Designing a new drug is often compared to finding a needle in a haystack, but the "haystack" of possible molecules is estimated at $10^{60}$ candidates—more than the number of atoms in the solar system. To navigate this space, researchers are moving away from building molecules atom-by-atom, instead adopting a "fragment-based" approach that assembles molecules from larger, chemically meaningful building blocks.
However, a recurring challenge across recent literature is the vocabulary bottleneck: if a generative model only knows a fixed set of "bricks," it can never build a truly novel "house." Current research is focused on breaking this constraint through dynamic fragment discovery, hierarchical modeling, and motif-agnostic generation.
The Shift from Atoms to Motifs
Traditional generative models, such as those described by Joongwon Lee in FragFM, often operate at the atomic level. While precise, this approach is computationally expensive and prone to generating "spaghetti" structures—long, floppy chains that are chemically unstable or impossible to synthesize.
The intuition behind the fragment-based shift is simple: it is easier to build a functional machine using pre-assembled gears and levers than by smelting individual screws from raw ore. By using "structural motifs"—frequently occurring multi-atom substructures—models like the one developed by Wengong Jin in Hierarchical Generation of Molecular Graphs using Structural Motifs can ensure 100% chemical validity and significantly improve the reconstruction of large, complex polymers.
The Vocabulary Bottleneck
Despite their efficiency, fragment-based models face a fundamental limitation. As Leon Hetzel points out in MAGNet: Motif-Agnostic Generation of Molecules from Shapes, these models are typically restricted by a predefined vocabulary of chemical building blocks.
If a model's library only contains benzene rings and simple chains, it will never "invent" a macrocycle or a novel fused-ring system. This dependency creates several issues:
- Limited Novelty: The model can only recombine existing motifs, preventing the discovery of entirely new chemical scaffolds TacoGFN.
- Rigidity: Introducing a new chemical element or functional group often requires retraining the entire model from scratch Jaechang Lim.
- Rare Structure Failure: Models often fall back on atom-by-atom construction for rare fragments, which frequently leads to invalid results MAGNet.
Beyond Static Libraries
To overcome this, researchers are developing methods that treat fragments as dynamic entities rather than static tokens.
Dynamic Discovery
Instead of using a fixed list, models like GEAM, introduced by Seul Lee in Drug Discovery with Dynamic Goal-aware Fragments, use genetic algorithms to discover and modify fragments on-the-fly. This allows the model to evolve its own building blocks based on the specific biological goal it is trying to achieve, such as high binding affinity to a protein.
Motif-Agnostic Scaffolding
Another strategy is to separate the "shape" of the molecule from its "identity." In MAGNet: Motif-Agnostic Generation of Molecules from Scaffolds, the authors propose generating an abstract geometric skeleton first and then filling in the specific atoms and bonds. This "motif-agnostic" approach allows the model to explore structural arrangements that don't exist in its training data.
Learned Distributional Vocabularies
Frameworks like FaST, developed by Xiang Fu in Fragment-based Sequential Translation for Molecular Optimization, replace static libraries with learned vocabularies. By using reinforcement learning to decide which substructures to add or delete, these models navigate chemical space more efficiently than those restricted to a fixed alphabet.
Integrating 3D Geometry and Biological Context
The ultimate goal of these generative models is to design molecules that work in the real world. This requires moving beyond 2D graphs into 3D space and biological systems.
Odin Zhang and colleagues introduced FragGen, which uses a "Deep Geometry Handling" protocol to ensure that fragment-based generation results in physically realistic 3D structures. This approach led to the successful wet-lab validation of a nanomolar-level kinase inhibitor, proving that fragment-based AI can produce high-quality drug candidates.
Furthermore, models are now being "conditioned" on biological data. Bonil Koo demonstrated in GGIFragGPT how transcriptomic signatures (the "fingerprints" of cellular activity) can act as prompts for fragment-based GPT architectures, tailoring molecule generation to induce specific biological responses.
Conclusion
The evolution of molecular generation is a transition from fine-grained but chaotic atomic assembly to coarse-grained but constrained fragment assembly, and finally toward flexible, learned substructure modeling. By breaking the vocabulary bottleneck, researchers are enabling AI to not just remember the chemistry it has seen, but to invent the chemistry we need.