acceptodds
Under review as a conference paper at ICLR 2027

Sequence-Conditioned Ligand Generation with Compact Protein Prefixes

Abstract

Designing ligands directly from protein sequences could enable the targeting of therapeutically relevant proteins for which well-characterized three-dimensional binding-pocket structures are not available. We develop a flexible autoregressive Transformer that conditions molecular generation on a compact protein prefix as well as separate affinity inputs. We use pretrained ESM3 residue embeddings to provide protein context without training a protein encoder from scratch. After a learned projection, we pool these embeddings over fixed sequence segments to provide regional protein summaries while keeping the conditioning sequence short. To test the contribution of protein context, we replace the protein prefix while keeping the intended target and oracle-based candidate-selection rule fixed. On a selected 42-target panel, correct protein conditioning yields a mean RF3 interface predicted TM-score (ipTM) of 0.920, compared with 0.829 for mismatched conditioning. A separate comparison also shows higher mean interface confidence for generated ligands compared to molecular-weight-matched ligands. Comparisons with PCMol and ProtoBind-Diff show numerically close mean RF3 ipTM scores for selected candidates on their respective panels. Our smaller ligand generator also uses less time and peak allocated GPU memory in cached-feature benchmarks of the tested implementations.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
structure-free ligand generation, target-aware drug design, protein language models, conditional molecular generation, prefix conditioning, SMILES generation, protein–ligand interaction, ESM3, autoregressive transformer, drug discovery

Reviews

Among:
Shape:
1234this paper, Bottom 9%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.50 over 4 reviews, above 9% of the others.
Reviewer_PgEPOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
3
Soundness
2
Read review
Paper summary

This paper presents a sequence-conditioned ligand generator that connects frozen ESM3 residue embeddings to an autoregressive SMILES transformer through a learned projection and mean pooling over 16 contiguous sequence segments. A separate token specifies a requested negative-log affinity proxy. The 6.85M-parameter generator is trained by next-token prediction on approximately 1.84M protein–ligand records partitioned using 30% sequence-identity clusters. Generated candidates are ranked by DeepPurpose AttentiveFP/CNN and adapted ConPLex regressors, then evaluated using RF3 complex confidence. On 42 targets, matched protein prefixes produce higher mean ipTM than mismatched prefixes (0.920 versus 0.829). An additional 82-target comparison places generated ligands between measured target binders and molecular-weight-matched off-target ligands. Comparisons with PCMol and ProtoBind-Diff show similar observed ipTM, higher yields of distinct valid molecules, and 4.3-6.1x faster generation with cached protein features. My reading is that the contribution is a compact implementation of protein-conditioned molecular generation, with computational evidence that the prefix carries useful target information; binding affinity and selectivity of the generated candidates remain unmeasured.

Critical strengths
  • The compact decoder offers a practical advantage for repeated sampling. The reported reductions in generation time and allocated GPU memory are substantial within the cached-feature setting, which could make this approach useful for building larger candidate pools.

  • The higher yield of distinct valid, connected molecules is a meaningful result independent of predicted binding scores. Table 16 reports 90.95% versus 86.10% for PCMol and 88.95% versus 65.33% for ProtoBind-Diff under a common validity procedure.

Critical weaknesses
  • There aren't very many methodological advancements here and also incomplete comparison with closely related work. Frozen protein embeddings, projection, segment averaging, and a standard next-token objective form a pretty standard architecture. Prot2Drug (Creanza et al., 2025) already connects pretrained protein representations to a two-layer transformer decoder, while DrugGPT and ZeroGEN are also missing from the discussion and experiments. The proposed regional pooling could still be useful, but Table 7 changes affinity fusion and training alongside pooling, and both confidence intervals include zero. This does not establish what the 16-token representation contributes.

  • Unfortunatley, the authors don't establish strong ligand design. The final affinity oracles achieve test Spearman correlations of only 0.395 and 0.321, and their average selects the candidates evaluated downstream. RF3 provides a separate structural prediction, but neither ipTM nor contact-restricted PAE establishes binding or selectivity, and the off-target controls are not verified non-binders. The docking results provide limited support, as native-ligand redocking succeeds within 2 Å on only 3/12 targets. I would need stronger validation against measured binding data to judge whether the computational improvements translate into useful candidates.

