Fragment-SELFIES

Definition

Fragment-SELFIES is a fragment-level molecular string representation in the SELFIES family. Instead of emitting only atom-level tokens, a model emits chemically meaningful fragments and attachment structure, which can shorten sequences and make generation closer to medicinal-chemistry building blocks.

Why It Matters

In molecular generation, standard SMILES can produce invalid strings, and atom-level SELFIES guarantees validity but generates long sequences for large molecules. Fragment-SELFIES offers a middle ground: shorter sequences with chemically interpretable tokens and maintained validity guarantees.

It is useful to treat Fragment-SELFIES as a representation choice, not as a complete model. A paper using it still needs a clear task, objective, data split, novelty rule, and chemical-state policy.

Modeling Contract

Let a molecule $M$ be decomposed into fragments:

$$M\to (f_1,\dots ,f_n,A)$$

where $f_i$ are fragment tokens and $A$ records attachment structure. A valid decoder must map the token sequence back to a chemically valid molecular graph:

$$\mathrm{decode}(f_1,\dots ,f_n,A)\in {\mathcal{G}}_{\mathrm{valid}}$$

The hard part is not only syntactic validity, but whether the fragment vocabulary covers the target chemistry without hiding stereochemistry, charge, tautomer, or rare functional groups.

Representation Boundary

Fragment tokenization changes what the model can easily learn:

Boundary	Meaning	Risk
Fragment vocabulary	allowed building blocks	rare chemistry may be unreachable
Attachment grammar	how fragments connect	ambiguous attachment can create invalid or unintended graphs
Chemical state	stereo, charge, aromaticity, tautomer, protonation	validity can hide state loss
Sequence policy	canonical, randomized, or augmented encoding	evaluation may reward tokenization artifacts
Reconstruction rule	how tokens become a molecular graph	decoder validity is not the same as usefulness

For a generative model, the effective hypothesis space is:

$${\mathcal{M}}_{\mathrm{reachable}}=\{\mathrm{decode}(z):z\in {\mathcal{Z}}_{\mathrm{frag}}\}$$

If the target chemistry is outside ${\mathcal{M}}_{\mathrm{reachable}}$, model quality is capped by representation coverage before architecture matters.

Evaluation Contract

Report generation metrics against the molecular object, not only the token string:

Metric Family	Check
Validity	decoded molecule is graph-valid and chemically sane under the chosen state policy
Uniqueness	duplicate handling is done after standardization
Novelty	generated molecules are compared against training molecules and near-duplicate fragment recombinations
Diversity	scaffold, fingerprint, or fragment diversity is reported with the chosen representation
Utility	property, docking, or synthesis proxy is evaluated under an explicit benchmark protocol

Checks

• Is the fragment vocabulary fixed, learned, or dataset-specific?
• Does tokenization preserve stereochemistry, charge, isotopes, and aromaticity?
• Are attachment points unambiguous?
• Does validity guarantee imply chemically useful molecules or only graph-valid molecules?
• Is novelty measured against fragment recombination as well as full-molecule identity?
• Are standardization, scaffold split, and duplicate policy applied after decoding?
• Is the generated distribution evaluated against the same chemical-state policy used for training?

Related Papers

• Molexar

Related Concepts

• SMILES
• Molecular graph
• Molecular standardization
• Molecular identity
• Molecular similarity
• Autoregressive model
• Molecular generation
• Scaffold split

Open Questions

• How does generation quality compare to atom-level SELFIES for property-conditioned tasks?
• Is there a standard fragment vocabulary, or is it task-dependent?
• Does fragment-level tokenization lose fine-grained control over individual atom placement?