MolEdit3D: 3D Ligand Editing for Drug Design

• MolEdit3D is a 3D fragment-based method for SBDD that directly edits ligands in 3D space to improve binding pocket complementarity.
• It leverages a hierarchical MPNN with generative pre-training on ChEMBL and annealed Bayesian sampling to achieve faster and more efficient ligand optimization.
• The framework integrates target-guided self-learning to optimize binding affinity while ensuring chemical validity, drug-likeness, and synthesizability.

MolEdit3D is a method for structure-based drug design (SBDD) that generates novel ligand molecules from the 3D structure of a target protein and its binding pocket by editing ligands directly in 3D rather than only in 2D graph space. It combines a 3D graph editing model, generative pre-training on abundant 3D ligands, and target-guided self-learning within a multi-chain annealed Bayesian sampling procedure. The method is motivated by what it calls a misalignment between the action space and the objective in prior optimization-based approaches: many earlier systems edit 2D molecules but are evaluated by 3D docking-related criteria, whereas MolEdit3D makes fragment and torsion edits in the same geometric domain as the scoring objective (Yang et al., 2024).

1. Problem formulation and conceptual motivation

MolEdit3D addresses SBDD under the assumption that ligand quality is fundamentally determined by 3D target–ligand complexes. In the formulation given for the method, generated ligands are expected to fit the target well and have high predicted binding affinity, remain drug-like and synthesizable, and be chemically valid and structurally plausible (Yang et al., 2024).

The work positions prior deep-learning approaches in two families. The first family consists of conditional 3D ligand generation methods conditioned on target binding sites or target–ligand complexes. The stated limitations are data scarcity and quality in experimentally measured 3D complexes, together with distributional conservatism: models trained to imitate observed conditional distributions may remain close to known ligands and may not exceed known actives. The second family consists of optimization-based molecule editing methods that edit 2D molecules and use molecular docking as the optimization objective. Here the stated problems are efficiency and action/objective mismatch, because the optimization target depends on 3D interactions while the edit actions occur in 2D graph space (Yang et al., 2024).

The central claim is that SBDD is intrinsically a 3D problem, so acting directly in 3D should improve alignment between the generator and the evaluator. MolEdit3D therefore represents molecules as linked rigid 3D fragments, explicitly predicts attachment sites and torsion angles, and keeps the generated ligand in a 3D pose in the pocket. This suggests a design in which changing a fragment or torsion angle has a direct effect on pocket complementarity, while target-aware scoring can use Vina local minimization on the current 3D pose rather than full docking for every proposal (Yang et al., 2024).

2. Molecular representation and edit operations

MolEdit3D represents a molecule as a set of rigid 3D fragments connected by rotatable single bonds. Its fragment library is built by breaking non-terminal single bonds in molecules from ChEMBL, labeling broken bonds as editable sites, adding hydrogen atoms to preserve valency, and generating 3D conformers for fragments using RDKit. The appendix specifies that the 1000 most frequent fragments are selected, each fragment has at most 10 heavy atoms, and different 3D conformers of the same 2D fragment can be treated as different rigid fragments (Yang et al., 2024).

The molecular state is encoded as a hierarchical graph with an atom layer and a fragment layer. Atomic node features include atomic number, element type, charge, and 3D coordinates; atomic edge features include bond type; and the hidden dimension is 64. At the fragment level, atoms separated by rotatable single bonds belong to different fragments, so the model reasons simultaneously over local atomic context and coarser fragment connectivity (Yang et al., 2024).

The action space distinguishes addable sites and deletable sites. Addable sites are editable sites with hydrogen atoms attached, while deletable sites are editable sites with non-hydrogen atoms attached. Two main edit types are supported:

• Add a fragment: choose an addable site in the current molecule, choose a fragment from the library, choose an addable site in that fragment, connect the skeleton and fragment, and predict the torsion angle around the new bond.
• Delete a fragment: choose a deletable site, break the corresponding bond, and replace the detached side with hydrogen to preserve valency.

The geometry of fragment addition is parameterized by a torsion angle over four consecutive atoms around the new connecting bond, and torsion is discretized into bins

$\Gamma = \{0, 10, 20, \cdots, 350\}.$

Accordingly, MolEdit3D does not predict all coordinates from scratch; it edits geometry through fragment assembly plus torsion prediction (Yang et al., 2024).

