SMILES Equivalence in Molecular ML

• SMILES Equivalence is the property where different SMILES strings represent the same molecular graph, ensuring consistent chemical representation.
• Enumeration techniques generate diverse, non-canonical SMILES that bolster neural network training and improve molecular property predictions.
• Empirical tests show that enforcing SMILES-Eq reduces errors and increases model robustness in QSAR and deep learning applications.

SMILES Equivalence (SMILES-Eq) denotes the property that different SMILES (Simplified Molecular Input Line Entry System) strings can encode identical molecular graphs, making them functionally interchangeable in computational chemistry and cheminformatics. In modern chemistry LLMs and LLMs for molecular tasks, effective utilization and explicit enforcement of SMILES-Eq are essential for representing molecular invariance, ensuring robust learning, and advancing property prediction, generative chemistry, and molecular informatics workflows.

1. Formal Definition and Theoretical Basis

SMILES-Eq is formally characterized via the mapping between SMILES strings and molecular graphs. Let $S$ be a valid SMILES string and $G(S) = (V, E, \ell, \beta)$ the induced molecular graph, where $V$ is the set of atoms, $E$ the bond set, $\ell$ atom labels, and $\beta$ bond labels. Two SMILES, $S_1$ and $S_2$, are equivalent if and only if $G(S_1)=G(S_2)$, i.e., they produce isomorphic, identically-labeled graphs. Thus, the equivalence relation is

$S_1 \equiv S_2 \iff G(S_1) = G(S_2)$

SMILES-Eq extends this classical definition to model architectures: different SMILES for the same graph should result in identical or near-identical model embeddings and property predictions, enforcing invariance with respect to token ordering, canonical representation, and serialization ambiguities (Jang et al., 22 May 2025, Park et al., 8 Dec 2025). In practice, this invariance is highly desirable, as a molecule may possess hundreds of unique SMILES due to reordering, ring and branch notations, and atom indexing.

2. SMILES Enumeration and Data Augmentation

Operationalizing SMILES-Eq for machine learning begins with SMILES enumeration, the process of generating many unique, non-canonical SMILES strings for a given molecular graph. This is often implemented using algorithms in packages such as RDKit:

function EnumerateSMILES(G, N): S ← ∅ for i in 1…N do G′ ← RandomPermuteAtoms(G) s ← MolToSmiles(G′, canonical=False) if s ∉ S then S ← S ∪ {s} end if end for return S end function

The expected number of unique SMILES after $N$ random permutations, if there are $m$ realizable orderings, is

$$E[|S|] = m\left(1 - \left(1 - \frac{1}{m}\right)^N\right)$$

In a QSAR modeling example, average enumeration per molecule reached $|S| \approx 130$, yielding a dataset augmentation factor of approximately $131.5\times$ for the training corpus (Bjerrum, 2017). This approach amplifies data diversity and exposes models to all possible tokenizations of identical chemistry, which is critical for enforcing SMILES-Eq in learning.

3. Neural Architectures and Representation Invariance

Sequence-based neural networks (LSTM, transformers, CLMs) must be invariant to SMILES permutations to guarantee consistent property predictions. LSTM-based QSAR models process SMILES as character sequences, typically with one-hot encoding and recurrent updates: $$i_t = \sigma(W_ix_t + U_ih_{t-1} + b_i), \quad ... \quad h_t = o_t \odot \tanh(c_t)$$ After processing, the final state $h_T$ is passed to a regression head. Training on enumerated SMILES per molecule, using dropout, L1/L2 regularization, and random orderings, ensures the network's internal representations do not overfit to a single serialization.

Empirical results demonstrate that enumeration improves model accuracy:

• Test $R^2$ increased from $0.56$ (canonical only) to $0.66$ (enumerated), with ensemble predictions over multiple SMILES further raising $R^2$ to $0.68$.
• RMS error fell from $0.62$ to $0.55$ using enumeration, and to $0.52$ with ensemble averaging (Bjerrum, 2017).

Foundation models for polymers, such as those using the CMDL Polymer Graph (CPG) representation, confirm the resilience of CLMs to even semantically or chemically invalid SMILES variants, as nearly all variants yield similar embeddings and prediction accuracy ($|\Delta \mathrm{RMSE}| < 0.1$ for most tested perturbations) (Park et al., 8 Dec 2025).

4. Empirical Testing and Control Experiments

Rigorous evaluation of SMILES-Eq in deep molecular models requires control experiments at both architectural and data representation levels. Techniques include:

• Generating atom-substituted or fully randomized SMILES variants (including semantically invalid forms).
• Measuring downstream regression/classification performance (e.g., via $\Delta \mathrm{RMSE}$) and attention-map divergence (RMSD).
• Embedding equivalence is typically assessed by training XGBoost regressors on latent representations and comparing performance across representations. Consistently small differences indicate empirical invariance.

For instance, in polymer foundation models, dielectric constant RMSEs are nearly identical for original and heavily perturbed CPG representations, confirming SMILES-Eq in the learned latent space. Attention pattern analyses reveal that fine-tuning smooths reliance on individual token artifacts, resulting in models that interpolate over sequence space rather than literal chemistry (Park et al., 8 Dec 2025).

5. Model Training Protocols for SMILES-Eq

Pretraining protocols which enforce SMILES-Eq deploy SMILES parsing as a multi-task curriculum:

• Subgraph matching: verifying presence/counts of functional groups, rings, and chains.
• Global graph matching: SMILES canonicalization (outputting can($S$) from $S$), and fragment assembly. Success in these tasks requires (a) canonicalization of arbitrary SMILES, and (b) matching subgraphs across different serializations. High accuracy in both categories (>0.92 for canonicalization and >0.88 for assembly) is achievable after dedicated multitask pretraining (Jang et al., 22 May 2025).

Curriculum learning and adaptive difficulty scoring optimize task presentation; data pruning ensures models do not overfit trivial representations.

6. Implications, Best Practices, and Limitations

SMILES-Eq supports robust, generalizable chemical property prediction and model interpretability:

• For neural QSAR and graph-to-sequence models, enumerative augmentation improves both $R^2$ and RMS, while ensemble averaging stabilizes inferences.
• For CLMs and foundation models, explicit testing with random and invalid representations provides assurance against artifact-driven interpolation.
• Best practices include balancing enumeration counts, normalizing sequence lengths, and benchmarking against both valid and invalid SMILES variants.
• For out-of-distribution performance, supplementing random splits with scaffold or time-splits is essential.

A plausible implication is that further increases in model size and pretraining corpus diversity may reinforce SMILES-Eq to the extent that fine distinctions between chemical graphs and serialization artifacts are universally suppressed, highlighting the necessity for control experiments during foundation model validation (Park et al., 8 Dec 2025).

7. Summary Table: SMILES-Eq Results Across Studies

Setting	Key Experiment/Method	Result/Metric
LSTM-QSAR, DHFR dataset (Bjerrum, 2017)	Enumerated SMILES, ensemble avg.	$R^2$: $0.68$, RMS: $0.52$ (test set, ensemble)
Polymer foundation (Park et al., 8 Dec 2025)	CPG, atom-substituted, randomized	$\Delta \mathrm{RMSE}$ typically $< 0.1$
CleanMol LLM parsing (Jang et al., 22 May 2025)	Canonicalization, assembly accuracy	$\sim$0.93–0.95 on canonicalization; $\sim$0.88–0.90 on assembly

In sum, SMILES Equivalence is both a fundamental formal property and a critical operational concept for modern molecular machine learning, enabling models to generalize across the multitude of possible textual representations of chemical graphs and to avoid artifacts of SMILES line notation in embedding, generation, and property prediction.