Development of NEural network simulator for structure-activity COrrelation of molecules:Neco. (3). Performance Evaluation of Self-organized Network and Perceptron.

康子 藤谷, 光永 小野寺, 芳美 井須, 長嶋 雲兵, Haruo Hosoya, 青山 智夫Journal of Chemical Software 1998 · DOI

Abstract

Researchers have developed a hybrid neural network simulator, Neco, designed specifically for structure-activity correlation (QSAR) in chemistry. By integrating a fast-learning self-organizing layer with a high-precision perceptron, Neco achieves a 4x speedup in reaching practical accuracy levels compared to standard models. The system was validated using the Exclusive-OR (XOR) problem and the prediction of 3D conformations in norbornane derivatives from $^{13}$C-NMR shifts.

The Challenge of Molecular Modeling

In computational chemistry, predicting a molecule's properties from its structure is often hindered by highly non-linear relationships. For example, the relationship between a chemical shift in an NMR spectrum and a specific 3D molecular shape is rarely straightforward. Traditional multi-layer perceptrons are accurate but can be agonizingly slow to train, while self-organizing maps are fast but often lack the fine-grained precision needed for chemical modeling. Neco bridges this gap by using a hybrid architecture that first organizes data into meaningful clusters before performing high-precision mapping.

Benchmarking: XOR and Norbornane NMR

The researchers tested Neco against two primary benchmarks:

  1. The XOR Problem: A classic non-linear classification task using 30,000 data points distributed across quadrants. This test measures the network's ability to handle non-linear boundaries and maintain data symmetry.
  2. Norbornane $^{13}$C-NMR Shifts: A real-world dataset of 38 compounds. The network used chemical shifts from seven carbon atoms to predict the exo vs. endo orientation of substituents.
Visualization of the XOR problem classification across four quadrants, demonstrating the network's ability to handle non-linear data distributions.

Performance Evaluation: Speed vs. Precision

A key finding of the study is the "practical accuracy" advantage. The hybrid model reaches a usable error threshold (e.g., an error of 0.2) approximately four times faster than a standard perceptron. In the XOR test, the hybrid model converged in approximately 220 steps, compared to 800 steps for the standard perceptron.

Comparison of error convergence over time. The hybrid model (dotted line) reaches practical accuracy significantly faster than the standard perceptron (solid line).

While the standard perceptron eventually achieves higher precision in the final stages of training, Neco's rapid initial convergence makes it ideal for rapid, iterative practical modeling. Furthermore, because Neco uses distance-based clustering, it maintains the inherent symmetry of the input data better than standard perceptrons, which often produce lopsided decision boundaries due to their weight-adjustment mechanisms.

Key Results and Conclusion

Both models performed exceptionally well on the chemical dataset, correctly predicting the conformation for 12 out of 13 unknown compounds (92% accuracy). However, the hybrid model required only 0.1 seconds of CPU time to reach stability, compared to 0.4 seconds for the perceptron—maintaining its 4x performance edge.

The Neco simulator offers a significant performance boost for chemical property prediction. Its hybrid architecture provides a robust balance between the speed of self-organization and the precision of supervised learning. Future developments will focus on scaling the system for larger molecular datasets and improving automated feature extraction to identify which structural elements most influence chemical properties.

Summary updated

Abstract

高速学習が可能な自己組織化モデルと、高精度のパーセプトロンモデルを組み合わせることにより、分子の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開発した。それを、2次元のExclusive-OR問題とノルボルナン類の13C-NMRシフトとの立体配置の問題に適用し、従来のパーセプトロンモデルと性能の比較を行なった。この組み合わせモデルは、実用的な精度までの収束は早いが、高精度の収束はパーセプトロンモデルより遅くなることがわかった。実用的な精度までの収束速度は、パーセプトロンモデルのそれより約4倍早い。パラメータスキャン法を用いて、挙動を調べると、組み合わせモデルのほうが、パーセプトロンのそれに比べて、非線形性の強い分類をすることがわかった。
BibTeX
@article{1998development,
  author = {康子 藤谷 and 光永 小野寺 and 芳美 井須 and 長嶋 雲兵 and Haruo Hosoya and 青山 智夫},
  title = {{Development of NEural network simulator for structure-activity COrrelation of molecules:Neco. (3). Performance Evaluation of Self-organized Network and Perceptron.}},
  journal = {Journal of Chemical Software},
  year = {1998},
  doi = {10.2477/jchemsoft.4.19},
}
Figure