Peptide2Mol: Equivariant Diffusion for Peptidomimetic Design

• Peptide2Mol is an E(3)-equivariant, non-autoregressive diffusion model that leverages peptide–protein interfaces and receptor pocket geometry to generate drug-like, peptide-mimetic molecules.
• It represents molecular structures as atomic graphs with explicit encoding of both covalent and noncovalent interactions, and employs partial diffusion to refine local binding features.
• Empirical results highlight that its fixed variant achieves the highest structural plausibility (83.80% PoseBusters rate), effectively mimicking key peptide binding modes in 3D.

Peptide2Mol is an E(3)-equivariant, non-autoregressive diffusion model for peptide-referenced, pocket-conditioned small-molecule design. It was introduced to address a gap in structure-based drug design: most target-conditioned small-molecule generators are trained only on protein–ligand complexes and therefore ignore structural information from peptide–protein and protein–protein interfaces, even though many endogenous interactions at protein surfaces are mediated by peptide segments or loops. In this framework, a peptidomimetic is defined as a small molecule that preserves the key noncovalent interaction pattern of a native peptide binder within a target pocket while adopting drug-like geometry and chemistry. Peptide2Mol learns from experimentally determined and AlphaFold-predicted peptide–protein interfaces, protein–ligand complexes, and standalone small-molecule conformations in order to generate small molecules that mimic peptide binding modes in 3D with pocket awareness (He et al., 7 Nov 2025).

1. Problem setting and design objective

Peptide2Mol is formulated around peptide-to-small-molecule translation in a structural setting rather than a purely ligand-centric one. The motivating claim is that training only on protein–ligand complexes neglects peptide-derived interaction logic, whereas many successful drugs have historically emerged by translating peptide interaction patterns into small molecules; the paper cites captopril and Saquinavir as examples (He et al., 7 Nov 2025).

The model’s central contribution is peptide- and pocket-referenced generation. During training, it references both the peptide binder and the receptor pocket environment, including partially diffused peptide side chains, so that generation is conditioned not only on a target site but also on the interaction motifs of the peptide occupying that site. This framing is intended to support the generation of small molecules in the specific context of the pocket that the peptide occupies, rather than treating pocket geometry as the only conditioning signal (He et al., 7 Nov 2025).

A common source of confusion is terminological. The 2025 toolkit paper on p2smi uses “Peptide2Mol” in a generic sequence-to-molecule sense and explicitly states that it does not describe a distinct tool named “Peptide2Mol”; in that paper, p2smi is presented as a practical peptide-specific FASTA-to-SMILES implementation for conversion and property analysis (Feller et al., 18 Apr 2025). By contrast, Peptide2Mol in (He et al., 7 Nov 2025) is a 3D generative diffusion model for pocket-conditioned small-molecule design from peptide interface information.

2. Molecular representation and network architecture

Peptide2Mol represents a ligand or ligand–pocket complex as an undirected atomic graph $M = (V, E)$. Each node $v_i$ carries an atom coordinate $r_i \in \mathbb{R}^3$ and an element-type feature $a_i \in \mathbb{R}^8$, with one-hot encoding over $\{ \mathrm{C, N, O, F, P, S, Cl, Br} \}$. Each edge $e_{ij}$ carries a bond/interaction feature $b_{ij} \in \mathbb{R}^6$ encoding single, double, triple, aromatic, non-bonded proximity, and an absorbing “no interaction” state (He et al., 7 Nov 2025).

The denoising backbone is a stack of six E(3)-equivariant GNN layers with rotation-equivariant convolutions that update both node features and coordinates. The equivariance condition is stated as follows: for any rigid transformation with rotation $R \in SO(3)$ and translation $t \in \mathbb{R}^3$, coordinate outputs transform consistently according to

$f(Rx + t) = R f(x) + t,$

while scalar features remain invariant. In practice, the model enforces this by using relative geometric quantities such as pairwise distances and unit direction vectors in message passing, by producing coordinate updates linear in direction vectors and normalized by distances, and by keeping atom-type and bond-type predictions invariant (He et al., 7 Nov 2025).

The architecture also performs explicit modeling of pocket–ligand contacts. Noncovalent interactions within $v_i$0 Å between ligand atoms and pocket atoms are encoded as edges. This means the graph is not restricted to covalent structure; it also embeds short-range contact geometry that is directly relevant to binding-site complementarity during denoising and generation (He et al., 7 Nov 2025).

Conditioning is implemented through a binary pocket indicator and a diffusion timestep embedding concatenated to node features. No explicit alignment or normalization is required at inference because coordinates are used in their native frames and E(3)-equivariance guarantees consistency up to rigid motions (He et al., 7 Nov 2025).

3. Diffusion formulation and partial diffusion

Peptide2Mol uses two Markov processes to define forward noising and reverse denoising. The Markov property is given as

$v_i$1

The forward process applies Gaussian noise to coordinates, atom-type features, and bond-type features with a predefined noise schedule $v_i$2:

$v_i$3

$v_i$4

$v_i$5

With $v_i$6 and $v_i$7, the closed-form marginals are

$v_i$8

$v_i$9

$r_i \in \mathbb{R}^3$0

As $r_i \in \mathbb{R}^3$1, the variables approach standard Gaussian priors used to initialize the reverse chain (He et al., 7 Nov 2025).

