Traditional generative models for chemistry often treat molecules like strings of text or clouds of individual points. To build a molecule, these models add one atom at a time, painstakingly deciding the element and position of every carbon, oxygen, and nitrogen.
However, this "atom-by-atom" approach faces a fundamental scaling problem. In a 30-atom molecule, a model must make dozens of perfect sequential decisions; a single misplaced atom can render the entire structure chemically impossible or synthetically unreachable. To solve this, a growing body of research has shifted the "unit of generation" from the individual atom to the molecular fragment.
The Curse of Atom-by-Atom Generation
When a model builds a molecule atom-by-atom, it must learn the laws of chemistry from scratch. It has to "discover" that carbon usually forms four bonds, that certain rings are strained and unstable, and that specific functional groups are required for biological activity.
As noted by Yuyang Wang in Graph Neural Networks for Molecules, atom-level models often struggle to capture the realism of complex functional groups. This leads to "hallucinated" structures—molecules that look plausible to a neural network but are physically unstable or impossible to synthesize in a laboratory. Furthermore, the search space for individual atoms is astronomical. By the time a model has placed 20 atoms, the number of possible configurations has exploded, making optimization via reinforcement learning or diffusion incredibly difficult.
The Fragment as a Chemical Primitive
Fragment-based generation changes the game by using pre-validated building blocks—rings, functional groups, and common medicinal scaffolds—as the primary "LEGO bricks" of construction.
The seminal Junction Tree Variational Autoencoder (JT-VAE), developed by Wengong Jin, Regina Barzilay, and Tommi Jaakkola, demonstrated that by decomposing molecules into a "junction tree" of chemically valid substructures, models could achieve 100% chemical validity. Because the building blocks themselves are valid, the model only needs to learn how to connect them correctly.
This approach drastically reduces the "computational path length." As Yulun Wu and colleagues show in their work on DGAPN, using fragments instead of atoms makes reinforcement learning more efficient because the agent takes fewer, more meaningful steps to reach a finished drug candidate.
Preserving Geometry through Rigid Assembly
The challenge becomes even steeper in 3D space, where models must predict precise spatial coordinates. Standard diffusion models often "tangle" molecules, creating unrealistic bond lengths that require expensive energy minimization to fix.
Fragment-based models bypass this by treating substructures as rigid bodies. In SQUID, Keir Adams and Connor W. Coley use an equivariant fragment-based decoder to ensure local geometric validity. Since the fragments (like a benzene ring) are rigid, their internal bond lengths and angles are fixed and perfect by definition. The model's only job is to find the correct relative orientation and rotatable bond angles.
This "coarse-to-fine" philosophy is pushed further by HierDiff, a hierarchical diffusion model by Bo Qiang and Yuxuan Song. It first generates the arrangement of fragments (the "coarse" level) and then refines the atomic positions (the "fine" level). This ensures that the global shape of the molecule is established before the model worries about individual atom placements.
Bridging the Gap to Synthesis
Perhaps the most practical advantage of fragments is synthetic accessibility. A molecule is only useful if a chemist can actually make it.
By using fragments derived from known chemical reactions or established catalogs, models like FragGen and 3D-MCTS ensure that the generated outputs are "born" from realistic precursors. Odin Zhang and Xujun Zhang demonstrated this by using fragment-wise assembly to discover a 75.4 nM kinase inhibitor that was successfully validated in a wet lab.
This logic extends to materials science as well. In the design of Metal-Organic Frameworks (MOFs), Xiang Fu and Tian Xie developed MOFDiff, which performs diffusion on coarse-grained building blocks. Similarly, researchers at MIT led by Aditya Nandy showed that reassembling fragments from experimentally stable MOFs leads to a 10-fold enrichment in "ultrastable" new materials.
Fragments as a Chain of Thought
The latest frontier in this area treats fragments not just as physical blocks, but as semantic units for Large Language Models (LLMs).
In FragGPT, Jie Yue and colleagues move away from character-based SMILES strings to FU-SMILES, a representation based on unordered fragments. This mitigates "exposure bias"—the tendency of LLMs to drift into nonsense after a few wrong characters—because each token represents a meaningful chemical unit.
As Li Hao argues in Navigating Chemical-Linguistic Sharing Space, fragments act as a "Chain of Thought" for the model. By planning the arrangement of functional groups before committing to an atomic structure, these models achieve a level of chemical realism that atom-level models simply cannot match.