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

Jian Lee; Changyun Cho; Sun Kim

cs.LGOct 30, 2026

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

Authors: Jian Lee, Changyun Cho, Sun Kim

Organizations: AIGENDRUG Co., Ltd., Seoul · Interdisciplinary Program in Artificial Intelligence, Seoul National University · Department of Computer Science and Engineering, Seoul National University · Interdisciplinary Program in Bioinformatics, Seoul National University

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.

Figures & tables

Figure 1: Overview of PIFT molecular generation framework. The interface encoder extracts representations from a reference peptide–target interface and an intermediate molecule–target state. The guidance objective derived from their embeddings jointly steers discrete building-block logits and continuous 3D coordinates during sampling.

Figure 2: Branch-specific mechanism of peptide-reference guidance.

Figure 3: Peptide–molecule embedding distance across target-residue overlap and IFA. Left: Embedding distance decreases as shared contacted target residues increase, clearly distinguishing overlap levels. Right: Higher IFA shows the same trend. Lines indicate the median and shaded regions the interquartile range (Q1–Q3).

Panel of Figure 4 — Figure 4: Peptide-reference joint guidance. (a) PIFT and Off populations in IFA– $-$ Vina space. Shared density regions show that guidance largely preserves the distribution supported by the DTM-adapted prior, while guided contours extend toward higher interface agreement. Full results for the 10 PPI targets are provided in Appendix C.4 . (b) Joint guidance ablation. Precision measures agreement of generated and reference $(\text{residue},\text{interaction type})$ contacts; success requires both IFA above and Vina below the target-specific Off median. (c) Fixed-budget discovery. The curve reports the cumulative number of additional qualified modes discovered by PIFT relative to Off; the line is the median and shading the approximate 95% interval. (d) 3D illustration of BCL2–BAX reference peptide, Venetoclax, and guided molecules Generated molecules demonstrate chemically distinct realizations of the reference interface.

		Average Vina Docking Score (kcal/mol, $\downarrow$ )
Category	Model	BRD2–H4	BCL2–BAX	Calpain–Calpastatin	cIAP1–SMAC	DAB1–ApoER2	Avg.	Med.
Autoregressive	Hot2Mol	-7.00 ( $\pm$ 0.04)	-7.61 ( $\pm$ 0.05)	-6.56 ( $\pm$ 0.03)	-6.59 ( $\pm$ 0.05)	-7.35 ( $\pm$ 0.02)	-7.01	-7.06
Search-based	SynGA	-6.47 ( $\pm$ 0.09)	-7.33 ( $\pm$ 0.05)	-6.50 ( $\pm$ 0.03)	-6.04 ( $\pm$ 0.09)	-6.78 ( $\pm$ 0.03)	-6.62	-6.64
GFlowNet-based	CGFlow	-7.20 ( $\pm$ 0.05)	-7.93 ( $\pm$ 0.03)	-6.91 ( $\pm$ 0.03)	-6.38 ( $\pm$ 0.03)	-6.99 ( $\pm$ 0.05)	-7.03	-7.10
Diffusion / Flow-based	TargetDiff	-5.30 ( $\pm$ 0.04)	–	-6.74 ( $\pm$ 0.03)	-3.50 ( $\pm$ 0.03)	-6.02 ( $\pm$ 0.08)	-6.27	-6.82
	MolCRAFT	-5.02 ( $\pm$ 0.02)	-6.84 ( $\pm$ 0.03)	-7.50 ( $\pm$ 0.04)	-3.90 ( $\pm$ 0.04)	-5.32 ( $\pm$ 0.01)	-5.81	-6.25

Table 2: Vina docking performance. Average Vina docking score for each targets, and the average and median values across 10 PPI targets. Results for remaining 5 targets are reported in Appendix C.4 . Best results are in bold.

Category	Model	IFA $\uparrow$	Validity $\uparrow$	AiZynth. $\uparrow$	Uniqueness $\uparrow$	Diversity $\uparrow$	PoseCheck $\uparrow$
Autoregressive	Hot2Mol	1.347	0.975	0.529	1.000	0.819	10.485
Search-based	SynGA	1.329	1.000	0.258	1.000	0.878	11.071
GFlowNet-based	CGFlow	1.182	0.978	0.575	0.972	0.828	9.193
Diffusion / Flow-based	TargetDiff	1.227	0.571	0.106	1.000	0.891	—
	MolCRAFT	1.420	0.864	0.226	0.976	0.854	11.340
	Flowr	1.290	0.909	0.023	1.000	0.842	—

Table 3: Molecular-generation quality. Best results are in bold.

