Towards a physically interpretable symbolic language of molecular recognition

Despite decades of structural biology and molecular modelling, molecular recognition remains predominantly interpreted through static structural representations, although it is inherently dynamic. Here we represent protein-ligand recognition as a language-like statistical process arising from sequential interaction dependencies over time. This is achieved through STRIPES (Spatio-Temporal Representation of Interactions in Protein-ligand Engagement Strings), a conceptual framework that encodes molecular interaction dynamics as symbolic sequences. By discarding explicit coordinates while preserving interaction semantics, STRIPES captures recognition-relevant temporal information inaccessible to conventional structural representations. Across ~15,000 molecular dynamics trajectories, STRIPES reveals non-random sequential organization and conserved interaction motifs that transcend protein families and molecular scaffolds. This abstraction compresses highdimensional trajectories into compact symbolic sequences while preserving mechanistic information, enabling interaction-driven similarity analysis and generative molecular design. Leveraging sequence-to-sequence deep learning, we demonstrate that interaction patterns can be translated into chemically valid and bioactive ligands, including compounds with experimentally observed bioactivity. By recasting molecular recognition as an emergent symbolic process, STRIPES establishes a generalizable conceptual framework for comparing, learning, and engineering biomolecular interactions beyond static structural information.