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

Jian Lee, Changyun Cho, Sun Kim

Oct 30, 2026arXiv:2610.41322v1

cs.LGcs.AI

Scorecard· 16/16

5.5/10 impact

Sensible, timely integration of peptide-reference guidance into a synthesis-aware generator with broad benchmarking, but gains from the guidance itself are small and evidence is purely docking-based with partly circular metrics.

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.

AI Impact Assessments

(1 model)

Core Contribution

PIFT targets small-molecule design against protein–protein interactions (PPIs). It transfers the interaction pattern of a known bound peptide to small molecules, avoiding atom-level conditioning on the peptide, which would over-constrain the search. The method has three parts:

1. A shared, target-conditioned "interface encoder". Peptide and small-molecule interfaces are both converted into typed pharmacophore point clouds (8 ProLIF interaction types), then embedded via invariant message passing with late FiLM conditioning on a target cloud. Training uses contrastive partial-view pretraining followed by ranking-based cross-modal alignment.

2. Target-specific adaptation of a pretrained synthesis-aware generator (SynCoGen) using Discrete Tilt Matching, with a Vina×QEPPI×size reward.

3. Sampling-time guidance with a single embedding-distance objective to the peptide reference. It acts on the discrete branch through GILC-style straight-through Gumbel-Softmax logit correction with Monte Carlo "worlds", and on the continuous branch through gradients on predicted clean coordinates.

The paper also introduces the Interface Agreement (IFA) metric.

The novelty is mostly in the composition and the problem framing. Peptide-to-small-molecule transfer via an abstract interaction space is conceptually sensible, and the joint discrete+continuous guidance in a synthesis-aware model is new. Each component is borrowed (PepMimic-style training, GILC, DTM, SynCoGen, FiLM), so the technical advance is an integration rather than a new primitive.

Methodological Rigor

Strengths:

Splits are by UniProt accession with GRIM-based clustering, which limits target leakage in the encoder.

The guidance ablation (off / coordinate / discrete / joint) holds the prior, seeds and budget fixed.

The comparison covers many baselines across several families (TargetDiff, MolCRAFT, FLOWR, CGFlow, SynCoGen, SynGA, Hot2Mol, Peptide2Mol).

A heavy-atom-count-adjusted docking analysis addresses size confounding.

A reference-specificity test on MDM2 (1YCR vs 3LNJ) is a useful sanity check.

Weaknesses:

Circular metrics. IFA is defined with the same ProLIF interaction profiles that feed the encoder and guidance objective, so the method optimizes something close to the evaluation metric. IFA validation against residue Jaccard (r≈0.99) is also near-tautological, since both derive from the same contact sets.

Vina as the only binding proxy. The prior is adapted using Vina as part of the reward, and Vina is also the headline evaluation. The comparison is therefore not neutral against baselines that were not trained on Vina. A large part of the Vina gain likely comes from DTM adaptation rather than peptide guidance. The paper only isolates guidance vs "Off" (the adapted prior), and the Off Vina is about −7.66 compared with PIFT's −7.70. The headline docking superiority over baselines is thus largely attributable to reward tilting, which is a confound for the paper's framing. Baselines were not given comparable reward adaptation.

Weak guidance effects. The guidance-ablation improvements are small: IFA 1.975→2.049, success 25.3%→29.3%. No significance tests are reported.

Weak encoder evidence. Encoder ablation differences are within noise: pair ranking 0.598±0.042 vs 0.572±0.029. Pair-ranking accuracy of about 0.6 is only modestly above chance. Alignment training uses only 82 queries across 18 targets, a very small supervision set for "cross-modal transfer".

Small, specific evaluation set. There are 10 targets, some of which (MDM2–p53, BCL2, KEAP1, cIAP1) have well-known small-molecule binders in the training data, which risks leakage between DLiP/2P2Idb training and test targets. The split protocol for the generative model targets is unclear.

No experimental validation or MD/FEP. The claims about "discovering PPI inhibitors" rest only on docking.

Unexplained baseline failures. Baseline degeneracies (e.g., FLOWR at −1.25 on BCL2) look more like failures than meaningful comparisons.

Diversity trade-off. Diversity is lowest for PIFT (0.709), which is only lightly discussed.

Potential Impact

If it holds up, the idea of using peptide complexes as abstract interaction references for small-molecule design is relevant to peptidomimetic design, an active area. The plug-in guidance on top of a pretrained synthesis-aware generator is reusable, and a combined discrete+continuous guidance recipe could transfer to other multimodal generative models. IFA may be adopted as a niche metric. Real-world impact is limited without wet-lab validation, and the pipeline is complex, with many hyperparameters and several external components. Code availability is not mentioned.

Timeliness

PPI targeting, synthesis-aware generation, and inference-time guidance of discrete diffusion are all current topics. Peptide2Mol and Pep2Mol are very recent, so the work sits in an emerging niche.

Strengths

A well-motivated problem with a clear abstraction (shared interaction space).

Joint guidance of discrete and continuous modalities with a single objective.

Extensive evaluation breadth with synthesizability and PoseCheck metrics.

Reference-specificity analysis.

Limitations

Mostly in-silico docking evidence, with metric circularity and a confounded attribution of the docking gains.

Small effect sizes for the paper's central component (guidance).

Weak encoder ablation.

A limited number of targets, with no statistics beyond seed variance.

Heavy reliance on prior components, limiting conceptual novelty.

Reproducibility is moderate: hyperparameters are listed, but code and data release are unspecified and the pipeline depends on a dataset merge with filtering.

Overall

A competent, timely applied methods paper with a sensible idea and thorough benchmarking. The core guidance gains are modest, and the evidence is entirely computational, with partially self-referential metrics. Its likely influence is moderate and within a narrow subfield.

Rating:5.5/ 10

Significance 5Rigor 5Novelty 5.5Clarity 6

Generated Nov 3, 2026

Comparison History (0)

No comparisons yet.

Scorecard· 16/16

5.5/10 impact