An important architectural distinction is that the model is 3D-aware but not described as equivariant. The paper explicitly states that MolEdit3D is not presented as an SE(3)- or E(3)-equivariant model; instead, 3D coordinates are included in atom features, geometry is modeled by rigid fragments and torsion angles, and the architecture is a hierarchical MPNN rather than an explicitly equivariant geometric network (Yang et al., 2024).

3. Neural architecture and generative pre-training

The editing model predicts where to edit in the current skeleton, whether the selected opportunity is addable or deletable through sampling over valid editable edges, which fragment to add when addition is selected, which site on the fragment to attach, and what torsion angle to use. Encoding is performed by a network $\phi$ built from Hierarchical Message Passing Neural Networks (HMPNNs), with one message-passing stage at the atom level and another at the fragment level (Yang et al., 2024).

At the atom level, the molecule is encoded as a graph

$g=(A, h^\text{atom-node}, h^\text{atom-edge}),$

and atom hidden states are computed by

$$h^\text{atom-node}_u = MPNN_1(g)_u \in \mathbb{R}^d.$$

Fragment embeddings are then formed by mean pooling atom embeddings:

$$h^\text{frag-node}_i = MeanPool_{u\in V_i}(h^\text{atom-node}_u) \in \mathbb{R}^d.$$

A second MPNN updates fragment nodes, so the model can score fragment-level edit sites and fragment choices using the hierarchical representation (Yang et al., 2024).

For edit-site prediction, the model forms separate distributions over addable and deletable fragment-level edges:

$$p_\text{add}(r|x_\text{skel}) = softmax(\{\text{score}_{j,k}^{\text{skel}}\}_{(j,k)\in E_a}),$$

$$p_\text{delete}(r|x_\text{skel}) = softmax(\{\text{score}_{j,k}^{\text{skel}}\}_{(j,k)\in E_d}),$$

and samples the selected edge from

$$\frac{1}{2}p_\text{add} + \frac{1}{2}p_\text{delete}.$$

If the selected site is addable, fragment selection is performed over the library $H$ through

$$p_\text{frag}(f|x,r) = softmax(\{\text{score}_{f}\}_{f \in H}),$$

followed by attachment-site prediction on the chosen fragment and discrete torsion-angle prediction over $\phi$0 (Yang et al., 2024).

Generative pre-training is used to learn target-independent properties before target-specific optimization. The pre-training source is ChEMBL: 150K molecules are used, their original 2D structures are converted to 3D conformers and minimized using RDKit, several iterative decomposition orders are sampled per molecule, and the total pre-training set size is 1.2 million data points. Training pairs are constructed by repeatedly removing a leaf fragment from a drug-like 3D molecule and asking the model to reconstruct the full molecule from the partial one. The pre-training objective maximizes the conditional likelihood

$\phi$1

where $\phi$2 denotes the editing model (Yang et al., 2024).

The work interprets these target-independent properties as including chemical validity, drug-likeness, synthesizability, conformational plausibility or energetic stability, and realistic ring systems and torsion preferences. A plausible implication is that pre-training acts as a strong prior over chemically acceptable local 3D moves before any target-specific adaptation is attempted (Yang et al., 2024).

4. Target-guided self-learning and annealed Bayesian sampling

Target specificity is introduced not by directly encoding protein features in the neural proposal model, but by placing generated molecules in the target pocket and evaluating them with a target-guided objective during sampling. The score is

$\phi$3

The appendix further states that the negative of Vina minimization energy is used in the objective and that SAScore is rescaled as $\phi$4, so higher $\phi$5 is better (Yang et al., 2024).

Sampling is described as multi-chain, stochastic, MCMC-like or Bayesian sampling with simulated annealing and online self-training. Each chain starts from an initial molecule such as methane, placed in the target binding site. At each step the model proposes one edited neighbor from the current state, and the proposal is accepted with probability

$\phi$6

If $\phi$7, the pair $\phi$8 is added to a self-training dataset, and the model is periodically updated using weighted maximum likelihood estimation with weight

$\phi$9

The paper explicitly contrasts this with standard policy gradient: MolEdit3D uses all chains simultaneously, uses annealed acceptance or rejection, and performs online weighted MLE rather than policy-gradient estimation over complete trajectories (Yang et al., 2024).

