TLDR: This research paper reviews the application of diffusion models in drug discovery, specifically comparing their use in generating small molecules versus therapeutic peptides. It highlights how these generative AI models adapt to distinct molecular representations and design objectives, excelling in structure-based design for small molecules and functional sequence/structure design for peptides. The paper discusses modality-specific challenges like synthesizability for small molecules and biological stability for peptides, as well as shared hurdles such as imperfect scoring functions and data scarcity. It concludes by emphasizing the need for integrated Design-Build-Test-Learn (DBTL) platforms to realize the full potential of diffusion models in creating novel therapeutics.
Drug discovery has long been a challenging and expensive endeavor, often taking over a decade and billions of dollars to bring a new medicine to market. The traditional methods, relying on high-throughput screening and combinatorial chemistry, struggle to explore the immense chemical space of potential drug-like molecules. However, a new era is dawning with the rise of generative Artificial Intelligence (AI), particularly diffusion models, which promise to transform this process by designing entirely new molecules tailored to specific needs.
A recent review, titled Diffusion Models at the Drug Discovery Frontier: A Review on Generating Small Molecules versus Therapeutic Peptides, by Yiquan Wang, Yahui Ma, Yuhan Chang, Jiayao Yan, Jialin Zhang, Minnuo Cai, and Kai Wei, provides a comprehensive comparison of how these powerful AI models are being applied to design two crucial types of therapeutics: small molecules and therapeutic peptides.
The Core of Diffusion Models
At their heart, diffusion models operate through a two-step process. First, a ‘forward diffusion’ process gradually adds noise to a data structure until it becomes pure noise. Then, a ‘reverse denoising’ process, guided by a trained neural network, learns to iteratively remove this noise, effectively generating new, high-quality data from random starting points. This framework is incredibly versatile, capable of handling both continuous data, like the 3D coordinates of atoms, and discrete data, such as atom types or amino acid sequences.
What makes diffusion models particularly exciting for drug design is their ability to perform ‘conditional generation.’ This means the models can be steered towards specific objectives by incorporating information like a target protein’s binding pocket geometry or desired physicochemical properties. This precise control allows for the multi-objective optimization crucial in developing new drugs, overcoming limitations faced by earlier generative AI models like VAEs and GANs, which often struggled with stability and generation quality.
Designing Small Molecules
For small molecules, which make up a significant portion of approved drugs, diffusion models excel at structure-based drug design (SBDD). This involves generating molecules that are geometrically and chemically complementary to a protein’s binding pocket, aiming to maximize binding affinity. Models like Pocket2Mol, DiffSBDD, and TargetDiff have shown impressive results, generating novel molecules with favorable predicted binding affinities and good drug-likeness. They can even generate molecules that form specific interactions, like hydrogen bonds, with critical residues in the protein pocket.
Another key application is property-based ligand design, where models generate molecules that meet multiple criteria simultaneously, such as high binding affinity, low toxicity, and good synthetic accessibility. While these models can generate chemically valid molecules, a significant hurdle remains: ensuring that these computationally designed molecules can actually be synthesized in a lab. Bridging this gap between computational design and practical synthesis is a major focus for future research.
Innovating Therapeutic Peptides
Therapeutic peptides represent a rapidly growing class of drugs, offering high specificity and potency for targets often challenging for small molecules. Diffusion models are being used to generate functional peptide sequences, such as antimicrobial peptides (AMPs) or cell-penetrating peptides (CPPs). Discrete diffusion models learn to create novel sequences with experimentally validated activity, often showing high diversity and validity compared to previous methods.
Even more ambitious is structure-guided de novo peptide design, where models generate peptides that fold into specific 3D structures or bind to target protein surfaces. Landmark models like RFdiffusion can generate protein backbones that, when combined with sequence design tools, achieve high experimental binding affinities. However, peptides face their own unique challenges, including ensuring biological stability against degradation, proper folding, and minimizing immunogenicity in the body. Integrating these complex biological constraints into the design process is a critical frontier.
Also Read:
- MOLCHORD: A Unified Framework for Protein-Guided Drug Design
- Unlocking Deeper Patterns: Fractional Diffusion Bridge Models in AI
Shared Challenges and Future Directions
Despite their distinct applications, both small molecule and peptide design with diffusion models face common obstacles. A major issue is the reliance on imperfect computational scoring functions, which often don’t perfectly correlate with real-world experimental results. This is compounded by the scarcity of high-quality experimental data that pairs molecular structures with validated biological activity. To truly unlock the potential of diffusion models, the field needs to move towards a ‘Design-Build-Test-Learn’ (DBTL) cycle, integrating AI-powered design with automated laboratory validation to create a self-optimizing discovery engine.
The future of diffusion models in drug discovery is bright, with opportunities to develop unified ‘foundation models’ that can design a wide range of therapeutics. Enhancing model reliability, making them more interpretable, and integrating physics-based simulations will be crucial. Ultimately, by overcoming these challenges and fully embracing automated DBTL paradigms, diffusion models are poised to fundamentally shift drug discovery from passive exploration to the active, purpose-driven creation of novel medicines.
