic concerns reference test fragments, generated outputs, or both.

  6. Atom retention: How should the element-wise atom-retention values in Appendix D.3 be interpreted given Equation (3), which partitions the original atom set? Which specific step removes or omits atoms, and how does that affect the target fragment sequences and evaluation?

  7. Structural features used by SAF: Figure 1 describes a 3D ligand input, while the SAF formulation lists atom identity, bond context, and local graph topology. Are three-dimensional coordinates, distances, or other geometric features used by the SAF encoder? Please define what “structure-aware” refers to in this module.

  8. Candidate geometry in FIPRA: Figure 1 lists residue position, distance, and orientation features. At what point in generation are these features computed, and do they encode candidate-fragment-specific relative geometry or only pocket/residue features? How is a candidate fragment’s pose represented when its compatibility score is calculated?

  9. Metric implementation and aggregation: Please specify the exact implementation and aggregation for SA, Lipinski, Diversity, and High Affinity, including normalization, molecular fingerprints/similarity, the reference ligand used for the affinity comparison, the number of generated and retained molecules, and whether reported values are aggregated per molecule or per pocket.

10 comments
AuthorsOpenReview ↗

Generator architecture, pocket conditioning, and tokenization

Thank you for raising these important questions about AFIMol's generator architecture, molecular representation, and end-to-end generation pipeline. We appreciate the opportunity to clarify these implementation details and explain how the protein pocket and molecular fragments are represented and integrated.

(1) Generator and pocket conditioning. AFIMol uses a custom decoder-only fragment-autoregressive Transformer implemented in PyTorch and initialized from a pretrained GPT checkpoint. It has 12 Transformer layers, 12 attention heads, a 768-dimensional hidden state, and a 3,072-dimensional FFN. Each block contains causal self-attention and protein-residue cross-attention. The pocket is represented by masked, 512-dimensional residue embeddings that integrate ESM embeddings with preprocessed structural and chemical features. After projection to 768 dimensions, residue representations serve as cross-attention keys and values, while molecular hidden states serve as queries.

(2) Fragment vocabulary and tokenization. Before Transformer optimization, we construct a fixed vocabulary from all unique SAF fragments in the training set, without frequency filtering, and add sequence-control tokens. Each fragment is canonicalized and mapped to its vocabulary ID. The vocabulary is fixed throughout training and inference; it is neither updated during training nor rebuilt for individual inputs.

Each autoregressive step represents:

  • Fragment identity: a discrete token identifying the chemical substructure.
  • Connection: tokens encoding bonds between generated fragments.
  • Position: pocket-centered Cartesian coordinates, quantized into 128 bins per axis.
  • Orientation: quaternion tokens from a 256-code vector-quantization vocabulary.

These tokens are serialized and embedded as inputs to the Transformer. The backbone predicts fragment identity and associated structural tokens autoregressively.

(3) Relation to the complete generation pipeline. The pocket representations condition the molecular sequence through cross-attention at every Transformer block. FIPRA provides an additional candidate-level refinement of fragment-identity logits; it does not replace the backbone's cross-attention or directly modify connection- and geometry-token logits. The generated fragment identities specify chemical substructures, while connection and geometric tokens provide the information needed for subsequent molecular reconstruction.

(Weakness 1&Questions 1 and 3 — Part 1 of 2)

AuthorsOpenReview ↗

SAF gradient pathway, joint training, and inference

Thank you for raising these important questions about how the language-model objective trains SAF and how the learned fragmentation mechanism interacts with the generator. We clarify the surrogate-gradient pathway, the straight-through estimator, the joint training procedure, and the complete inference process below.

(1) Hard fragmentation and differentiable surrogate.

In the forward pass, SAF predicts a boundary response b_ij for each chemically admissible bond. Top-B selection and connected components produce the hard fragments:

S=TopB⁡(Ecut;b),F=CC⁡(V,E∖S).S=\operatorname{TopB}(E_{\mathrm{cut}};\mathbf b),\qquad \mathcal F=\operatorname{CC}(V,E\setminus S).

We do not differentiate through these discrete operations. Instead, we define soft bond-retention weights:

Wij={1−bij,(i,j)∈Ecut,1,(i,j)∈E∖Ecut,0,(i,j)∉E.W_{ij}= \begin{cases} 1-b_{ij}, &(i,j)\in E_{\mathrm{cut}},\\\\ 1, &(i,j)\in E\setminus E_{\mathrm{cut}},\\\\ 0, &(i,j)\notin E. \end{cases}

The symmetrically normalized adjacency and graph-diffusion kernel are:

W~=D−1/2WD−1/2,Dii=∑jWij+ϵ,C=(I−λW~)−1,0<λ<1.\widetilde W=D^{-1/2}WD^{-1/2},\quad D_{ii}=\sum_jW_{ij}+\epsilon,\qquad C=(I-\lambda\widetilde W)^{-1},\quad 0<\lambda<1.