Two sampling scales are reported: MolEdit3D (L) uses 1000 chains, and MolEdit3D uses 5000 chains. Additional sampling details include 1000 total sampling steps, temperature annealing every 5 steps by a factor of 0.97, model training every 5 steps, a self-training buffer with maximum size 75K and FIFO replacement, up to 1000 gradient updates in each self-training phase, a final restriction to at most 40 heavy atoms, and random selection of 1000 molecules from the final step for evaluation (Yang et al., 2024).

A major practical claim is that MolEdit3D uses Vina local minimization rather than full docking during optimization because the ligand is already maintained in a 3D pose in the pocket. The appendix claims that this makes MolEdit3D about 100x faster than relying on docking for every proposal. This suggests that the method’s efficiency gain is tied directly to the choice of 3D edit representation, rather than only to the neural network architecture (Yang et al., 2024).

5. Experimental setting, metrics, and empirical results

Evaluation uses 10 protein targets from CrossDocked2020: 1FKG, 2RD6, 3H7W, 3P0P, 3VRJ, 4CG9, 4OQ3, 4PS7, 5E19, and 5MKU. All methods generate 1000 ligands per target. Baselines include the conditional generation methods liGAN, AR, GraphBP, DESERT, and Pocket2Mol, together with the optimization baselines MARS adapted to SBDD using Vina score, AutoGrow4, and RGA (Yang et al., 2024).

The reported metrics are Validity, Uniqueness, Diversity, High Affinity, Vina, QED, SA, and Success Rate. Validity is the percentage of generated molecules readable by RDKit and connected as one component; High Affinity is the percentage whose affinity is higher than the reference ligand; Vina is a docking-based score after a full docking procedure for more accurate evaluation; and Success Rate is the percentage of molecules satisfying all of $g=(A, h^\text{atom-node}, h^\text{atom-edge}),$0, $g=(A, h^\text{atom-node}, h^\text{atom-edge}),$1, and $g=(A, h^\text{atom-node}, h^\text{atom-edge}),$2 kcal/mol. The thresholds are described as the 10th percentile of approved drugs in DrugCentral for QED and SA, with the Vina threshold corresponding to affinity better than about $g=(A, h^\text{atom-node}, h^\text{atom-edge}),$3M (Yang et al., 2024).

The main quantitative result is that MolEdit3D achieves state-of-the-art performance on the majority of the evaluation metrics. The abstract highlights an improvement from the best prior Vina score of -9.77 for Pocket2Mol to -10.16 for MolEdit3D, together with an absolute 13.8% increase in Success Rate (Yang et al., 2024).

Metric	Pocket2Mol	MolEdit3D
Validity	not stated in the comparison summary	100.0%
Uniqueness	not stated in the comparison summary	99.2%
Diversity	not stated in the comparison summary	0.880
High Affinity	61.1%	70.3%
Vina	-9.77	-10.16
QED	not stated in the comparison summary	0.55
SA	0.74	0.78
Success Rate	68.2%	82.0%

A lighter variant, MolEdit3D (L), still reports Vina of -10.00 and Success Rate of 78.7. The paper attributes the overall gains to four components: a 3D action space aligned with a 3D objective, fragment-based editing for validity and local plausibility, pre-training for target-independent properties, and self-learning for target specificity (Yang et al., 2024).

Ablation on target 5MKU separates the effects of pre-training and self-learning. Without pre-training and without self-learning, the reported metrics are High Aff 10.7, Vina -8.83, QED 0.68, SA 0.67, and Success 59.5. Without pre-training but with self-learning, the values become High Aff 16.3, Vina -9.07, and Success 67.0. With pre-training but without self-learning, they become High Aff 24.1, Vina -9.35, QED 0.70, SA 0.83, and Success 85.6. With both pre-training and self-learning, the reported values are High Aff 53.6, Vina -10.30, QED 0.64, SA 0.84, and Success 94.1. The paper’s stated interpretation is that pre-training contributes most to target-independent quality and overall success rate, especially synthesizability, whereas self-learning contributes most strongly to target-dependent affinity metrics; combined, they are synergistic (Yang et al., 2024).

Additional target-independent analysis reports that MolEdit3D generated molecules have the best overlap with DrugCentral ring-size frequencies and that MolEdit3D and DESERT show the best alignment to CrossDocked2020 reference ligands for CCCC and Cccc torsion-angle distributions. The paper interprets these patterns as support for the claim that fragment-based 3D generation with torsion modeling yields more realistic conformational preferences (Yang et al., 2024).

