MolChord: Structure-Based Molecule Generation

• MolChord is a multimodal structure-based drug design framework that aligns 3D protein structures, molecular graphs, sequence data, and text to enable conditional small-molecule generation.
• It integrates a diffusion-pretrained structure encoder with an autoregressive language model and adapter module for effective cross-modal alignment and molecule design.
• The framework uses staged training with Direct Preference Optimization to refine outputs on affinity, synthesizability, drug-likeness, and diversity, achieving state-of-the-art benchmarks.

MolChord is a structure-based drug design (SBDD) framework that addresses protein-guided ligand generation as a multimodal structure–sequence alignment problem spanning 3D protein and molecule structures, biological and chemical sequence representations, and natural language. Its distinguishing features are large-scale alignment of structural and sequence/text modalities using a diffusion-pretrained encoder and autoregressive LLM, and property-aware refinement through Direct Preference Optimization (DPO). This integrated architecture enables conditional small-molecule generation for proteins with state-of-the-art performance on affinity, synthesizability, drug-likeness, and diversity metrics on the CrossDocked2020 benchmark (Zhang et al., 31 Oct 2025).

1. Structure–Sequence Alignment in Protein-Guided Molecule Generation

MolChord frames SBDD as an alignment problem between protein structural representations (3D pocket coordinates), molecular representations (SMILES or atom-level graphs), protein/molecule sequences (FASTA, SMILES), and textual functional/natural language. The central challenge is that protein–ligand binding data are limited, and protein/molecule representations differ both structurally and semantically. MolChord bridges these gaps by leveraging foundation-model-scale pretraining and cross-modal alignment tasks. This alignment spans:

• Encoding of protein structures as residue-level 3D coordinates and annotations.
• Encoding of ligand structures as atom-level 3D graphs.
• Language-model-based tokenization of protein sequences, molecule SMILES, and text.
• Insertion of structural features as latent embeddings in an autoregressive language-model context.

This multimodal alignment allows conditioning the sequence generator on protein structure, molecule structure, and textual prompts in a unified autoregressive generation space—substantially expanding the effective data for conditional molecule design beyond scarce complex-labeled structures (Zhang et al., 31 Oct 2025).

2. Model Architecture: Diffusion Encoder, Autoregressive Generator, and Modal Alignment

MolChord combines three principal components:

1. Diffusion-pretrained structure encoder: A FlexRibbon-style architecture pretrained on large-scale protein and molecule structures. It processes residue/atom sequences and denoises noisy coordinates, yielding rich latent structure-aware features.
2. Autoregressive multimodal sequence generator (NatureLM): An autoregressive transformer LLM (16 layers, 2048 hidden, 32 heads) trained to model protein FASTA, SMILES, and natural language, augmented with domain-specialized tokens.
3. Adapter module: A gated MLP mapping structural encoder outputs to the generator's hidden space, trained while encoder and generator weights are frozen (alignment stage).

Modalities are fused by interleaving embedded textual tokens and adapted structural features into a single sequence, enabling the generator to attend jointly over 3D structure and sequence/language representations in generation. This architecture enables prompts such as "Generate a compound based on the pocket <3d pocket>," where the <3d pocket> tag is replaced by the structural encoding.

A Variational Autoencoder (VAE) module is used post-alignment to stochasticize conditioning (VAE perturbation) during supervised fine-tuning and DPO, improving robustness and diversity by offsetting structural features with noise sampled from a learned latent distribution.

3. Training Paradigm: Staged Alignment, Supervised Fine-tuning, and DPO-Based Refinement

MolChord's training follows a three-stage protocol:

