Reproduce “Multimodal Mixture-of-Experts with Retrieval Augmentation for Protein Active Site Identification” on OpenResearch

About this paper

Jiayang Wu, Jiale Zhou, Rubo Wang, Xingyi Zhang, Xun Lin, Tianxu Lv, Leong Hou U, Yefeng Zheng

MERA, a framework for protein active site identification, integrates multi-expert retrieval with reliability-aware multimodal fusion to enhance prediction robustness and accuracy. The system achieved an AUPRC of 0.90 and an Fmax of 0.88 on ProTAD-Gen, improving upon existing methods by up to 7% in Fmax, and increased Hits@1 by 3-6% across benchmarks.

Key insights

• Hierarchical multi-expert retrieval, incorporating chain, sequence, and active-site perspectives, effectively extracts relevant contextual signals at the residue level, mitigating noise from broader homologous data.
• Quantifying modality trustworthiness inspired by Dempster–Shafer evidence theory provides a principled way to adaptively weigh multimodal information, ensuring unreliable modalities do not degrade overall prediction quality.
• The residue-level Mixture-of-Experts (MoE) gating within the retrieval module enables dynamic and fine-grained integration of retrieved information, crucial for handling diverse local protein contexts.

Summary by alphaXiv.