Compact molecular encodings cut storage by up to 80% in chemistry-AI tests

Researchers at King’s College London and partner institutions report molecular encodings that cut disk storage by up to 80% and RAM use by up to 50% in evaluated settings. The methods make chemical representations more compact while retaining their information, potentially allowing larger datasets to fit on the computing equipment a laboratory already has.[1]

The work targets a routine cost in computational chemistry. Before a model can predict a molecule’s properties, its structure must be represented numerically. Widely used encodings can be large and sparse, carrying many zeros through storage and processing.

Why chemistry turns molecules into enormous rows

A molecule cannot be fed directly into most prediction algorithms as a name such as aspirin. Chemists convert its structure into computer-readable features: a fingerprint may indicate whether certain arrangements of atoms or bonds appear; a one-hot encoding may indicate which categorical choice is present. These methods make chemical structures comparable and searchable. They also frequently produce vectors with enormous numbers of empty positions.[1]

The researchers developed embedded versions of Morgan fingerprints and one-hot encodings. King’s describes a reversible conversion to a numerical representation, designed to preserve the original information while avoiding the bit collisions that can make different chemical fragments share a fingerprint position. That is more specific than simply discarding zeros or selecting fewer chemical features.

This is conceptually distinct from reducing a molecule to a shorter description by throwing away features. Feature deletion can damage accuracy, especially when rare chemistry is important. Sparse compression instead looks for redundancy in the way the same information is represented. Exactly how large the advantage is depends on sparsity, file format, batch size and how downstream code handles decoding.

Storage, RAM and energy are three different claims

The reductions are maxima from the evaluated representations and settings. File size and working memory measure different costs, and the benefit depends on the encoding and downstream workload. A laboratory needs the result for its own dataset, rather than a universal compression percentage.[1]

Energy is a further measurement. Smaller inputs can reduce storage and transfer work, but decoding and model execution also consume power. The announcement does not establish a general percentage reduction in electricity use or carbon emissions.

For a group whose dataset exceeds available RAM, a more compact representation can change the practical options: fit it on the current machine, stream smaller portions or seek additional hardware. That makes data representation an access issue as well as an efficiency issue.

The hidden infrastructure beneath scientific breakthroughs

The story also reveals an asymmetry in AI research. New model architectures attract attention because their outputs are visible: a predicted protein interaction, a promising material or a ranked candidate drug. Representation formats rarely make the news. Yet a laboratory could improve throughput simply by packing its inputs more intelligently, independently of whether the predictive model gets any smarter.[1]

A better representation can remove a computing bottleneck before the learning algorithm changes at all. The decisive benchmark is how much useful chemistry a laboratory can run with the same memory, time and equipment.

Report an error or suggest a correction