| 摘要 |
<p>A signal processing system is used in the manufacture of chemical compound to predict which molecules may have a required property for the compound. In the signal processing system, signals defining molecules in a training set and a measured property of each molecule are processed to produce fixed length signals encoding each molecule by (i) identifying for each molecule active atoms which may cause the molecule to react, and physical properties; (ii) defining each unique group which contacts a given number of the active atoms as a 'key' in terms of the atoms and their relative positions; (iii) determining the contribution each key makes to the measured property of molecules containing the key; (iv) forming a histogram of the contributions of the keys in a molecule; and (v) encoding the molecule using the histogram values and the physical properties. The training data is used to train a neural network. A further molecule is encoded as before, but using the individual key contributions defined during training. The trained neural network is used to predict the property of the further molecule.</p> |
