Core Concepts
Incorporating structural information through a simple, efficient, and scalable adapter architecture can significantly improve the performance and training efficiency of protein language models across diverse downstream tasks.
Abstract
The paper introduces SES-Adapter, a model-agnostic adapter architecture that integrates protein language model (PLM) embeddings with structural sequence embeddings through cross-modal fusion attention. The key highlights are:
SES-Adapter can be applied to various PLM architectures, including ESM2, ProtBert, ProtT5, and Ankh series, and is evaluated on 9 benchmark datasets across 4 downstream tasks: protein localization prediction, solubility prediction, function prediction, and annotation prediction.
Extensive experiments show that SES-Adapter outperforms vanilla PLMs, with a maximum performance increase of 11% and an average of 3%. It also significantly accelerates training speed by up to 1034% and an average of 362%, and improves convergence efficiency by approximately 2 times.
The serialization strategy using FoldSeek and DSSP effectively mitigates potential prediction errors and is insensitive to structural quality, as verified by comparative tests using structures folded by ESMFold and AlphaFold2.
Ablation studies confirm the contributions of each component in the SES-Adapter design, including the FoldSeek sequence, DSSP sequence, and rotary positional encoding (RoPE).
The SES-Adapter demonstrates superior performance compared to other state-of-the-art hybrid models that combine sequence and structure information, such as MIF-ST, ESM-GearNet, and SaProt-GearNet.
Overall, the SES-Adapter provides a simple, efficient, and scalable approach to enhance the representational quality of PLMs and improve their performance on diverse downstream tasks, while being robust to potential errors in predicted protein structures.
Stats
The training dataset for DeepSol is 6 times larger than DeepSoluE.
The pLDDT score difference between AlphaFold2 and ESMFold structures is up to 10 for some datasets.
Quotes
"Incorporating structural information through a simple, efficient, and scalable adapter architecture can significantly improve the performance and training efficiency of protein language models across diverse downstream tasks."
"The serialization strategy using FoldSeek and DSSP effectively mitigates potential prediction errors and is insensitive to structural quality, as verified by comparative tests using structures folded by ESMFold and AlphaFold2."