acceptodds
Under review as a conference paper at ICLR 2027

Transferring Peptide–Target Interaction Patterns to Small Molecules by Joint Guidance for Diffusion and Flow Matching

Abstract

Generating small molecules to target protein-protein interactions (PPIs) remains one of the most challenging yet high-reward frontiers in drug discovery. Unlike traditional small-molecule drugs that bind to deep, well-defined pockets, PPI interfaces often comprise broad, shallow binding surfaces. A natural strategy is to leverage known peptide–target interactions, as peptide ligands provide rich interface-level constraints. However, a major algorithmic challenge is that rigid atom-level conditioning on peptide structures can over-constrain the generative space and exclude valid small-molecule solutions. To address this challenge, we present PIFT (Peptide InterFace Transfer), a novel framework that transfers peptide–target interaction patterns to small molecules through joint sampling-time guidance of masked discrete diffusion and continuous flow matching. PIFT first maps peptide–target and small-molecule–target interfaces into a shared, target-conditioned interaction space, capturing transferable interface information across molecular modalities. The resulting peptide-derived reference then jointly steers building-block logits and 3D coordinates toward peptide-compatible molecules while preserving a strong adapted molecular prior. Extensive experiments across diverse PPI targets show that the learned representation captures cross-modal interaction compatibility, while peptide guidance improves interface recovery and achieves strong docking performance and favorable molecular properties.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
drug discovery, peptide, small molecule, generation, diffusion, masked discrete diffusion, flow matching, guidance

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_h4noOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This work proposes PIFT (Peptide InterFace Transfer), a framework that transfers peptide–target interaction patterns to small molecules through joint sampling-time guidance of masked discrete diffusion and continuous flow matching.

Critical strengths
  1. The discrete chemical space and continuous geometric space are guided by the same interface objective, resulting in a relatively complete and coherent design.
  2. The baselines cover standard 3D generation, synthesis-aware generation, and PPI/peptide-guided generation, providing a relatively broad comparison.
Critical weaknesses
  1. It is difficult to clearly separate the Vina advantage of PIFT from the effect of target-specific prior adaptation.
  2. The features used to train and guide the model are closely related to those used in the main interface-recovery metric, which raises some concern about the independence of IFA as an evaluation metric.
  3. PPI interfaces are highly heterogeneous, and evaluation on only 10 targets may be insufficient to support a broad claim about PPI inhibitor generation.
  4. The writing has noticeable clarity issues. The paragraph organization is rather fragmented, which makes the overall presentation difficult to follow. A number of symbols used in the equations are not explicitly defined.
Questions for authors

How much of the performance improvement of PIFT comes from peptide guidance rather than DTM-based target-specific prior adaptation? Is it fair to use Vina as a major benchmark metric after using a Vina-based reward for target-specific adaptation? The improvement from peptide guidance over the Off setting is relatively limited. Is this sufficient to support “interaction pattern transfer” as the main contribution? The method is evaluated on only 10 PPI targets. How well does it generalize to substantially different interface types, new target families, or cases without a high-quality reference peptide?

Reviewer_S2ZqOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes PIFT, a framework for transferring peptide–target interaction patterns to small-molecule generation for protein–protein interaction (PPI) targets. It learns a shared, target-conditioned representation of peptide and small-molecule interfaces using pharmacophore features and geometric information. Building on a synthesis-aware molecular generator adapted through Discrete Tilt Matching, PIFT jointly guides discrete building-block selection and continuous 3D coordinates during sampling using peptide-derived interface objectives. Experiments across ten PPI targets demonstrate improved interface agreement and competitive docking performance compared with existing generative baselines, while maintaining reasonable molecular validity, diversity, and predicted synthesizability.

Critical strengths

S1. The paper addresses a valuable and challenging molecular generation problem. S2. The cross-modal interaction abstraction is intuitive and meaningful. S3. Joint guidance of discrete chemistry and continuous geometry is an interesting design.