  • Generalization and comparative performance are difficult for me to assess. Appendix B.2 states that the nominal test split informed model development, and the 42-target prefix panel comes from targets selected in earlier candidate screens. The Pocket2Mol/TargetDiff panel contains nine training and three validation targets. These choices substantially weaken the evidence on unseen targets. Moreover, Table 2 reports ipTM differences of -0.0108 against PCMol and +0.0083 against ProtoBind-Diff, with both intervals crossing zero.

Questions for authors
  1. The authors should compare against Prot2Drug (Creanza et al., 2025), DrugGPT (Li et al., 2023), and ZeroGEN (Chen et al., 2025) on a common target panel. I'd prioritize Prot2Drug because its pretrained protein representation plus small decoder directly tests the proposed contribution. Please match generation and selection budgets, and document training overlap.

  2. Can the authors provide an explanaiton for benefit comes specifically from contiguous regional pooling in Eq. 2? Please compare fully trained models using global mean pooling, 16 learned queries, and 16 randomly assigned residue groups, holding ESM3 features, decoder size, affinity conditioning, and training budget fixed.

  3. Which decisions used the nominal test split? Please evaluate on protein clusters unused during development, with targets selected before generating candidates, and report nearest training-protein similarity and generated-ligand similarity to the full training set, including scaffold overlap. To me, exact-SMILES novelty alone does not establish substantial chemical novelty.

  4. How well does the ensemble average of the DeepPurpose and ConPLex predictions rank measured ligands within each held-out protein? Please report within-target correlations and top-ranked enrichment, then compare ensemble, single-oracle, and random selection from identical candidate pools.

  5. How should a user interpret the requested affinity proxy, a=−log⁡10(affinity in M)a=-\log_{10}(\mathrm{affinity\ in\ M}), when assay endpoint types were discarded? Table 13 gives a fitted response of only 0.0749 predicted pAff per requested unit, and Table 14 shows no clear RF3 benefit. Please evaluate an endpoint-consistent subset and a model trained without the affinity token.

  6. Can RF3 ipTM distinguish experimentally measured actives from measured inactives on unseen targets? Please test this with chemically matched controls and report discrimination and enrichment to help me see this.

  7. Could you repeat timing with models already resident on the GPU, comparable precision, and consistent output-length limits? ProtoBind-Diff currently incurs Trainer creation and device placement inside the timer. If you can report distinct valid molecules per second and separately include protein-feature extraction and downstream scoring, it would show the practical cost of the full workflow.

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

The paper presents a sequence-conditioned ligand generator. Frozen ESM3 embeddings from the full protein sequence are divided into 16 contiguous segments and mean-pooled into 16 protein tokens. An additional affinity token conditions a small causal Transformer that generates SMILES. Candidates are ranked by two affinity predictors and evaluated mainly with RF3 ipTM.

The experiments include matched versus mismatched protein prefixes, a three-arm comparison, comparisons with PCMol and ProtoBind-Diff, generation efficiency, affinity control, and a limited docking analysis.

Critical strengths
  • The sequence-only setting is clear and practically relevant, especially when reliable pocket structures are unavailable.
  • The model is lightweight and computationally efficient at the generation stage.
Critical weaknesses

Binding Effectiveness Is Not Established

No generated ligand is experimentally validated. The main conclusions rely on RF3 ipTM and two affinity predictors. ipTM measures confidence in a predicted interface, not binding affinity or free energy. The two predictors achieve Spearman correlations of only 0.395 and 0.321 on the protein-cluster test set, and their training labels mix different assay endpoints. The off-target ligands in the three-arm experiment are also not confirmed non-binders.

I recognize that experimental validation is difficult. At minimum, the paper should report standard SBDD metrics such as docking score, QED, SA, pbvalid, diversity, and novelty. It should also measure the correlation between RF3 ipTM and experimental affinity on an appropriate training or validation subset rather than treating ipTM as a direct proxy for binding.

