MolChord

MolChord is a structure-based drug design (SBDD) model that generates small-molecule ligands tailored to a target protein pocket. The central problem in SBDD is producing molecules that both fit a binding site geometrically and carry drug-like properties, while keeping the chemistry synthetically realistic. Many generative approaches optimize binding affinity at the expense of these other qualities, or struggle to connect a protein's three-dimensional structure with the sequential and textual descriptions that large language models handle well.

MolChord addresses this by aligning structural and sequential representations of both proteins and molecules. It uses NatureLM, an autoregressive foundation model that unifies natural-language text, small molecules (as SMILES), and proteins (as FASTA), as its molecule generator, paired with a diffusion-based structure encoder that injects pocket geometry into the generation process. To steer molecules toward desired properties, the authors build a property-aware preference dataset and refine the model with Direct Preference Optimization (DPO).

The model was developed by researchers at Zhongguancun Academy and the University of Science and Technology of China (USTC) and released as a preprint in October 2025. It sits alongside diffusion-based SBDD methods such as TargetDiff and DecompDiff, but distinguishes itself by grounding generation in a text-aware multimodal language model rather than a purely geometric generative process.

#Key Features

• Structure-sequence alignment: MolChord jointly aligns protein and molecule 3D structures with their sequential (FASTA/SMILES) and textual descriptions, bridging geometric context and language-model generation.
• Multimodal generator: It leverages NatureLM, an autoregressive model that unifies text, small molecules, and proteins, so ligand generation can draw on a broad chemical and biological prior.
• Diffusion structure encoder: A diffusion-based encoder conditions generation on the protein pocket's three-dimensional geometry.
• Preference-based property control: A curated property-aware preference dataset and Direct Preference Optimization guide molecules toward favorable drug-likeness and synthesizability without sacrificing binding.

#Technical Details

MolChord couples a diffusion structure encoder, which represents the target pocket, with the NatureLM autoregressive generator that emits candidate ligands. Property guidance is achieved by assembling preference pairs and applying Direct Preference Optimization to the alignment stage. On the CrossDocked2020 benchmark, MolChord reaches a 33.2% overall success rate, substantially above Pocket2Mol (24.4%), TargetDiff (10.5%), and DecompDiff (24.5%). It reports a median Vina Dock score of -7.62 kcal/mol with 55.1% high-affinity molecules, alongside strong drug-likeness and synthesizability (QED 0.56, SA 0.77) and high diversity (0.76). These results reflect its emphasis on balancing binding affinity against the property and synthesizability metrics that other methods often trade away.

#Applications

MolChord targets early-stage drug discovery, where researchers need to generate novel candidate molecules for a specific protein target. By conditioning on pocket structure while optimizing drug-like properties, it supports hit generation and lead ideation for medicinal chemists and computational drug designers who want candidates that are both potent and tractable to synthesize.

#Impact

MolChord illustrates how multimodal foundation models that unify text, sequence, and structure can be adapted to three-dimensional drug design, and how preference optimization borrowed from language-model alignment can enforce property constraints in molecular generation. Its balanced CrossDocked2020 results suggest that language-model-based generators can compete with specialized geometric diffusion models on structure-based design. As a recent preprint awaiting peer review, its generalization beyond the CrossDocked2020 setting and experimental validation of generated molecules remain to be established.