Critical weaknesses

W1. To my understanding, the core idea of transferring peptide interaction patterns is closely related to existing pharmacophore-based approaches. PIFT primarily integrates established techniques, including SynCoGen, DTM, and GILC-style guidance. Moreover, it lacks comparisons with simpler alternatives, such as directly guiding the same generative backbone using a differentiable pharmacophore-matching loss without the learned cross-modal encoder. W2.The learned interface embeddings show weak correlations with target-residue overlap and IFA (Spearman ρ = −0.19 and −0.18, Figure 3). Moreover, the proposed encoder achieves only 59.8% pair-ranking accuracy, marginally outperforming the target-agnostic baseline (57.2%, Table 7). I wonder whether the learned representation effectively captures transferable peptide–small-molecule interaction patterns.

Questions for authors

Q1. Figure 4(b) shows relatively modest improvements from joint guidance over the adapted prior. Could the authors provide some additional discussion on the contribution of peptide-reference guidance and whether its effectiveness varies across different PPI targets? Q2. Table 7 reports a pair-ranking accuracy of 59.8% for the proposed encoder. Could the authors elaborate on how this representation facilitates peptide-to-small-molecule interaction transfer, despite the relatively modest ranking performance? Q3. The proposed joint guidance involves additional gradient computations during sampling. Could the authors discuss the computational overhead compared with sampling without peptide guidance?

Reviewer_sdu2OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

PIFT uses a bound peptide as a reference for small-molecule design at protein–protein interaction interfaces. A shared encoder maps peptide and small-molecule pharmacophore clouds into a target-conditioned embedding space. A SynCoGen backbone is first adapted to each target using a reward based on Vina, QEPPI, and molecular size. At sampling time, distance to the reference-peptide embedding guides both discrete building-block choices and continuous coordinates. Experiments cover ten PPI targets, external generators, guidance ablations, interface similarity, molecular quality, and a two-reference MDM2 example.

The scientific motivation is strong. Transferring a pattern of useful contacts without copying peptide atoms is an attractive way to address the mismatch between peptides and small molecules. The main question is whether the learned representation actually identifies transferable binding determinants, rather than mainly reproducing the contact rules used to train and evaluate it.

Critical strengths
  1. The problem formulation is useful and well motivated. PPI interfaces are difficult targets, and a peptide can provide information that receptor geometry alone does not specify. An interface-level representation is a plausible alternative to rigid structural imitation.
  2. The paper includes meaningful internal controls. Comparing guidance with the same adapted prior, separating discrete and coordinate guidance, switching the peptide reference on MDM2, and examining size-adjusted docking are all valuable. These controls should be retained and made more quantitative.
  3. The aggregate computational results are encouraging. PIFT reports the best aggregate Vina and IFA among the compared methods. Table 8 shows the best docking mean on eight of ten targets. The synthesis-aware backbone also provides a practical constraint beyond unconstrained atom generation.
  4. The data split is more carefully considered than a random complex split. Appendix A.1.1 uses accession-disjoint partitions and additionally groups accessions by pharmacophore similarity. This is a useful attempt to reduce target leakage.
Critical weaknesses

1. The validation mostly measures recovery of the same contact rules used for supervision

Phase-II positives and negatives are defined by contact and interaction-type agreement (Appendix A.1.1). The representation is then evaluated against residue overlap and IFA, which are built from the same general contact information. The high correlations in Appendix C.1 therefore establish consistency among related geometric descriptors; they do not independently establish preservation of binding determinants or PPI inhibition.

This distinction matters for the central claim. Two molecules can make similar contacts while differing greatly in affinity because of desolvation, strain, conformational entropy, or the energetic importance of particular residues. Conversely, a useful small molecule may reproduce a small subset of peptide hotspots rather than a large fraction of its contacts.

Please add a direct, nonlearned pharmacophore/contact-matching guidance baseline under the same adapted prior and sampling budget. This would test whether the learned embedding contributes beyond its input representation. An independent evaluation using known active/inactive compounds, experimentally characterized hotspots, or held-out pose discrimination would make the biological interpretation substantially stronger. Experimental assays would be valuable, but my concern is the lack of independent validation, not a blanket requirement for wet-lab work.