Appendix figures & tables10 assets

Supplementary material from the paper’s appendix.

Appendix

Figure 5: Target-conditioned Cross-Modal Interface encoder architecture.

Stage I: modality-shared geometric pretraining
Hyperparameters	Values
Interface encoder hidden dimension	128
Interface embedding dimension	64
Target encoder	hidden dim. 64, output dim. 32
Epochs	30
Steps per epoch	2,000

Appendix

Table 4: Default hyperparameters.

Hyperparameters	Discrete branch	Coordinate branch
Guidance start $t_{\mathrm{start}}$	0.5	0.5
Schedule exponent $p$	2.0	2.0
Gradient clipping	5.0	5.0
Guidance scale $\gamma_{\max}$	0.1	0.1
Margin fraction	1.2	—
ST-Gumbel MC worlds $M$	4	—

Appendix

Table 5: Default hyperparameters.

Hyperparameters	Values
Reverse sampling steps	100
Batch size	64
Samples per target	256
Docking	UniDock
Docking seeds	5

Appendix

Table 6: Default hyperparameters.

Encoder variant	Pair ranking $\uparrow$	Pair margin $\uparrow$
No target conditioning	0.572 ( $\pm$ 0.029)	0.067 ( $\pm$ 0.032)
Concatenation	0.586 ( $\pm$ 0.030)	0.054 ( $\pm$ 0.056)
Early FiLM	0.579 ( $\pm$ 0.018)	0.044 ( $\pm$ 0.024)
PIFT (Late FiLM)	0.598 ( $\pm$ 0.042)	0.099 ( $\pm$ 0.064)

Appendix

Table 7: Ablation of target conditioning. Pair ranking measures the fraction of peptide queries for which the compatible small-molecule interface is embedded closer than its mismatched counterpart. Pair margin is the difference between mismatched- and compatible-pair distances,

d_{\mathrm{neg}}-d_{\mathrm{pos}}

, where larger values indicate stronger separation.

Panel of Figure 6 — Figure 6: Reference-specificity of peptide guidance on MDM2. (a) Self- and cross-reference IFA scoring. Changes in IFA relative to Off for the 1YCR- and 3LNJ-guided ensembles, evaluated against their own (self) or alternate (cross) peptide reference. Self-reference gains are larger for both guided ensembles. (b) Reference peptides and representative guided binding modes. Guided molecules occupy similar MDM2 regions while adopting reference-dependent local interaction geometries. (c) Differential population-level interaction fingerprints. Localized shifts reveal reference-dependent interaction patterns.

Figure 7: Full target-level results for guided and unguided molecular populations in IFA– $-$ Vina space.

		Average Vina Docking Score (kcal/mol, $\downarrow$ )
Category	Model	BRD2–H4	BCL2–BAX	Calpain–Calpastatin	cIAP1–SMAC	DAB1–ApoER2
Autoregressive	Hot2Mol	-7.00 ( $\pm$ 0.04)	-7.61 ( $\pm$ 0.05)	-6.56 ( $\pm$ 0.03)	-6.59 ( $\pm$ 0.05)	-7.35 ( $\pm$ 0.02)
Search-based	SynGA	-6.47 ( $\pm$ 0.09)	-7.33 ( $\pm$ 0.05)	-6.50 ( $\pm$ 0.03)	-6.04 ( $\pm$ 0.09)	-6.78 ( $\pm$ 0.03)
GFlowNet-based	CGFlow	-7.20 ( $\pm$ 0.05)	-7.93 ( $\pm$ 0.03)	-6.91 ( $\pm$ 0.03)	-6.38 ( $\pm$ 0.03)	-6.99 ( $\pm$ 0.05)
Diffusion / Flow-based	TargetDiff	-5.30 ( $\pm$ 0.04)	–	-6.74 ( $\pm$ 0.03)	-3.50 ( $\pm$ 0.03)	-6.02 ( $\pm$ 0.08)
	MolCRAFT	-5.02 ( $\pm$ 0.02)	-6.84 ( $\pm$ 0.03)	-7.50 ( $\pm$ 0.04)	-3.90 ( $\pm$ 0.04)	-5.32 ( $\pm$ 0.01)

Appendix

Table 8: Average Vina docking score for 10 PPI targets. Best results are in bold.

Figure 8: Heavy-atom-count-adjusted docking comparison. Bars show mean residual Vina score after fitting a separate Vina–heavy-atom-count regression within each target; the final group averages residuals across all 10 targets. Lower residuals indicate more favorable docking scores than expected at matched molecular size.

