Multimodal Multi-Granularity Fusion Model with Mamba Architecture for Ames Mutagenicity Prediction

Traditional Ames tests for chemical mutagenicity are
slow, costly,
and often yield inconsistent results between in vitro and in vivo
assays, hindering high-throughput safety screening. To address these
limitations, we propose AMPred-LWN, a multimodal multi-granularity
model that fuses atomic-level graphs, functional group sequences,
and molecular fingerprints for Ames mutagenicity prediction. Our model
integrates enhanced graph neural networks (GIN and GAT) with the Mamba-2
sequence modeling architecture and a novel bidirectional ConBiMamba
module that synchronously processes forward and reverse paths to mitigate
unidirectional biases to capture multiscale and long-range chemical
features efficiently. AMPred-LWN achieves state-of-the-art performance
on Ames data set, with AUROC of 0.922 and ACC of 0.852, outperforming
baselines and generalizing well to external sets while reducing inference
time by over 30%. Interpretability analysis shows that our model highlights
mutagenic substructures and recognizes features of non-mutagenic molecules
like polyhydroxylation, offering valuable structure–activity
insights.