For hard fragment k, with atom set V_k held fixed during backpropagation, define:

qki=1∣Vk∣∑s∈VkCsi,Asoft[k,i]=exp⁡(qki/τA)∑rexp⁡(qri/τA).q_{ki}=\frac{1}{|V_k|}\sum_{s\in V_k}C_{si},\qquad A_{\mathrm{soft}}[k,i]= \frac{\exp(q_{ki}/\tau_A)} {\sum_r\exp(q_{ri}/\tau_A)}.

(2) Straight-through estimator and gradient verification.

We use:

AST=Asoft+sg⁡(Ahard−Asoft).A_{\mathrm{ST}}=A_{\mathrm{soft}}+ \operatorname{sg}(A_{\mathrm{hard}}-A_{\mathrm{soft}}).

Thus, the forward pass uses the hard assignment, while the backward gradient follows A_soft. Fragment representations are pooled as:

zk=∑iAST[k,i]ui∑iAST[k,i]+ϵ.\mathbf z_k= \frac{\sum_iA_{\mathrm{ST}}[k,i]\mathbf u_i} {\sum_iA_{\mathrm{ST}}[k,i]+\epsilon}.

They are fused with fragment-identity and geometric-token embeddings and projected into the Transformer space. The surrogate path is:

LLM→zk→Asoft→q→C→W~→bij→θSAF.\mathcal L_{\mathrm{LM}}\rightarrow\mathbf z_k \rightarrow A_{\mathrm{soft}}\rightarrow q\rightarrow C \rightarrow\widetilde W\rightarrow b_{ij} \rightarrow\theta_{\mathrm{SAF}}.

The graph-diffusion derivative is:

dC=λC(dW~)C.dC=\lambda C(d\widetilde W)C.

Discrete component membership and token lookup are not differentiated.

We independently backpropagated each objective and measured the SAF gradient normalized by the SAF parameter norm: 0.012 for LM loss alone, 0.006 for fragment-interaction loss alone, 0.035 for SAF regularization alone, and 0.041 for the joint objective. The nonzero LM-only gradient directly confirms that generation loss updates SAF boundary-scoring parameters.

(3) Joint training and inference.

SAF and the generator are jointly optimized on teacher-forced sequences using language-model, fragment-interaction, and SAF-regularization losses. Training uses AdamW (learning rate 1e-6), cosine scheduling, and linear warm-up.

SAF is not rerun at inference because the input is a pocket, not a ligand graph to partition. The pocket is encoded and the sequence starts with a start token. At each step, the Transformer predicts fragment logits; FIPRA adjusts only fragment-identity logits using candidate-conditioned residue compatibility. The backbone then generates connection and geometric tokens until termination. Decoding maps fragment tokens to chemical substructures, restores inter-fragment bonds from connection tokens, and places fragments using geometric tokens in the pocket-centered frame. The complete 3D molecule is then reconstructed and checked for chemical validity.

(Weakness 1&Questions 2 and 4 — Part 2 of 2)

AuthorsOpenReview ↗

OOV Fragments and Vocabulary Coverage

Thank you for asking us to clarify the OOV statistics and the completeness of the fragment representation.

(1) OOV statistics and evaluation scope.

Figure 2(d) evaluates reference fragments obtained by applying SAF to held-out test ligands and matching them against the fixed training vocabulary. It does not measure generated outputs.

  • Test OOV-token rate: 25.3% for SAF, compared with 3.7% for conventional fragmentation.
  • Test OOV-type rate: 39.5% for SAF, compared with 18.2% for conventional fragmentation.

The 25.3% OOV-token rate means that 25.3% of reference-fragment occurrences are absent from the vocabulary. Generated outputs do not have OOV fragment IDs because their fragment identities are selected from the fixed vocabulary.

(2) OOV handling.

No UNK fallback is used for held-out OOV reference fragments. They are counted in the test OOV statistics and excluded only from the numerator of exact-match vocabulary-coverage analysis.

Importantly, this filtering is not applied during training: training uses complete fragment sequences and retains all ligand atoms.

(3) Fragment-level and complete-ligand coverage.

The 25.3% OOV-token rate corresponds to 74.7% exact vocabulary coverage at the fragment-occurrence level. Complete-ligand coverage is evaluated separately: a ligand is considered covered only if every SAF fragment has an exact vocabulary match.