The IFA definition itself needs clarification. It sums residue-wise cosine similarities over the union of contacted residues, including cases where one vector is zero; cosine is undefined there unless a convention is specified. With the natural zero-contribution convention, additional contacts to residues absent from the reference are not penalized. Moreover, when both vectors contain only the same single interaction type, their cosine is one regardless of the distance weights. Thus, IFA is neither a normalized interface similarity nor generally a measure of distance agreement. Reporting precision is helpful, but the paper should explicitly characterize these limitations, compare normalized alternatives, and test sensitivity to contact thresholds. A common IFA threshold of 1.5 can also represent different levels of recovery for peptides with different interface sizes.

2. The central differentiable interface construction is not sufficiently specified

Equations (12)–(13) write the pharmacophore cloud formula, but the manuscript does not define how Φ\Phi supports gradients through changes in building-block identity. Straight-through Gumbel–Softmax supplies a surrogate derivative for a categorical choice; it does not automatically make reaction assembly, atom indexing, pharmacophore typing, or ProLIF contact detection differentiable.

Different building blocks can contain different numbers of atoms and different reactive centers. Please specify how their atoms and coordinates share a tensor representation, how reaction validity and feature masks are handled, and what continuous relaxation carries gradients from interface features back to logits. For coordinate guidance, explain whether contacts and feature identities are frozen during each update or recomputed through a smooth surrogate. Hard contact thresholds can yield a piecewise-defined objective, which is not necessarily unusable, but its behavior needs to be described.

The geometric features ρi\rho_i and ρij\rho_{ij} also need explicit formulas. For fixed point types and masks, separate encoders using only within-cloud invariant geometry would be insensitive to rigidly moving the binder relative to the target. The target-dependent contact extraction can already encode some relative information by changing which features are present, so the complete pipeline need not have this limitation. Please explain whether pose sensitivity comes from that extraction, explicit cross-cloud geometric features, or both, and how it supplies coordinate gradients. A displaced/rotated-binder control with contacts recomputed would be informative. This is a request to clarify the pipeline, not a claim that it ignores the target.

These details are central to the claimed contribution. A precise algorithm, tensor dimensions, and a small gradient sanity check would be more helpful than additional high-level diagrams.

3. Most of the docking improvement cannot currently be attributed to peptide guidance

PIFT uses a target-specific prior optimized with a reward containing Vina and QEPPI, while the headline comparison is largely against generators without the same adaptation. This is a reasonable system comparison, but it does not isolate the contribution of the peptide representation or joint guidance. Optimizing the reported docking function is also a different setting from generalizing to an independent binding evaluator.

The controlled Off comparison is therefore especially important. Figure 4b reports:

MetricOffJoint guidanceChange
IFA1.9752.049+0.074, approximately 3.75%
Contact precision0.4480.460+0.012
Vina−7.662−7.703−0.041 kcal/mol
Joint success25.3%29.3%+4.0 percentage points

These are potentially useful improvements, but much smaller than the full-system comparison with SynCoGen. Coordinate-only guidance also gives a slightly better docking mean than joint guidance. The results suggest a modest multi-objective benefit; they do not yet justify attributing the broad docking advantage to peptide transfer.

Please report the full target-level matrix for the original backbone, adapted Off prior, direct-contact guidance, learned guidance, and each guidance branch. Include paired differences and confidence intervals across targets and independent runs. Appendix C.4 already provides a useful size-regression control, but this controls molecular size, not adaptation budget, reward optimization, or the value of the learned representation.

The MDM2 reference-switch experiment is a good control, although the displayed self/cross IFA gains are small: approximately 0.019/0.005 and 0.021/0.009 in Figure 6a. Without uncertainty and additional targets, “clear reference-dependent guidance” is stronger than the evidence currently warrants.

4. The amount of effective cross-modal supervision and its generalization need closer examination

The initial corpus is substantial, but Appendix A.1.1 reports only 82 eligible peptide queries across 18 targets for target-conditioned alignment. This is the relevant supervision scale for the central cross-modal stage, and it should be visible in the main text.