The reverse transition is parameterized by an E(3)-equivariant neural network:

$r_i \in \mathbb{R}^3$2

where $r_i \in \mathbb{R}^3$3 denotes any of $r_i \in \mathbb{R}^3$4, $r_i \in \mathbb{R}^3$5, or $r_i \in \mathbb{R}^3$6. Training minimizes denoising losses at a randomly sampled timestep:

$r_i \in \mathbb{R}^3$7

$r_i \in \mathbb{R}^3$8

$r_i \in \mathbb{R}^3$9

$a_i \in \mathbb{R}^8$0

with $a_i \in \mathbb{R}^8$1 (He et al., 7 Nov 2025).

Generation is full-graph and non-autoregressive. The model starts from Gaussian noise over all ligand variables and iteratively denoises over timesteps $a_i \in \mathbb{R}^8$2, jointly predicting atom coordinates, atom types, and bond types. The stated advantages are avoidance of exposure bias, faster sampling than sequential atom placement, and greater global coherence with respect to pocket geometry (He et al., 7 Nov 2025).

A distinctive feature is partial diffusion. During training, peptide side chains are partially diffused while pocket residues remain fixed; ligand geometries undergo progressive noising. At inference, selected variables can be frozen while others are diffused and denoised, enabling local edits without full resampling. The intended optimization workflow is to start from a generated ligand pose, select variables to refine, run reverse denoising from a mid-noise level while keeping fixed atoms and pocket context unchanged, and optionally apply Pocket2Mol-based refinement afterward (He et al., 7 Nov 2025). The paper explicitly notes that a masked update pattern can be summarized conceptually but that the exact formula and schedule are not specified.

4. Data sources, preprocessing, and evaluation protocol

Training data were drawn from three sources. The small-molecule component used the GEOM drug dataset with 304,322 molecules. Protein–ligand complexes were drawn from PDBBind with 38,860 complexes. Peptide–protein interfaces were drawn from BioLiP2 and AlphaFold DB, using monomeric models where buried loops are treated as ligands, for a total of 39,499 interfaces (He et al., 7 Nov 2025).

Preprocessing uses RDKit parsing and sanitization to ensure that training molecules are chemically parsable and sensible, including valencies and basic sanitization. Graph construction follows the node and edge definitions already noted, with explicit encoding of noncovalent contacts within $a_i \in \mathbb{R}^8$3 Å (He et al., 7 Nov 2025).

The main evaluation set is a CrossDock2020 subset of 10 protein–ligand complexes with PDB IDs 1BVR, 1ZYU, 2ATI, 4BNW, 5G3N, 1U0F, 2AH9, 2HW1, 4I91, and 5LVQ, matching prior work. Antibody–antigen pairs from SAbDab were used for residue replacement analysis via generated fragments (He et al., 7 Nov 2025).

Evaluation spans chemical and physical criteria. Chemical and drug-likeness metrics are QED, SA, and LogP. Physical plausibility is assessed by the PoseBusters passing rate, defined over 19 structural criteria for docking plausibility and molecular integrity. Geometry fidelity is assessed through bond-length distribution analysis across nine bond types: C–C, C=C, C–O, C=O, C–N, C=N, C–Cl, C–S, and C–F. Mimicry analysis uses PMI between residue types and generated small-molecule fragments (He et al., 7 Nov 2025).

The reported training details are deliberately narrow. The architecture contains six E(3)-equivariant GNN layers, conditioning uses timestep embeddings and binary pocket indicators, and the loss weights are $a_i \in \mathbb{R}^8$4. The paper does not specify optimizer, batch size, learning rate, epochs, regularization, augmentation, or exact noise schedule values. It does state that equivariance reduces the need for rotational augmentation (He et al., 7 Nov 2025).

5. Empirical performance and physical plausibility

On the reported test set, Peptide2Mol and its partially refined variant, Peptide2Mol-Fixed, were compared with LiGAN, Pocket2Mol, TargetDiff, and PocketFlow (He et al., 7 Nov 2025).

Method	QED	SA	LogP	PBrate
LiGAN	0.428	0.546	1.224	39.50%
Pocket2Mol	0.587	0.758	1.063	71.60%
TargetDiff	0.430	0.550	1.249	36.90%
PocketFlow	0.497	0.769	3.521	46.00%
Peptide2Mol	0.501	0.612	0.638	45.30%
Peptide2Mol-Fixed	0.509	0.637	0.729	83.80%

The most emphasized result is that Peptide2Mol-Fixed achieves the highest PoseBusters passing rate among the compared methods, at 83.80%, which the paper interprets as strong structural plausibility under rigorous physical filters (He et al., 7 Nov 2025). Peptide2Mol’s QED is described as competitive with pocket-conditioned baselines such as PocketFlow and better than LiGAN and TargetDiff. Its SA is lower than Pocket2Mol and PocketFlow but comparable to LiGAN and TargetDiff, which the paper associates with peptide-like chemistry. Its LogP is lower, hence more polar, and this is explicitly described as consistent with peptide mimicry (He et al., 7 Nov 2025).

