Neural Network Definitions of Highly Predictable Protein Secondary Structure Classes

NeurIPS 1993

TL;DR

This paper introduces a novel machine learning approach using two co-evolving neural networks to discover new, more predictable classes of protein secondary structure. By forcing these networks to agree on predictions derived from both amino acid sequence and structural information, they identify classes that are significantly better predicted from sequence than traditional alpha-helix, beta-strand, and coil classifications.

What Problem Does This Paper Address?

Predicting protein secondary structure (like alpha-helices and beta-sheets) from its amino acid sequence is a long-standing, important problem in computational biology. Despite decades of effort and various methods, including neural networks, the accuracy of predicting the conventional secondary structure classes has remained disappointingly low. This suggests that the traditional classifications might not be the most naturally predictable units from sequence alone. The authors aim to find new ways to classify local protein structure that are inherently more predictable from the primary sequence.

Methods

The core of the paper's methodology lies in using two neural networks that are trained simultaneously to "co-evolve." * Left Network: Takes the amino acid sequence of a local window as input and predicts a "secondary structure" classification. * Right Network: Takes local structural information (specifically, backbone dihedral angles, represented by sine and cosine of ΦΨ angles) as input and also predicts a "secondary structure" classification.

The key innovation is the constraint imposed during training: the outputs of both networks must agree for each input. This agreement is not simply about minimizing the squared difference between their outputs, which can lead to trivial solutions where both networks output a constant value. Instead, they explore several objective functions to enforce agreement, including maximizing correlation and, most successfully, maximizing the Mathews correlation coefficient. This process effectively allows the networks to discover new, mutually consistent definitions of secondary structure that are predictable from both sequence and local structure.

Datasets

The study utilized a database of 105 proteins, identical to those used in previous investigations. This database was split into:

  • A training set of 91 proteins.
  • A prediction set of 14 proteins.

For comparison with conventional methods, the standard alpha, beta, and coil states were defined using the Kabsch and Sander definitions. A window size of 13 residues was used, resulting in 16028 training set examples and 3005 prediction set examples.

Key Insights and Ideas

  • Beyond Conventional Classes: The traditional alpha-helix, beta-strand, and coil classifications are not necessarily the most predictable units from amino acid sequence.
  • Co-evolutionary Learning: Training two networks simultaneously with an agreement constraint is a powerful paradigm for discovering new, predictable classifications from data.
  • Sequence-Structure Synergy: By forcing agreement between networks processing sequence and structural information, the method can uncover structural definitions that are well-supported by both types of data.
  • New Classes are Related but Different: The newly discovered "Xclass" shows a significant correlation with conventional alpha-helices but is not identical. It appears to be a broader or slightly shifted definition, offering higher predictability.
  • Random Initialization is Crucial: Starting with random initial conditions for the network weights is necessary to evolve interesting, non-trivial new classes, rather than just converging to existing ones.

Key Results

The paper reports that their novel "Xclass" (a dichotomous classification) achieved a Mathews correlation coefficient of -0.425 on the prediction set. This is a negative correlation, indicating anti-correlation, which is a valid outcome of their objective function and can be easily compensated for. For comparison, predicting conventional classes (Alpha/NotAlpha, Beta/NotBeta, Coil/NotCoil) using standard backpropagation achieved coefficients of 0.33, 0.26, and 0.39, respectively.

While the new Xclass is negatively correlated with the conventional classes, it is significantly more predictable from sequence than the conventional ones. The Xclass is also found to comprise about 10% of the data, while conventional alpha-helices make up roughly 25%. The new Xclass typically extends the definition of helix past standard boundaries.

Limitations and Future Work

  • Local Minima: The authors acknowledge that objective functions can be susceptible to local minima, and they often resorted to trying new random initial conditions to escape uninteresting solutions.
  • Limited Window Size Exploration: The effects of window sizes other than 13 residues were not extensively tested.
  • Extending to Multiple Classes: While the method is conceptually extendable to more than two classes, the paper focuses on a dichotomous classification for clarity.
  • Physical Meaning: Further work is needed to fully analyze the physical meaning of the newly evolved secondary structure classes.
  • Broader Applications: The authors suggest that this co-evolving network paradigm could be applied to other tasks requiring the correlation of different data representations.

Summary updated

BibTeX
@inproceedings{1993neural,
  title = {{Neural Network Definitions of Highly Predictable Protein Secondary Structure Classes}},
  booktitle = {Advances in Neural Information Processing Systems},
  year = {1993},
}
Figure