Molecules Enter the Model Before the Model Enters Chemistry

TL;DR for operators

A molecule does not enter an AI system in a neutral form. Before a model predicts a property, generates a candidate, plans a reaction, or searches for similar compounds, a team has already decided which aspects of the molecule the model can readily see.

The review by Sanjanasri JP, Pratiti Bhadra, N. Sukumar, and Soman KP makes this representation problem accessible through an NLP-oriented lens.¹ Its central finding is not that one representation wins. It is that no reviewed representation captures every structural detail required across applications.

For an AI chemistry team, that moves molecular encoding into architecture review. Compact strings integrate naturally with language-model tooling and sequence architectures, but they can introduce ambiguity or invalid generated outputs. Graph and matrix formats express connectivity more directly and can carry additional structural attributes, but they consume more memory and still require choices about what information to encode.

Cognaptus therefore reads the paper as a framework for selecting inputs against the downstream task. It does not justify claims that SELFIES, graphs, or another representation will deliver higher accuracy, lower R&D cost, or better commercial outcomes. The paper is a narrative review, not a comparative benchmark.

Representation is already a modeling decision

An AI chemistry workflow must decide how a molecule enters the model before deciding how that model will reason about it.

That choice matters because molecules are not merely sequences of atom labels. The review highlights several sources of representational difficulty: three-dimensional structure, multiple conformations, atom-labeling conventions, treatment of hydrogen, and the possibility that small structural changes produce large property changes. A machine-readable format must simplify this complexity without discarding information needed for the target task.

The paper organizes the available options into two broad families. String representations serialize molecular structure into character sequences. Graph and matrix representations encode atoms as nodes and bonds as connections. Neither family is complete.

That makes representation suitability conditional. A property-prediction system may need different structural information from a generative model. A retrosynthesis pipeline may place different demands on connectivity and sequence compatibility than a similarity-search system. The format that is convenient for one model can impose unnecessary loss or cost on another.

String formats buy NLP compatibility—and inherit sequence problems

A compact molecular string behaves enough like text to make familiar NLP machinery available. Embeddings, recurrent networks, transformers, tokenization pipelines, and transfer-learning techniques can operate on molecular sequences with relatively little conceptual modification.

SMILES is the established example: a compact text notation for molecular structure and, according to the review, a continuing cheminformatics workhorse. Its convenience comes with several failure modes. The same molecule can have multiple valid SMILES strings. Canonicalization does not remove every representational complication. More importantly for generative systems, a model can emit a sequence that is syntactically malformed or chemically invalid.

DeepSMILES modifies the syntax to reduce specific problems such as unmatched branches and paired ring-closure symbols. The review nevertheless notes remaining semantic errors, limited standardization, representation-size concerns, and constraints on applicability.

SELFIES goes further. It is a molecular string language designed around formal grammar and valence-related constraints so that generated sequences avoid the syntactic and semantic invalidity emphasized for earlier formats.

That property is operationally relevant for molecular generation. If an encoding prevents a class of invalid outputs before downstream validation, a team may reduce wasted generation and filtering work. But the paper does not measure that reduction across production systems, and it does not establish that greater validity translates into better candidate quality.

InChI addresses a different need. Its layered representation can include topology, stereochemistry, and hydrogen-related information and offers a standardized identifier. The review also reports costs: longer strings, more complex syntax, computational overhead in generation and parsing, and possible non-uniqueness in some cases.

The string family therefore contains several different engineering compromises. Treating them as interchangeable because they all look like text misses the point.

Graphs make connectivity explicit, but not representation loss disappear

A graph representation starts closer to molecular connectivity: atoms become nodes and bonds become edges. Matrix forms can encode adjacency or connectivity, while additional parameters or weights can attach further structural information.

This gives graph-based systems flexibility that a simple linear string does not provide as directly. Bond lengths, node attributes, and other task-specific information can be incorporated into the representation.

The price is computational and architectural. The review notes that graph and matrix representations can require more memory, while implementation becomes tied to decisions about which node, edge, or matrix features actually matter.

More importantly, moving from strings to graphs does not solve representation once and for all. The paper explicitly concludes that graph and matrix encodings can also omit structural detail.

That corrects a common shortcut in AI system design: richer topology is not synonymous with complete chemistry. A graph still reflects choices about what gets represented.

The downstream task determines whether the tradeoff matters

The review’s selected applications show why representation cannot be separated cleanly from model design.

Mol2Vec operationalizes the NLP analogy by treating molecular fragments and bonds as word-like units and mapping them into a learned vector space. Those embeddings can support property prediction and similarity search.

Other reviewed examples use molecular strings with recurrent neural networks for generation and transfer learning, while Graph2SMILES connects graph-based molecular inputs to sequence outputs for chemistry tasks such as retrosynthesis.

These are examples from prior work, not systems introduced or newly evaluated by this paper. Their role in the review is illustrative: the representation determines what kind of learning machinery can be applied conveniently and what information that machinery receives.

For operators, the resulting decision is more concrete than “string or graph?” A team should ask which representation best fits the required structural information, expected failure modes, computational budget, existing cheminformatics stack, and intended model family.

Cognaptus infers that this choice can propagate into validation workload and R&D architecture. A generative system that routinely produces invalid strings creates one kind of downstream burden. A graph representation that carries richer structural data but materially increases memory requirements creates another. Existing databases, tokenization tools, transfer-learning pipelines, and graph-model infrastructure can also make one option cheaper to integrate than another.

Those are architecture consequences supported by the tradeoffs described in the review. Their magnitude is not measured here.

What this review can—and cannot—settle

The paper is most useful as an introductory decision map for researchers crossing from NLP into chemistry or materials AI.

Its evidence base is a narrative synthesis of prior literature, qualitative comparison of representation formats, illustrative encodings, and selected application examples. It reports no systematic search protocol, formal quality assessment, original benchmark, ablation study, or standardized head-to-head evaluation.

That boundary matters because several tempting conclusions remain unsupported. The paper does not show that SELFIES produces better molecules than SMILES. It does not show that graphs outperform strings in property prediction. It does not estimate whether a more structurally expressive representation produces enough model improvement to justify additional compute or engineering cost.

The review instead supports a more disciplined question: what must this downstream system know about the molecule, and which representation preserves that information with acceptable computational and workflow costs?

That is a smaller claim than declaring a winner, but it is the one teams can act on without pretending the evidence is stronger than it is.

Molecular representation sits upstream of almost every subsequent model decision. Once that is recognized, encoding stops being a file-format preference and becomes part of system architecture.

Cognaptus: Automate the Present, Incubate the Future.