| Tartalmi leírás: |
Canonical SMILES representations are ubiquitous in molecular machine learning, yet the algorithmic rules that produce them introduce systematic positional patterns unrelated to molecular chemistry. We present a comprehensive statistical analysis demonstrating that canonical SMILES contain significant positional artifacts: 29% higher mutual information I(position; token) compared to random SMILES, dramatically elevated position-topology correlations (up to 17.5× higher for ring membership), and algorithm-specific fingerprints enabling 65% classification accuracy for identifying which canonicalizer produced a given string. These findings suggest that models trained exclusively on canonical SMILES may learn positional shortcuts rather than genuine chemical representations, with implications for generalization and robustness. We propose random SMILES augmentation and representation-invariant evaluation as standard practice for molecular ML.
|