6. Distinction from similarly named and adjacent systems

MolEdit3D is distinct from MolEdit, despite the name overlap. MolEdit is described as a knowledge editing framework for multimodal molecule LLMs operating over molecular graphs, SMILES strings, and captions or text descriptions, and it evaluates edits on molecule-to-caption and caption-to-molecule generation. Its tasks use inputs and outputs $g=(A, h^\text{atom-node}, h^\text{atom-edge}),$4, $g=(A, h^\text{atom-node}, h^\text{atom-edge}),$5, and $g=(A, h^\text{atom-node}, h^\text{atom-edge}),$6, its evaluation uses BLEU, ROUGE, METEOR, Levenshtein, and MACCS-based metrics, and it does not formulate tasks over 3D coordinates, conformers, distance matrices, SE(3)-equivariant features, or geometry-conditioned generation or editing. In that sense, MolEdit is squarely a 2D/text/graph/SMILES-centered framework, not a 3D editing framework (Lei et al., 16 Nov 2025).

That distinction matters because MolEdit3D solves a different problem. Rather than correcting factual associations inside a pretrained multimodal LLM without retraining the full model, it performs target-aware 3D ligand optimization by add/delete fragment actions, torsion prediction, Vina local minimization, and annealed Bayesian sampling (Yang et al., 2024). A common misconception is therefore to treat the two systems as successive versions of the same method; the available descriptions indicate that they address different levels of abstraction and different representations.

A second adjacent system is El Agente Estructural, which is a multimodal, natural-language-driven geometry-generation and manipulation agent for autonomous chemistry and molecular modelling. It performs tool-based, geometry-aware edits directly on Cartesian coordinates, supports atom- and fragment-level replacement, connectivity manipulation, stereochemical control, isomer interconversion, ligand binding and ligand exchange, and mechanism-guided geometry generation and modification. Unlike MolEdit3D, it is not presented as a monolithic generative model with a single optimization objective; instead, it is an agentic orchestration layer over chemistry-aware tools (Choi et al., 4 Feb 2026). This suggests that MolEdit3D and El Agente Estructural occupy adjacent but nonidentical positions in the 3D molecular editing landscape: the former is a fragment-editing optimizer for SBDD, whereas the latter is a natural-language-controlled molecular workbench for direct xyz manipulation.

7. Limitations, assumptions, and practical scope

Several limitations are explicit in the description of MolEdit3D. First, like many SBDD methods, it depends heavily on the reliability of Vina as an affinity proxy. If Vina is poorly correlated with true activity, optimization may overfit the scoring function. Second, target information is indirect: the protein is not encoded directly by the neural model in the described architecture, and target-specificity enters primarily through the objective during sampling and self-learning. Third, the fragment vocabulary is constrained to the top 1000 frequent fragments from ChEMBL, with at most 10 heavy atoms per fragment, which helps validity but may restrict novelty or unusual chemistry. Fourth, torsions are discretized into 10-degree bins, which is simpler than continuous geometry modeling. Fifth, the method requires per-target sampling and self-training, making deployment heavier than one-shot conditional generation. Finally, the paper notes that metrics such as QED and SAscore have known limitations and may not perfectly reflect medicinal chemistry value (Yang et al., 2024).

Within those assumptions, the method is presented as especially useful when a 3D protein structure is available, novel ligands rather than database retrieval are desired, docking-based or minimization-based target guidance is available, and a balance of affinity and chemical plausibility is important. It is also positioned as particularly attractive when full target–ligand training complexes are scarce but abundant unlabeled molecular structures are available for pre-training (Yang et al., 2024).

In summary, MolEdit3D is best understood as a 3D fragment-editing framework for SBDD that seeks to repair a specific methodological mismatch in earlier optimization systems: they optimize a 3D objective while acting in 2D. Its response is to make the edit space itself geometric, to learn a prior over plausible 3D ligands by generative pre-training, and to adapt that prior to a target by annealed Bayesian sampling with target-guided self-learning. The reported benchmark results indicate strong performance on affinity-related and composite metrics, while its limitations locate it clearly within docking-guided, fragment-based, target-specific optimization rather than universal 3D molecular design (Yang et al., 2024).