Incomplete Related Work and Baselines

  • Additional protein-sequence-conditioned baselines are needed.
  • A more practical baseline is an end-to-end pipeline of protein structure prediction, blind pocket detection, and modern structure-based ligand generation.

Unreliable Docking Evaluation

Appendix G.3 evaluates only 12 targets, whereas the standard CrossDocked2020 protocol contains 100 pockets. Only 5/12 targets for the proposed method and 1/12 for TargetDiff completed docking for all 100 molecules. Native-ligand redocking achieved RMSD < 2 Å for only 3/12 targets... These results suggest that the evaluation pipeline was not sufficiently validated and cannot support model comparisons.

Limited Evidence for Affinity Control and Efficiency

  • Increasing the requested affinity produces only a small gain in oracle score while substantially increasing molecular weight. Masking or permuting the affinity input also does not consistently improve RF3 results. The current experiments therefore do not establish controllable affinity generation.
  • The efficiency comparison measures only generation with cached protein features. It excludes feature extraction, model loading, oracle ranking, and RF3 evaluation, and the baseline implementations do not use matched precision or process lifetimes. This result should not be presented as an end-to-end efficiency advantage.
Questions for authors
  1. The model pools the full protein sequence. Does it learn to identify binding pockets? Please consider residue- or segment-level masking, permutation, or attribution analyses on proteins with known pockets, including hidden- or multi-pocket cases.
  2. What is the advantage of 16 fixed segments over global pooling, other values of K, attention or BoM-style pooling, and residue-level conditioning?
  3. Does RF3 ipTM correlate with experimental affinity? Please report this relationship on a training, validation, or independent structure-labeled subset and test whether ipTM is biased by ligand size or training-set chemical space.
Reviewer_gfAVOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The authors present a hybrid architecture for the generation of SMILES conditioned on a given input protein sequence. The input sequence is passed to frozen ESM3 (sequence only), which then passes to a decoder-only architecture in the form of MolGPT after pooling into a lower-dimensional space. The decoder has around 6M parameters, making the architecture rather light, as appreciated by the low memory usage and the fast inference. They report 89% validity, with 99% of the generated molecules absent from the training set, which is also a positive outcome. The benchmark relies extensively in RFD3 as a proxy, reporting comparisons in ipTM for ground truth versus decoy sets. While the idea is well-exposed, the current benchmark does not establish that the generated molecules bind their intended targets, and it the model is insufficiently compared against other baselines and proper ablation studies. Gnerally speaking, my other major concern is the novely of the approach, as all methodology in the manuscript comes from previously established (some a few years ago) architectures and methodologies.

Critical strengths
  • The small architecture has practical value for the generation of SMILES. With a very small architecture, the (6M params, excluding the frozen protein encoder) it achieves numerically similar RF3 interface confidence to substantially larger generators and provides similar values for generated SMILES and ground-truth. This provides results for that (very) specific benchmark with faster inference times and fewer computational needs.

The matched vs. mismatched protein control is a useful experiment. Both conditions undergo the same candidate selection and RF3 evaluation, with matched inputs scoring higher on 39/42 targets. This supports the conclusion that the compact prefix retains information relevant to target-dependent generation.

Critical weaknesses

Limited methodological novelty: The approach combines frozen ESM-3 embeddings, fixed segment pooling, and a standard autoregressive SMILES decoder-only model, inthe fashion of MolGPT. While this is a reasonable design, I do not see a sufficiently clear technical advance over existing protein-conditioned molecular generators. The architectures themselves have been published for several years, with ESM3 being the newest (2024) (which is taken with frozen weights). The main proposed contribution (the protein prefix) is not supported by controlled ablations establishing its advantage over simpler pooling schemes. A smaller generator could have practical value, but the current experiments do not convincingly establish what is gained through the proposed design (see below).

The computational and experimental validation is insufficient: Although comparisons with PCMol and ProtoBind-Diff are included, the evaluation relies heavily on predicted affinity and RF3 interface confidence, without establishing how these translate into successful binding. There is for example no docking performed with other tools, such as GNINA, GLIDE, boltz2. In addition, broader comparisons with existing generators (DrugFlow, chemical language models) and evaluation on targets unused during model development are needed. While harder to accomplish, there is no experimental validation of generated compounds with the targets (e.g SPR, NMR-1D..). Without such evidence, a stronger independent computational benchmark would be needed to support the claimed utility for ligand design.