The measured test-set coverage is:

  • Exactly 2 fragments: 52.83%.
  • Exactly 3 fragments: 39.68%.
  • Exactly 4 fragments: 28.14%.
  • Overall: 40.22%.

These are directly measured test-set values, not estimates obtained by exponentiating the fragment-level coverage.

Thus, Figure 2(d) measures reference-test vocabulary generalization, while complete-ligand coverage separately quantifies whether all fragments of an individual ligand can be represented exactly.

(Question 5)

Reviewer_3wbQOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes AFIMol, a fragment-based autoregressive framework for pocket-conditioned 3D molecular generation. It introduces two components: Structure-aware Adaptive Fragmentation (SAF), which learns molecule-specific fragmentation boundaries, and a Fragment Interaction-aware Pocket Residue Adapter (FIPRA), which incorporates residue-level interaction predictions into fragment selection. Experiments on CrossDocked2020 demonstrate improvements in synthetic accessibility and molecular diversity while maintaining competitive docking performance.

Critical strengths
  1. The paper addresses relevant limitations of existing fragment-based molecular generation methods. Combining adaptive fragmentation with residue-level interaction guidance is well motivated.

  2. The proposed framework integrates both components into an autoregressive generation pipeline. The methodology is generally clear and technically interesting.

  3. The experimental evaluation includes component ablations and mechanistic analyses. The improvements in SA (0.788) and diversity (0.771) are encouraging, and the authors appropriately acknowledge trade-offs in other metrics.

Critical weaknesses
  1. In Section 4.2, the boundary scores in Eq. (2) are computed from MPNN-based atom representations, but no explicit 3D geometric features are specified. It is therefore unclear whether SAF captures spatial structural information or primarily learns from molecular graph topology. Furthermore, Eq. (3) involves discrete Top-B selection and fragment construction, while Eq. (4) only regularizes the cutting ratio and boundary confidence. The paper does not clearly explain how the generation objective guides the learned fragmentation decisions.

  2. FIPRA is trained using residue-contact labels associated only with the ground-truth fragment (Eq. (8)), whereas it scores all candidate fragments during inference (Eq. (10)). This creates a potential gap between training supervision and inference-time fragment ranking. Moreover, Figure 3 mainly reports geometric proximity statistics rather than actual contact prediction accuracy. The relatively small improvement in polar-atom proximity (+0.007) also makes it difficult to determine whether FIPRA effectively captures the intended residue-specific interactions.

  3. The experimental evidence does not fully establish the robustness of the reported improvements. In Table 1, AFIMol achieves a Vina Score of −7.241, which is worse than Frag2Seq (−7.366) and several other baselines.

Questions for authors

see the weakness

5 comments
AuthorsOpenReview ↗

Structural Information in SAF and Generation-Objective Guidance

Thank you for raising these important questions about the structural information encoded by SAF and how the generation objective guides its fragmentation decisions. We clarify SAF's structural input and how the generation objective trains it. (1) Structural information. SAF captures molecular structure through graph representations, not explicit 3D coordinates or interatomic distances. The MPNN encodes atom identity, bond context, and local topology. Let uiu_i denote atom ii's representation. SAF computes:

\xi_{ij}=[a_i,a_j,|a_i-a_j|,u_i\Vert u_j].$$ A learnable pair encoder predicts the boundary response: $$d_{ij}=\psi_\theta(\xi_{ij}),\quad b_{ij}=\sigma\left(\frac{d_{ij}-\mu_b}{\tau_b}\right).$$ Thus, cutting priorities depend on local molecular context, not merely bond type. SAF selects the top $B$ admissible bonds and constructs fragments through connected components: $$S=\operatorname{TopB}(E_{\mathrm{cut}};\mathbf b),\quad \mathcal F=\operatorname{CC}(V,E\setminus S).$$ SAF therefore learns molecule-dependent fragmentation boundaries. The generator's geometric tokens model 3D placement, while FIPRA provides pocket-residue interaction guidance. (2) Generation-objective guidance. Top-B and connected components produce the hard assignment $A_{\mathrm{hard}}$. Continuous boundary responses yield a differentiable soft assignment $A_{\mathrm{soft}}$. We use the straight-through estimator (STE): $$A_{\mathrm{ST}}=A_{\mathrm{soft}}+ \operatorname{sg}(A_{\mathrm{hard}}-A_{\mathrm{soft}}).$$ The forward pass uses hard fragments, while gradients follow the soft assignment. Atom representations are pooled to obtain fragment representations: $$\mathbf z_k= \frac{\sum_i A_{\mathrm{ST}}[k,i]\mathbf u_i} {\sum_i A_{\mathrm{ST}}[k,i]+\epsilon}.$$ These representations are fused with fragment-identity and geometric-token embeddings before Transformer input. The gradient pathway is: $$\mathcal L_{\mathrm{LM}} \rightarrow A_{\mathrm{soft}} \rightarrow b_{ij} \rightarrow\theta_{\mathrm{SAF}}.$$ We do not differentiate through discrete Top-B selection, connected-component construction, or token lookup. Instead, gradients propagate through differentiable structural pooling. The normalized SAF gradients are 0.012 for the LM loss alone, 0.006 for the fragment-interaction loss alone, 0.035 for SAF regularization alone, and 0.041 for the joint objective. Crucially, the nonzero LM-only gradient directly verifies that the generation objective independently updates SAF's bond-boundary parameters. Thus, SAF is a trainable, generation-coupled component rather than merely a fixed preprocessing operation. **(Weakness1)**
AuthorsOpenReview ↗

