Structure-Aware Fusion with Progressive Injection for Multimodal Molecular Representation Learning
Zihao Jing, Yan Sun, Yan Yi Li, Sugitha Janarthanan, Alana Deng, Pingzhao Hu · 2025
Read Structure-Aware Fusion with Progressive Injection for Multimodal Molecular Representation Learning by Zihao Jing, Yan Sun, Yan Yi Li, Sugitha Janarthanan, Alana Deng, Pingzhao Hu (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-2510-23640.txt.
Description
Zihao Jing and five co-authors present MuMo, a framework that learns molecule representations from both 2D structure and 3D shape. Multimodal models promise richer descriptions for drug discovery, but two problems dog them. Predicted 3D conformers are unreliable, and naive fusion lets one data type swamp the other until the model collapses onto it. Both flaws hurt robustness on new molecules.
MuMo first merges the 2D topology and 3D geometry into one structural prior, which softens harm from shaky conformers. An injection step then feeds that prior into the sequence stream unevenly, enriching it while each data type keeps its own modeling. Built on a state space backbone, the system handles long-range links well. Across twenty-nine tasks, MuMo beat the best rival by about three percent on average. It ranked first on twenty-two tasks, including a twenty-seven percent jump on a toxicity task.
Catalog record
| CREATOR | Zihao Jing, Yan Sun, Yan Yi Li, Sugitha Janarthanan, Alana Deng, Pingzhao Hu |
|---|---|
| YEAR | 2025 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 167 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2510-23640