AI Designs Better Drugs by Aligning Protein and Molecule Structures

Drug discovery is notoriously slow and expensive, often taking over a decade and billions of dollars to bring a single treatment to market. A new AI framework called MolChord offers a faster, more precise way to design potential drugs by aligning the 3D structures of proteins and molecules, enabling the generation of compounds that are both effective and safe. This approach could accelerate the search for new therapies, reducing reliance on traditional trial-and-error methods.

MolChord's key innovation lies in its ability to align structural representations of proteins and small molecules with their sequential descriptions, such as FASTA sequences for proteins and SMILES strings for molecules. Researchers integrated a diffusion-based structure encoder with an autoregressive sequence generator, using a lightweight adapter to bridge these components. This allows the model to generate molecules conditioned on protein binding pockets, ensuring the compounds are tailored to specific biological targets. The training involved three stages: cross-modal alignment with diverse biological data, supervised fine-tuning on protein-ligand pairs, and direct preference optimization to refine molecules based on binding affinity and drug-like properties.

In experiments on the CrossDocked2020 dataset, MolChord outperformed existing methods across critical metrics. It achieved a high affinity score, with 74.6% of generated molecules binding as well or better than known ligands, and maintained strong drug-likeness with a QED score of 0.78 and synthetic accessibility (SA) of 0.71. The success rate, which combines affinity, QED, and SA, reached 53.4%, indicating a balanced output of viable drug candidates. For example, MolChord produced molecules with an average of 1.75 fused rings, closely matching approved drugs like those in the FDA database (1.78 rings), whereas other methods often overproduced complex ring systems that hinder synthesizability. The model also showed robust generalization, improving performance on out-of-distribution proteins by 0.17 in docking scores, thanks to its large-scale pretraining on diverse structural data.

This advancement matters because it addresses a core challenge in drug design: creating molecules that not only bind strongly to targets but are also easy to synthesize and have low toxicity. By leveraging AI to optimize multiple properties simultaneously, MolChord reduces the risk of late-stage failures in drug development. It could streamline early-stage research, allowing scientists to explore more candidates computationally before costly lab tests.

Limitations include the model's reliance on curated datasets like CrossDocked2020, which may not cover all protein types, and the trade-off observed when aggressively optimizing for binding affinity alone, which can slightly reduce molecular diversity. Future work could expand to more diverse biological targets and incorporate real-world validation to ensure practical applicability.