Abstract:Deep learning has revolutionized the field of protein structure prediction. AlphaFold2, a deep neural network, vastly outperformed previous algorithms to provide near atomic-level accuracy when predicting protein structures. Despite its success, there still are limitations which prevent accurate predictions for numerous protein systems. Here we show that sparse residue burial restraints from deep mutational scanning (DMS) can refine AlphaFold2 to significantly enhance results. Burial information extracted from DMS is used to explicitly guide residue placement during structure generation. DMS-Fold was validated on both simulated and experimental single-mutant DMS, with DMS-Fold outperforming AlphaFold2 for 89% of protein targets and with 253 proteins having an improvement greater than 0.1 in TM-score. DMS-Fold is free and publicly available: https://github.com/LindertLab/DMS-Fold.

What problem does this paper attempt to address?

This paper aims to address the limitations in protein structure prediction. Although AlphaFold2 has achieved great success in protein structure prediction, it still faces challenges in dealing with dynamic proteins, predicting mutation effects, orphan proteins or disordered proteins, etc. For this reason, the paper proposes DMS - Fold, a deep - learning model that combines single - point mutation deep mutational scanning (DMS) data to improve the prediction results of AlphaFold2. Specifically, the main contributions of the paper include: 1. **Extract buried information**: Analyze large - scale single - point mutation DMS data to extract the buried information of protein residues. This information can reflect the relative position of residues inside the protein, thus guiding the generation of protein structures. 2. **Develop DMS - Fold**: Based on the AlphaFold2 framework, develop a new deep neural network DMS - Fold. This model embeds the buried information in DMS data into the pair representation of AlphaFold2 to guide the position arrangement of residues, thereby improving the prediction accuracy. 3. **Verify the effectiveness of DMS - Fold**: Verify the performance of DMS - Fold through simulation and experimental data. The results show that DMS - Fold outperforms AlphaFold2 on 89% of protein targets, and the TM - score of 253 protein targets has increased by more than 0.1. 4. **Explore the influence of MSA**: Study the influence of different levels of MSA sampling on the prediction results of DMS - Fold and AlphaFold2. It is found that in the case of shallower MSA, DMS - Fold is more dependent on buried information and thus shows better performance. In conclusion, this paper proposes a new protein structure prediction model DMS - Fold by combining DMS data and deep - learning methods, which significantly improves the accuracy of protein structure prediction, especially when dealing with some complex protein systems.

Deep-Learning Structure Elucidation from Single-Mutant Deep Mutational Scanning

Dissecting AlphaFolds Capabilities with Limited Sequence Information

Effective High-Accuracy Prediction of Protein Structures from Easily Obtainable Artificial Homologous Sequences by Structure-Stability-Based Selection

Topological deep learning based deep mutational scanning

Improving deep learning protein monomer and complex structure prediction using DeepMSA2 with huge metagenomics data

AF2-Mutation: Adversarial Sequence Mutations against AlphaFold2 on Protein Tertiary Structure Prediction

Single-sequence protein structure prediction using a language model and deep learning

So ManyFolds, So Little Time: Efficient Protein Structure Prediction With pLMs and MSAs

Enhancing the Protein Tertiary Structure Prediction by Multiple Sequence Alignment Generation

Ins and outs of AlphaFold2 transmembrane protein structure predictions

AlphaFold2 transmembrane protein structure prediction shines

SPEACH_AF: Sampling protein ensembles and conformational heterogeneity with Alphafold2

MoDAFold: a strategy for predicting the structure of missense mutant protein based on AlphaFold2 and molecular dynamics

AlphaFold 2: Why It Works and Its Implications for Understanding the Relationships of Protein Sequence, Structure, and Function

Deep Learning-Based Advances in Protein Structure Prediction

Structure-based, deep-learning models for protein-ligand binding affinity prediction

Unsupervisedly Prompting AlphaFold2 for Few-Shot Learning of Accurate Folding Landscape and Protein Structure Prediction

AlphaFold3, a secret sauce for predicting mutational effects on protein-protein interactions

DeepDriveMD: Deep-Learning Driven Adaptive Molecular Simulations for Protein Folding

High-accuracy protein structures by combining machine-learning with physics-based refinement

AlphaFold2 in biomedical research: facilitating the development of diagnostic strategies for disease