Using Knowledge-Based Neural Networks to Improve Algorithms: Refining the Chou-Fasman Algorithm for Protein Folding

Richard Maclin, Jude W. ShavlikAAAI 1992

TL;DR

The authors introduce FSKBANN, a method that converts existing state-based algorithms into neural network architectures. By seeding a network with the logic of the classic Chou-Fasman algorithm for protein folding and refining it with data, they achieved higher accuracy than both the original human-designed algorithm and standard neural networks trained from scratch.

Introduction: Bridging Domain Theory and Machine Learning

In the early 1990s, machine learning models were typically built "from scratch," often ignoring decades of expert-derived "domain theory." While the KBANN system could translate simple "if-then" rules into neural networks, it lacked the ability to handle state—a critical component of most real-world algorithms.

In protein folding, for instance, the classification of an amino acid often depends on its neighbors. To address this, the authors developed a system capable of ingesting complex, state-based algorithms and using machine learning to fine-tune them, effectively bridging the gap between symbolic logic and connectionist learning.

The FSKBANN Methodology

The core contribution is FSKBANN (Finite-State Knowledge-Based Artificial Neural Networks). This approach follows a three-stage pipeline:

  1. Representation: An existing algorithm (e.g., Chou-Fasman) is modeled as a generalized Finite-State Automaton (FSA).
  2. Translation: The FSA is mapped onto a recurrent neural network topology, specifically a variant of an Elman network. The states and rules of the algorithm dictate the initial hidden unit count and connectivity.
  3. Refinement: Unlike standard networks that begin with random weights, FSKBANN starts with weights that encode the algorithm's logic. Standard backpropagation then adjusts these weights based on empirical data to correct the human-designed algorithm's practical failures.
The FSKBANN architecture, where domain theory determines the hidden unit topology and state feedback enables memory.

Experimental Framework

The researchers evaluated FSKBANN using the Qian & Sejnowski (1988) Protein Dataset, which contains 21,623 amino acids across 106 proteins. The objective was to predict three secondary structures: $\alpha$-helix (25.2% of the data), $\beta$-strand (20.3%), and random coils (54.5%).

Key Insights: The Power of Structured Knowledge

  • Knowledge as a Scaffold: Starting the search in "weight space" from a position of existing knowledge is significantly more effective than random initialization. Even imperfect theories provide a structured starting point.
  • The Necessity of State: Simply adding state information to a standard network is insufficient. The specific structure provided by domain theory (the Chou-Fasman logic) was essential to leveraging that state effectively.
  • Targeted Improvements: FSKBANN notably improved the prediction of $\beta$-strands, a structure historically difficult for both human experts and standard AI models to identify accurately.

Performance Analysis

FSKBANN outperformed all baselines in predicting protein secondary structure:

  • Overall Accuracy: FSKBANN reached 63.4% accuracy, surpassing standard ANNs (61.8%) and the original Chou-Fasman algorithm (57.3%).
  • Statistical Significance: A t-test confirmed that the 1.6% improvement over standard ANNs was significant at the 99.5% confidence level.
  • Structural Precision: FSKBANN demonstrated a much higher correlation for $\beta$-strands (0.33) compared to standard ANNs (0.25), indicating a better ability to identify actual structural regions rather than merely predicting the most frequent class.
Detailed performance comparison across secondary structure types, highlighting FSKBANN's superior handling of beta-strands.

Limitations and Future Directions

Despite its success, FSKBANN faces several challenges. A primary issue is the "Black Box" problem: once trained, the original human-readable rules are distributed across the network's weights, making it difficult to extract the refined rules back into a human-readable format. Additionally, the system is currently limited to propositional logic, preventing the ingestion of algorithms based on predicate calculus. Finally, the authors suggest that future evaluations should move beyond "per-amino-acid" accuracy toward metrics that better reflect the biological importance of identifying correct structural regions and their sequences.

Summary updated

BibTeX
@inproceedings{maclin1992using,
  author = {Richard Maclin and Jude W. Shavlik},
  title = {{Using Knowledge-Based Neural Networks to Improve Algorithms: Refining the Chou-Fasman Algorithm for Protein Folding}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {1992},
}
Figure