Topological Feature Compression for Molecular Graph Neural Networks

Rahul Khorana · 2025

Read Topological Feature Compression for Molecular Graph Neural Networks by Rahul Khorana (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2508-07807.txt.

Description

Rahul Khorana built in 2025 a neural network for molecules that squeezes more signal from shape traits most models skip. Work on molecular traits must trade off accuracy, clarity, and cost of compute time. His design merges squeezed higher-order topological traits, which hold global shape, with plain atom-and-bond traits. The squeeze keeps the extra traits cheap to use and clear to read for working chemists.

On tests that ran from small molecules to hard materials, his model came first or near first in accuracy and robustness. It did so with few parameters to train. All code and all results are open to all. Khorana argues that topology can give chemists broad insight, not just black-box calls. That holds when it is squeezed well and shown in plain form that a human can grasp, check, and use with trust.

Catalog record

CREATORRahul Khorana
YEAR2025
COLLECTIONPAPERS
FORMATTXT
FILE SIZE80 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2508-07807

Browse the full library · About the Tower of Babel