	PoseCheck Metrics ( $\uparrow$ )
Model	H-Bond Acceptors	H-Bond Donors	Van der Waals	Hydrophobic	Sum
Hot2Mol	0.201 ( $\pm$ 0.004)	0.128 ( $\pm$ 0.008)	7.187 ( $\pm$ 0.044)	2.968 ( $\pm$ 0.046)	10.485 ( $\pm$ 0.066)
SynGA	0.241 ( $\pm$ 0.024)	0.071 ( $\pm$ 0.006)	7.797 ( $\pm$ 0.051)	2.961 ( $\pm$ 0.045)	11.071 ( $\pm$ 0.088)
CGFlow	0.130 ( $\pm$ 0.006)	0.054 ( $\pm$ 0.009)	6.583 ( $\pm$ 0.032)	2.427 ( $\pm$ 0.013)	9.193 ( $\pm$ 0.048)
MolCRAFT	0.232 ( $\pm$ 0.022)	0.141 ( $\pm$ 0.024)	7.736 ( $\pm$ 0.093)	3.231 ( $\pm$ 0.044)	11.340 ( $\pm$ 0.124)
SynCoGen	0.262 ( $\pm$ 0.012)	0.139 ( $\pm$ 0.008)	7.364 ( $\pm$ 0.041)	2.883 ( $\pm$ 0.037)	10.648 ( $\pm$ 0.071)

Appendix

Table 9: PoseCheck metrics. Average interaction counts for the 10-target benchmark, restricted to models with successful evaluation on all 10 targets. Values are means

\pm

standard deviations over three runs; the main-text table reports only the corresponding Sum column. Best results are in bold, and the second-best H-bond acceptor value is underlined.

3D Molecular Generation

3D molecular generation creates molecules by modeling their chemical structures and three-dimensional geometries.

See all 59 papers

Joint Discrete and SE(3) Diffusion for Protein Interface Design
Oct 10, 2026Zhenqiao Song, Jianwen Xie, Lei LiProtein Binder Design Molecular Generation
Language Technologies Institute, Carnegie Mellon University · Lambda Inc.
An Integrated Deep-Learning Framework for Peptide-Protein Interaction Prediction and Target-Conditioned Peptide Generation with ConGA-PepPI and TC-PepGen
Apr 20, 2026Chupei Tang, Junxiao Kong, Moyu Tang +5Molecular Generation Conditional Generative Modeling
School of Mathematics and Statistics, Lanzhou University, 222 South Tianshui Road, Lanzhou 730000, Gansu, China · Cuiying Honors College, Lanzhou University, Lanzhou, Gansu, China · Shanghai Innovation Institute, Shanghai, China +1
Accurate structural modeling of chemically diverse molecular interfaces with Vilya-2
Jul 28, 2026Vilya Research, :, Pascal Sturmfels +10Molecular Docking Drug Discovery
Vilya Research

Oct 10, 2026J/K move · Enter open · S save

Joint Discrete and SE(3) Diffusion for Protein Interface Design

Zhenqiao Song, Jianwen Xie, Lei Li

Language Technologies Institute, Carnegie Mellon University · Lambda Inc.

Protein Binder Design Molecular Generation Multimodal Diffusion Models 3D Diffusion Models

Open paper AskPDF arXiv

The design of high-affinity, target-specific protein binders is central to the development of modern therapeutics. However, existing protein-generation methods remain limited in this setting, as many generated designs fail to simultaneously achieve geometric and physicochemical complementarity at binding interfaces. To address this challenge, we introduce PIDiff, a physicochemically enhanced multimodal diffusion framework for protein-protein complex design. PIDiff jointly designs target-conditioned protein sequences and structures by coupling diffusion in discrete sequence space with diffusion on the continuous (SE(3)) structural manifold. At its core, PIDiff employs PIFormer, a novel architecture that interleaves geometry-aware invariant point attention with global sequence self-attention. We further curate PIBench, a protein-protein complex dataset with diverse target scaffolds derived from the Protein Data Bank (PDB). We first pretrain PIDiff on PIBench and subsequently fine-tune it for two therapeutically relevant applications: minibinder design and antibody design. PPFrame consistently outperforms baseline methods across sequence design success rate, structure binding energy, and sequence-structure co-designability on all tasks.

Abstract

3D Molecular Generation

Explore similar work

Joint Discrete and SE(3) Diffusion for Protein Interface Design

An Integrated Deep-Learning Framework for Peptide-Protein Interaction Prediction and Target-Conditioned Peptide Generation with ConGA-PepPI and TC-PepGen

Accurate structural modeling of chemically diverse molecular interfaces with Vilya-2