Abstract
構造活性相関解析のためのニューラルネットワークシミュレータをC言語を用いて開発した。本シミュレータでは、中間層の層数、中間層のユニット数を自由に設定できるようにした。また、学習方法を通常のバックプロパゲーション法と青山らが提唱した再構築学習法の2通りを選択できるようにした。また、本シミュレータの応用例として、ここでは、ノルボルナン誘導体の13C-NMRケミカルシフトとその立体配置の関係を学習し、未知誘導体の立体配置を予測することを行なった。この結果、実験と非常に良い結果を示した。また、再構築学習法により、ノルボルナンの立体配置の決定には7つの炭素のうちの2つの炭素のケミカルシフト値が大きく関わっていることが判った。このことはパラメータスキャン法を用いて確かめられた。
Introduction
In the early 1990s, Structure-Activity Relationship (SAR) analysis primarily relied on linear regression. However, the relationship between a molecule's shape and its chemical behavior is often highly non-linear, making linear models hit a "performance ceiling." While neural networks (NNs) were known to handle non-linearity well, they were often criticized for being "black boxes." This 1994 paper introduces Neco, a neural network simulator designed to bridge the gap between complex molecular data and structural predictions by moving beyond "black box" predictions.
Methods: The Neco Simulator
The authors built the Neco (NEural network simulator for structure-activity COrrelation) using the C programming language for portability across Unix workstations. Key technical features include:
- Flexible Architecture: Users can freely set the number of hidden layers and units.
- Dual Learning Modes:
- Standard Backpropagation: The classic method for training NNs.
- Reconstruction Learning: A more advanced method that "prunes" or zeroes out unnecessary connection weights during training. This effectively performs automated feature selection.
- Parameter Scan Method: A sensitivity analysis tool where the simulator systematically varies input values to visualize how the network’s "opinion" changes, helping to map out decision boundaries.
Application: Norbornane Stereochemistry
The researchers evaluated Neco using a dataset of norbornane derivatives, a classic class of bridged bicyclic compounds. The goal was to predict the stereochemical configuration (exo vs. endo) based on $^{13}$C-NMR chemical shifts.
Dataset and Inputs
- Inputs: $^{13}$C-NMR chemical shifts for seven specific carbon atoms ($C_1$ through $C_7$).
- Target: Predicting the stereochemical configuration.
- Size: 25 compounds were used for training, and 13 "unknown" compounds were used for testing.
Key Insights and Results
- From 7D to 2D: The most striking finding was that although they provided data for seven carbon atoms, the "reconstruction learning" method showed that the network ignored most of them. It discovered that the chemical shifts of only $C_6$ and $C_7$ were truly necessary to determine the exo/endo configuration.
- High Accuracy: Neco correctly predicted the configuration of 12 out of 13 unknown test compounds.
- Visualizing the "Brain": By using the Parameter Scan method, the authors could plot the decision regions for $C_6$ and $C_7$. This turned the "black box" into a "glass box," showing a clear, non-linear boundary that separated the two stereoisomers.
- Data Quality: The network failed on one specific compound (No. 31). Upon investigation, they found its $C_6$ shift was an extreme outlier, highlighting that NN reliability is strictly tied to training data representation.
Conclusion
Neco demonstrates that neural networks can be both powerful predictive tools and interpretable models for chemical structure analysis. By identifying $C_6$ and $C_7$ as the primary descriptors for norbornane stereochemistry, the simulator provided insights that align with chemical intuition while handling the inherent non-linearity of NMR data.
Summary updated