Please report the number of held-out queries, targets, and pairs, the contribution of each structural source, and the exact relationship between the ten generation targets and every training stage. Accession-disjoint encoder splits do not automatically establish that the evaluation targets are unseen by the generative backbone. Conversely, target-specific reward adaptation is intentional and should be described as adaptation, rather than conflated with zero-shot generalization.

The Phase-II labels are also coarse: a positive needs nonzero agreement, whereas a negative has zero agreement and a nearby pocket centroid. This can teach the encoder to distinguish some overlap from none without learning how to prioritize energetically important interactions. More challenging negatives with similar contact counts, matched molecular size, and perturbed poses would test the intended capability.

Table 7 gives a pair-ranking accuracy of 0.598 ± 0.042, compared with 0.572 ± 0.029 without target conditioning. This is encouraging but modest. Please define what the uncertainties represent and report a paired comparison. The strong pooled trends in Figure 3 should not replace this harder query-level generalization result.

5. Evaluation budgets, failures, and molecular trade-offs need a more complete account

Table 6 specifies 256 samples per target and five docking seeds, whereas Figure 4c extends beyond 50,000 sampling steps/candidates on its budget axis. Please explain whether that figure pools targets, uses a separate large experiment, counts proposals or accepted molecules, and includes the samples used for prior adaptation. The “approximate 95% interval” needs a defined resampling unit and construction. Because discovery modes are accepted sequentially by Tanimoto distance, ordering can change the count; random-order repetitions or a fixed clustering procedure would help.

The baseline failures also need a concrete diagnosis. Peptide2Mol has validity 0.088 and several missing targets, while FLOWR has a degenerate BCL2 result. Such results may be real, but the general explanation that diffusion and flow matching require a localized cavity is too broad—PIFT itself uses both mechanisms, and the paper describes baselines that can accept peptide or interaction information. Report versions, conditioning setup, docking boxes, validity filters, and failure counts, with suitable positive controls for the baseline implementations.

Missing-target handling changes the interpretation of averages. For example, TargetDiff's −6.27 average is consistent with averaging its eight successful targets, while PIFT's −7.68 uses all ten. Report conditional docking means alongside unconditional valid/docked yield and common-target comparisons. Restricting PoseCheck to complete methods is transparent, but the other aggregate metrics need equally explicit denominators.

Finally, Table 3 shows a meaningful trade-off: PIFT's diversity is 0.709, the lowest reported, compared with 0.804 for SynCoGen. Its AiZynthFinder success rate is 0.612 versus 0.643, and validity is 0.947 versus 0.957. These do not negate the docking gains, but they qualify the claim of broadly favorable molecular properties. PoseCheck interaction counts alone do not establish pose plausibility; report clashes and strain as well. Successful route prediction or construction from reaction templates should be described as predicted synthetic feasibility.

Questions for authors
  1. How is the pharmacophore construction differentiated through variable building-block identities and reaction assembly, and what target-relative geometry enters the encoder?
  2. Does learned guidance outperform direct contact/pharmacophore guidance with the same adapted prior and total computational budget?
  3. Are the gains over Off consistent across targets and independent runs, and do they persist under an independent evaluator?
  4. How many held-out queries support the cross-modal claim, and how do the ten generation targets relate to all training and adaptation datasets?
1 comment

We sincerely thank the reviewer for their time and careful assessment of our results. Although some experiments are still ongoing, we are continuing this work based on the reviewer’s suggestions for further evaluation. Below, we report additional analyses addressing the limitations of IFA.

W1. Clarifying and Complementing IFA

We agree that the current IFA definition has limitations. Although residue-wise cosine similarity captures differences in interaction-type profiles, it is insensitive to distance weights when both profiles contain only the same single interaction type. With zero-vector contributions defined as zero, contacts to residues absent from the peptide reference do not reduce the summed IFA.

To address the latter limitation, we additionally report normalized IFA and interaction-pair precision. Normalized IFA divides the summed score by the number of residues contacted by either binder. Interaction-pair precision measures the fraction of ligand (residue, interaction-type) pairs also present in the peptide reference. We include recall and F1 to characterize the balance between interaction recovery and additional contacts.

