DeepStruct: Multi-modal deep learning integrating omics and energy scoring for protein structure prediction in...

Area of Science:

• Plant biology
• Computational biology
• Structural bioinformatics

Background:

• Predicting plant protein structure and function is challenging due to limited templates and environmental stress effects.
• Current deep learning methods often overlook biological context and stress impacts on protein stability.

Purpose of the Study:

• To develop DeepStruct, a multimodal deep learning framework for contextually informed plant protein structure modeling.
• To integrate omics-derived stress signatures and energy-based scoring for improved structural prediction.

Main Methods:

• DeepStruct utilizes a hybrid CNN-BLSTM-Transformer system combining sequence embeddings, PSSMs, physicochemical descriptors, and abiotic stress omics features.
• An energy-based scoring module enhances structural plausibility with residue-level functional weights.
• The model was pre-trained on large corpora and fine-tuned on PDB and Swiss-Prot data with controlled splits.

Main Results:

• DeepStruct achieved a TM-score of 0.78 ± 0.02 and RMSD of 2.6 ± 0.3 Å on Capsicum annuum under abiotic stress.
• The model demonstrated over 80% precision for long-range contacts.
• Incorporating stress-responsive omics features significantly improved TM-score by up to 12% (p < 0.05).

Conclusions:

• Contextual biological signals, particularly stress signatures, significantly enhance plant protein structure prediction.
• DeepStruct offers a novel approach for modeling proteins conditioned by environmental stress, complementing sequence-based methods.
• This framework advances our understanding of plant protein behavior under environmental challenges.