Rob Owen’s presentation explores Pfizer’s journey in managing chemical data formats, beginning with the transition from V2000 to V3000 and the interoperability challenges this created—particularly around stereochemistry, reactions, and degeneracy in molecular representations. It details the company’s strategic decision to maintain compatibility with both formats, the reliance on Chemdraw interpretations, and the move toward CXSMILES to address the “lossy” nature of SMILES and InChI. The shift toward web-based, text-friendly formats like CDXML, the complexities of copy-paste and reaction handling, and the need for consistent rendering across multiple toolkits are emphasized. Owen advocates for open standards, broader vendor support, and focusing on functionality rather than file-format lock-in, all while acknowledging the evolving role of SaaS solutions and the importance of enabling chemists with flexible, interoperable tools.

The challenges of data exchange in a FAIR world

In his presentation, Gerd Blanke of StructurePendium Technologies GmbH addresses the persistent challenge of interoperability in cheminformatics, focusing on the critical role of reliable chemical exchange formats for structures and reactions. He explains that existing formats are often lossy, leading to data quality issues, higher costs, and unpredictable workloads—especially during database mergers where discrepancies and legacy structures emerge. Poorly implemented formats undermine FAIR principles and hinder AI readiness. Blanke calls for regular cross-industry dialogue among data engineers, vendors, and format owners to share experiences, identify limitations, and agree on standardized improvements, positioning the Pistoia Alliance as an ideal forum to coordinate these efforts.

This presentation by Ramil Nugmanov addresses the challenge of maintaining linear compute scaling when working with ever-growing datasets in AI-assisted drug discovery, particularly for DNA-encoded libraries (DELs) containing billions of molecules. By breaking combinatorial libraries into fragments and using SMILES concatenation with placeholder atoms, the method avoids storing every entity, reducing memory use from gigabytes to ~10 MB and compute time to minutes. Search efficiency is improved by fragmenting queries, reconstructing fingerprints on the fly, and using bitwise operations to calculate Tanimoto similarity without full molecule reconstruction. The approach enables rapid similarity search, efficient CPU cache usage, and parallelization, while avoiding deep neural architectures that are inefficient for such data. The key message: don’t apply standard solutions blindly—design fragment-based, resource-efficient methods for massive chemical spaces.