B-factor prediction in proteins using a sequence-based deep learning model

Akash Pandey, Elaine Liu, Jacob Graham, Wei Chen, Sinan KetenPatterns 2023 · DOI

TL;DR

This paper introduces a novel deep learning model that predicts protein B-factors solely from their amino acid sequence, achieving state-of-the-art accuracy. This sequence-based approach overcomes limitations of previous methods and offers a powerful tool for understanding protein dynamics and designing new proteins.

What Problem Does This Paper Address?

Proteins are dynamic molecules, and their flexibility and movement are crucial for their function. The B-factor (or Debye-Waller factor) is a key experimental measure of an atom's displacement from its mean position, providing insights into a protein's flexibility, thermal stability, and activity. However, accurately predicting B-factors has been a significant challenge. Previous methods often relied on complex physics-based models requiring detailed structural information, which are not always available and don't generalize well. Machine learning approaches have shown promise, but often struggle to capture the influence of neighboring atoms effectively, leading to limited predictive power, especially on unseen data. This paper aims to develop a more robust and generalizable method for B-factor prediction that leverages the power of deep learning and relies only on the protein's primary amino acid sequence.

Methods

The core of this work is a sequence-based deep learning model utilizing a bidirectional Long Short-Term Memory (LSTM) network. Unlike traditional methods that might require 3D structural information, this model takes the primary amino acid sequence as input. The input sequence is first processed by an encoder that transforms each amino acid's information (including its type, secondary structure, and 3D coordinates) into a vector representation. This vector sequence is then fed into the bidirectional LSTM. The bidirectional nature of the LSTM is crucial, as it allows the model to capture dependencies from both the beginning and the end of the sequence simultaneously, effectively accounting for long-range interactions within the protein. Finally, a decoder translates the LSTM's output into the predicted B-factor for each atom. The model is trained using a mean squared error (MSE) loss function and optimized via backpropagation.

Datasets

The study utilizes data extracted from the Protein Data Bank (PDB). The dataset was curated by excluding proteins with extreme B-factor values (above 80), unphysical B-factors (less than or equal to 0), or identical B-factors across all residues. Additionally, proteins exceeding 500 amino acids were excluded to manage computational costs. The final dataset comprised approximately 61,000 proteins, which were split into:

  • 56,000 proteins for training.
  • 2,400 proteins for validation.
  • 2,400 proteins for testing.

This represents the widest dataset used for B-factor prediction to date, enabling a robust assessment of the model's generalizability.

Key Insights and Ideas

  • Primary Sequence Sufficiency: The most striking insight is that the primary amino acid sequence alone, when processed by a sufficiently powerful deep learning model like LSTM, is remarkably effective for predicting B-factors. This suggests that the sequence implicitly encodes crucial structural and dynamic information.
  • Neighboring Atom Influence: The model implicitly learns that atoms within a certain radius (around 12-15 Å) have the most significant impact on an atom's B-factor. This aligns with physical intuition about local interactions.
  • Window Size Matters: The optimal window size for considering neighboring amino acids in the sequence varies between proteins. The study found that a window size of around 12-15 Å for Euclidean distance is generally optimal, and larger windows don't necessarily improve predictions, suggesting a localized influence.
  • LSTM's Crucial Role: Removing the LSTM component from the model drastically reduces predictive accuracy, highlighting its importance in capturing sequential dependencies and long-range interactions.
  • Amino Acid Type and Secondary Structure Correlation: While the primary sequence is sufficient, the model's predictions correlate with known trends: hydrophobic amino acids (often buried) tend to have lower B-factors, while charged amino acids (often on the surface) tend to have higher B-factors. Similarly, more ordered secondary structures like beta-sheets and alpha-helices generally show lower B-factors than disordered coils.

Key Results

  • State-of-the-Art Accuracy: The developed sequence-based LSTM model achieved an average Pearson Correlation Coefficient (PCC) of 0.8 for normalized B-factor prediction on the test dataset. This is a significant improvement over previous state-of-the-art models, which reported PCCs between 0.53 and 0.61. The model also achieved a PCC of 0.73 for un-normalized B-factors.
  • 30% Improvement: When tested on a subset of proteins not used by prior SOTA models, the proposed model achieved a 30% higher PCC compared to the average PCC of those models.
  • Robustness: The model demonstrated robustness by achieving consistent PCC values across different random seeds used for training.
  • Cutoff Radius Identification: The study identified a critical cutoff radius of approximately 12-15 Å for the influence of neighboring atoms on B-factor prediction, which is in agreement with findings from protein network models.

Limitations and Future Work

  • Protein Size Limit: The current dataset excluded proteins larger than 500 amino acids. The authors suggest that future work could explore whether the model architecture can be extended to handle larger proteins.
  • Experimental Error: The study acknowledges that B-factors can be influenced by experimental conditions (e.g., resolution, solvent content), which can introduce noise and uncertainty. While normalization helps, it doesn't entirely eliminate this issue.
  • Future Applications: The authors envision this model being used for predicting other mechanical properties of protein-based materials, identifying active regions for drug discovery, and assisting in the design of de novo proteins with specific functionalities. The sequence-based nature makes it particularly suitable for in silico design where only the sequence is known.

Summary updated

BibTeX
@article{pandey2023b,
  author = {Akash Pandey and Elaine Liu and Jacob Graham and Wei Chen and Sinan Keten},
  title = {{B-factor prediction in proteins using a sequence-based deep learning model}},
  journal = {Patterns},
  year = {2023},
  doi = {10.1016/j.patter.2023.100805},
}
Figure