The PoseBusters waterfall analysis further differentiates model behavior. Peptide2Mol is reported to excel at controlling intermolecular distance constraints and reducing steric clashes, whereas Pocket2Mol excels at bond length distribution and internal energy checks. The paper therefore argues that combining pocket-aware diffusion with local refinement yields the strongest overall plausibility (He et al., 7 Nov 2025).

Bond-length distributions for Peptide2Mol are reported to closely match the training distribution across the nine analyzed bond types while still capturing peptide-like bond-length patterns. This is presented as evidence that the generated structures are chemically realistic rather than merely passing post hoc filters (He et al., 7 Nov 2025).

The runtime profile is not quantitatively reported. The paper only states conceptually that non-autoregressive diffusion should yield faster sampling than autoregressive placement and improve global consistency, which would be advantageous for scaling and for iterative optimization (He et al., 7 Nov 2025). This should be read as a design rationale rather than a benchmarked speed claim.

6. Mimicry analysis, case studies, limitations, and related tooling

Peptide2Mol includes a residue-to-fragment analysis aimed at actionable peptidomimetic design. In antibody CDR–antigen contexts, the study analyzes four residues—TYR, ASP, ARG, and LEU—and uses PMI statistics to associate them with preferred small-molecule fragments. High-PMI fragments preserve the expected chemical motifs: TYR is associated with aromatic rings and hydroxyl-containing groups; ASP with polar oxygen-rich groups described as carboxylate-like; ARG with nitrogen-rich cationic motifs described as guanidinium-like mimicry; and LEU with carbon-rich hydrophobic chains (He et al., 7 Nov 2025). The intended implication is not simple atom-level substitution, but identification of fragment classes that can serve as small-molecule side-chain surrogates in a structural context.

Two representative examples are highlighted. In PDB 7WXO, a peptide binder was converted into a small molecule that mimics its binding interface in the target pocket. In PDB 3NGB, an antibody CDR was transformed into a small-molecule mimic that preserved key interactions with the antigen. The qualitative interpretation is that hydrogen-bond donors and acceptors, aromatic stacking, and hydrophobic contacts were preserved in ways consistent with the original binders (He et al., 7 Nov 2025).

The paper also states several limitations. Generated molecules inherit peptide-like physicochemical properties, including lower LogP and modest QED and SA, which may not optimally align with oral small-molecule drug criteria without further optimization. No quantitative metric for peptide–small-molecule similarity is provided; the authors suggest that future work could introduce pharmacophore or interaction-fingerprint similarity scores. Atom and bond types are noised using Gaussian relaxations of one-hot features, and the paper notes that more sophisticated categorical diffusion might improve discrete consistency. Performance also depends on accurate pocket representation, so errors in peptide-derived pocket context may mislead generation. Finally, mixing protein–ligand and peptide–protein interfaces may mitigate ligand-only bias but may also introduce interface-specific biases from the data sources (He et al., 7 Nov 2025).

The practical workflow specified for use is also narrowly defined. One provides a protein pocket structure, optionally includes the peptide structure during training, constructs the ligand–pocket graph with atom coordinates, atom types, edge features, binary pocket indicators, and noncovalent edges within $a_i \in \mathbb{R}^8$5 Å, initializes ligand variables from Gaussian noise, runs non-autoregressive denoising, then applies RDKit sanitization and Pocket2Mol refinement. Validation includes QED, SA, LogP, PoseBusters, and optionally docking; optimization proceeds by freezing pocket atoms, optionally freezing well-contacting ligand parts, and diffusing the remainder from a mid-noise level (He et al., 7 Nov 2025).

For reproducibility, code and pretrained models are reported at https://github.com/BLUE-Flowing/Peptide2Mol/, with data sources including GEOM drug dataset, PDBBind, BioLiP2, AlphaFold DB, the CrossDock2020 subset, and SAbDab (He et al., 7 Nov 2025). A related but distinct tool is p2smi, which supports FASTA-to-SMILES conversion, noncanonical amino acids, cyclization chemistries, N-methylation, PEGylation, and property analysis for peptides (Feller et al., 18 Apr 2025). That toolkit is relevant to peptide representation and cheminformatics preprocessing, but it is not the same methodological object as the Peptide2Mol diffusion model.

A plausible implication of the Peptide2Mol framework is that peptide–protein interface knowledge can be used to bridge peptide-inspired recognition logic and small-molecule design for difficult protein surfaces, including protein–protein interaction targets. The paper itself frames future directions in terms of physics-based evaluation loops such as MD and free energy perturbation, explicit peptide–small-molecule similarity metrics, improved discrete diffusion, and broader benchmarking on diverse PPI targets (He et al., 7 Nov 2025).