Questions for authors

No questions would materially affect my assessment.

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

This paper proposes a compact autoregressive decoder that generates ligand SMILES from a protein sequence, without needing a binding pocket structure. Frozen ESM3 residue features are projected and mean-pooled into 16 segment tokens, which together with one affinity token form a fixed-length conditioning prefix instead of a variable-length prefix based on sequence length. The paper shows that ligands generated with a target's own prefix score higher on affinity oracles and RF3 interface confidence than those generated with another protein's prefix. It also reports RF3 confidence comparable to PCMol and ProtoBind-Diff with a much smaller generator and lower generation time and memory.

Critical strengths
  1. The sequence-only setting is less explored than pocket-conditioned ligand generation, and it matters for targets without a solved pocket structure.
  2. The three-arm comparison in Section 3.3 is well designed. The protein sequence, the MSA and the ligand molecular weight are all held fixed, so only the ligand identity changes.
Critical weaknesses
  1. Very limited comparison to baselines. Only PCMol and ProtoBind-Diff are actually run, and the model loses to PCMol. Peer-reviewed sequence-only generators with public checkpoints are neither compared nor cited. See questions for detail.

  2. Limited ablation study. Neither design choice, the compact prefix and the affinity input, is supported by the authors' own controls in Section B.3 and Appendix D.3. See questions for detail.

  3. Nothing in the evaluation measures binding, and no panel is held out. Selection uses two in-house oracles with cluster-split Spearman 0.395 and 0.321, the test split was consulted during development, and the 12 targets in Appendix G are all training or validation targets.

Questions for authors
  1. Two published journal papers that tackle protein sequence conditioned ligand SMILES generation are completely ignored and not benchmarked in this paper, including: Uludoğan et al. (Bioinformatics, 2022) and Chen & Bajorath (J. Cheminformatics, 2024). Both released their trained weights.

  2. Pocket2Mol and TargetDiff are evaluated in Appendix G, so structure-based generators are clearly in scope. Then why not benchmark stronger ones such as UniMoMo and MolCRAFT, both of which outrank Pocket2Mol and TargetDiff on CBGBench? Also, if this comparison is made, why not exclude the targets that are in the proposed model's training set? All 12 targets in Appendix G are training or validation targets, 9 and 3 respectively.

  3. Can the model generate ligands with both conditioning inputs removed or masked (unconditional generation)? Section 3.2 shows the prefix changes the output, but the effect of conditioning over no-condition is not shown at all. Since the oracle selection is target-aware in both arms, an unconditional arm, or just 25 random molecules from the existing test pool run through the same selection, would separate the prefix from the selection.

  4. The ablation on the prefix size is not conclusive. Section B.2 compares K=1,16,64K = {1, 16, 64} only in reduced-budget pilots with no numbers reported, and Section B.3 compares K16 against a predecessor that also differs in affinity fusion and training implementation. Can the authors report K=1,16,64K = {1, 16, 64} at the full training budget, and add a run with uncompressed per-residue conditioning?

  5. Generation of tokenized SMILES may result in invalid molecules. Why not use a molecular sequence representation that guarantees validity, such as SELFIES?

  6. The mismatched arm in Section 3.2 draws all 25 generations from a single donor protein (Appendix D.1), and the bootstrap resamples targets but not the donor assignment. This also favors the matched arm, since one donor gives a narrower candidate pool and the oracle takes the best of 25. Can the authors redo this with the 25 mismatched candidates drawn from 25 different donors?

  7. In Equation 3, masking the affinity request leaves the affinity token equal to the bias term bab_{a}, and Appendix D.3 states this is the same input as a request at the training-label mean. The masking control therefore does not test removal of the affinity token. Why was it designed this way instead of zeroing the projected output or dropping the bias term? Also, was m=0m = 0 ever used during training, or only at inference time for this control?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.