• Stage A (Alignment Pretraining): The adapter is trained using large unlabeled or weakly labeled corpora to inject protein or molecule structural features into the sequence model's hidden state, optimizing next-token prediction over language targets (FASTA, SMILES, and text). Only the adapter is updated—both structure encoder and generator are frozen. This alignment includes structure-to-sequence and structure-to-text instruction-following tasks, performed over proteins, molecules, and protein–ligand complexes.
• Stage B (Supervised Fine-tuning): The aligned model is fine-tuned on curated protein–ligand pairs from CrossDocked2020, with a VAE providing latent noise perturbations to the conditioning structure for improved output diversity and robustness. The generator is now trained to produce ligand SMILES given encoded protein pocket structures.
• Stage C (Direct Preference Optimization): Model parameters are further refined using DPO on a curated property-aware preference set drawn from protein–ligand pairs with low data coverage. Rewards combine docking score (AutoDock Vina) and a chemical plausibility penalty (penalty for excessive fused rings). The DPO objective encourages the model to raise the relative likelihood of higher-reward (docking- and property-optimized) ligands, compared to a frozen reference policy.

No explicit CLIP-style or contrastive loss is used; "alignment" is implemented autoregressively via structure-inserted language modeling and preference-based policy refinement (Zhang et al., 31 Oct 2025).

4. Evaluation: Affinity, Drug-Likeness, Synthesizability, Diversity, and OOD Generalization

MolChord is evaluated on standard benchmarks and metrics in SBDD:

Method	Vina Dock ↓	High Affinity ↑	QED ↑	SA ↑	Diversity ↑	Success Rate ↑
MolChord	-7.62	55.1%	0.56	0.77	0.76	33.2%
MolChord-RL^dock	-9.29	83.7%	0.44	0.77	0.63	59.3%
MolChord-RL (full)	-8.59	74.6%	0.56	0.78	0.71	53.4%

MolChord exceeds or matches leading benchmarks in Vina Dock (docking proxy), QED (Quantitative Estimate of Drug-likeness), SA (Synthetic Accessibility), output molecule diversity, and a composite success rate involving all three properties. The RL variant with fused-ring penalty achieves a balanced optimum, preventing the common pathology of optimizing only for docking at the cost of chemical plausibility (Zhang et al., 31 Oct 2025).

Performance remains stable or even improves on non-homologous (remote) protein targets compared to homologous ones, indicating generalization to out-of-distribution structures.

Efficiency is high: generating 100 compounds for a target in approximately 4 seconds on a single A100 GPU. Qualitative analyses (fused-ring count, structure plausibility) confirm that outputs are not only high-affinity but realistic by medicinal chemistry standards.

5. Innovations and Distinctions

Key innovations of MolChord relative to prior work and benchmarks are:

• Cross-modal structural–sequence alignment that leverages both abundant structure/sequence data and scarce paired protein–ligand complexes, increasing effective supervision.
• Autoregressive instruction-following LLM fine-tuned for molecular sequence generation conditioned on multimodal structural input.
• VAE-perturbed conditioning to inject stochasticity and improve output diversity and robustness.
• Property-aware Direct Preference Optimization with composite reward functions, enabling fine-grained control of trade-offs between affinity, synthesizability, diversity, and chemical realism.
• Empirical demonstration of improved OOD transfer, diversity of generated chemotypes, and fast generation.

6. Practical Considerations, Limitations, and Future Directions

MolChord requires protein pocket structures as input; performance in the absence of such structures is not addressed. All binding evaluation is docking-score–based, a practical proxy but only an approximation to actual affinity. The system is substantial in size (4.2B parameters, multiple pretrained modules), with pretraining on ~78M protein structures and an alignment corpus of ~1.1M sequence–structure examples, which may constrain reproducibility. Limiting reward functions to docking and fused rings may not capture all relevant medicinal chemistry criteria; the framework could naturally extend to multi-objective or broader property conditioning. Current generation outputs 1D molecular sequences (SMILES); direct 3D pose generation is not implemented (Zhang et al., 31 Oct 2025).

7. Significance for SBDD and Generative Molecular Design

MolChord's unified structural–sequence alignment and property-aware refinement mark a shift from narrow pocket-to-ligand mapping to a foundation-model-style, multimodal generative approach. Its demonstration that explicit cross-modal alignment and reward-based preference optimization substantially improve the practical yield of SBDD generative models sets a new technical baseline in the field. The architecture is extensible to further structural modalities, richer textual conditioning, or more complex reward integration for scalable, biologically and chemically plausible molecular design.