FIPRA Training Supervision and Inference-Time Candidate Ranking

Thank you for distinguishing residue-contact supervision from inference-time ranking. FIPRA uses the same candidate-conditioned scoring mechanism during training and inference.

1. Shared candidate scoring in training and inference.

At each fragment-selection step, the current generation state is combined with each candidate embedding. The shared interaction network compares each query with the pocket-residue representations and aggregates residue-level evidence into a candidate-specific score, which adjusts the corresponding logit under Eq. (10). Thus, all candidates are scored in both training and inference.

2. Different roles of supervision.

  • Contact supervision: Eq. (8) uses reference-complex contact annotations to supervise the teacher-forced ground-truth fragment.
  • Candidate ranking: The autoregressive objective, through the adjusted logits, trains the relative ranking of the full candidate set.
  • Unobserved candidates: They are not treated as contact negatives, since absence from one reference ligand does not establish incompatibility with the pocket. Such labeling would introduce false negatives.

3. Geometric evidence and ablation results.

Figure 3 provides an outcome-level geometric analysis, not a contact-prediction accuracy benchmark. Adding FIPRA changes:

  • Median nearby-residue density by +0.060.
  • Carbon-associated proximity by +0.047.
  • Polar-atom proximity by +0.007.
  • Ligand N/O fraction by +0.063.

We therefore interpret these statistics as geometric evidence, not direct contact-classification accuracy. Consistently, removing FIPRA in Table 2 worsens Vina from -7.241 to -7.082 and High Affinity from 0.589 to 0.520.

Conclusion. Candidate ranking is trained through the shared scorer and autoregressive objective, while direct contact labels provide additional supervision only for the observed reference continuation. (Weakness2)

AuthorsOpenReview ↗

Robustness of Experimental Improvements and Vina Score

Thank you for raising this concern. We acknowledge that AFIMol does not achieve the best mean Vina Score in Table 1. However, our objective is balanced molecular generation rather than optimizing docking scores at the expense of other essential molecular properties.

(1) Better docking scores do not necessarily indicate better overall molecular quality.

Two baselines in Table 1 clearly illustrate this trade-off.

  • AMDiff: Although it achieves a better Vina Score than AFIMol (-7.466 vs. -7.241), its QED (0.479 vs. 0.596), SA (0.681 vs. 0.788), and Diversity (0.729 vs. 0.771) are substantially lower.

  • PMDM: It obtains a better Vina Score (-7.572 vs. -7.241) and higher High Affinity (0.628 vs. 0.589), but its SA (0.611 vs. 0.788) and Diversity (0.709 vs. 0.771) are considerably lower. Its QED is also slightly lower (0.594 vs. 0.596).

These comparisons demonstrate that better predicted docking scores do not necessarily translate into better synthetic accessibility or structural diversity.

(2) AFIMol achieves a more balanced multi-objective profile.

AFIMol achieves the best reported SA (0.788) and Diversity (0.771) in Table 1, while maintaining a competitive Vina Score of -7.241 and the second-best High Affinity (0.589) among methods reporting this metric.

Our goal is not to claim superiority in every individual metric, but to avoid sacrificing essential molecular properties for improvements in predicted docking scores. This balance is particularly important for practical drug discovery, where generated candidates must be not only compatible with the target pocket but also chemically tractable and synthetically accessible.

(3) Controlled ablations further support the complementary contributions of our modules.

  • Removing SAF reduces SA from 0.788 to 0.765 and Diversity from 0.771 to 0.753.
  • Removing FIPRA worsens Vina Score from -7.241 to -7.082 and High Affinity from 0.589 to 0.520.

These results support the distinct and complementary roles of adaptive fragmentation and pocket-conditioned interaction guidance.

In summary, AFIMol aims to achieve balanced improvements across complementary molecular-generation objectives, rather than universal dominance in docking scores. (Weakness3)

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.