ModelIFANormalized IFAInteraction-pair precisionInteraction-pair recallInteraction-pair F1
Ours2.0400.14650.4620.1850.239
Off2.0120.14370.4570.1830.236
FLOWR1.6400.12220.4420.1460.196
MolCRAFT1.6010.11410.4610.1380.195
SynGA1.5210.11280.4720.1370.191
SynCoGen1.4560.10610.4480.1290.179
Hot2Mol1.4320.10450.4570.1220.177
TargetDiff1.4140.10150.4940.1240.181
CGFlow-QED1.2930.10050.4880.1200.174
Peptide2Mol0.9910.08940.6560.1030.162

Our method has the highest normalized IFA, recall, and F1, though the gains over guidance-off sampling are small. Its precision is lower than that of several baselines, suggesting that it recovers more reference interactions but also makes additional contacts. The added metrics account for these contacts, while IFA’s limited sensitivity to distance differences remains.

We are also considering how to distinguish additional target interactions arising from small-molecule chemistry that may support binding from those that are unnecessary, and how to assess their compatibility with the peptide reference. Developing a concrete metric that captures this distinction remains an important direction for our work. We thank the reviewer for raising this question and helping us identify this direction for improvement.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Contribution of the Learned Representation

We thank all reviewers for their constructive feedback and suggestions to improve our work. We appreciate their recognition of the challenges of targeting PPI interfaces and the motivation for transferring peptide interaction patterns to small molecules.

As all reviewers highlighted the need to examine whether the learned embedding captures useful interaction features, we conducted additional analyses using harder negative pairs and a direct, nonlearned pharmacophore-guidance baseline.

Sensitivity to Negative Pairs

We evaluated whether the encoder distinguishes compatible pairs from negative pairs when contact count and molecular size are controlled. Across all four conditions, compatible pairs had smaller embedding distances in 32–33 of 40 comparisons. This separation persisted when both contact count and molecular size were matched, suggesting that the representation captures differences beyond these two factors.

Negative-pair selectionMedian dposd_{\mathrm{pos}}Median dnegd_{\mathrm{neg}}dneg/dposd_{\mathrm{neg}}/d_{\mathrm{pos}}dpos<dnegd_{\mathrm{pos}} < d_{\mathrm{neg}}Relative margin (macro)One-sided pp-value
Unmatched0.1370.3442.01×32/40+0.742 [0.055, 1.599]2.9×10−62.9 \times 10^{-6}
Contact-count matched (±1)0.1370.3001.71×33/40+0.757 [0.176, 1.490]1.2×10−61.2 \times 10^{-6}
Molecular-size matched (±3)0.1370.3632.03×32/40+0.839 [0.209, 1.550]3.6×10−73.6 \times 10^{-7}
Both matched0.1370.3181.74×33/40+0.762 [0.241, 1.344]2.6×10−72.6 \times 10^{-7}

Comparison with Direct Pharmacophore Guidance

For BCL2–BAX, we compared learned guidance with a nonlearned baseline that matches pharmacophore geometry directly. The baseline constructs distance histograms for pharmacophore-type pairs, such as donor–donor and donor–hydrophobic, and uses the difference between the peptide and ligand histograms as its guidance loss.

Learned guidance achieved higher IFA, normalized IFA, and interaction-pair precision, recall, and F1 than both unguided sampling and direct histogram guidance, with a similar mean Vina score. Validity was lower with learned guidance. These single-target results provide preliminary evidence that the learned representation offers an advantage over the tested direct-matching baseline.

MethodnnValidityVina (kcal/mol)IFANormalized IFAInteraction-pair precisionInteraction-pair recallInteraction-pair F1
Guidance off2090.852−8.2961.49340.08680.34160.11340.1672
Direct histogram guidance2080.848−8.2811.49450.08640.33060.11200.1644
Learned guidance (ours)2070.828−8.2951.61010.09400.34540.11690.1709

Although our experiments indicate that the interface encoder can distinguish compatible from mismatched pairs, the reviewers appropriately note that its ranking performance remains modest. We therefore view these results as encouraging but preliminary, and agree that further experiment is needed to establish the model’s biological relevance.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.