Feature representation

Using the hydrophobicity scale as an example, the amino acid residues can be represented by the real-numbered scalar value. They can also be classified with respect to their side chains as polar, nonpolar, or amphipathic, depending on the range of the hydrophobicity in the scale, such as in 

ExchangeGroup 6 Conservative Substitution HRK DENQ (C) STPAG MILV FYWJ  

Structural 3 Surface Exposure (DENQHRK) CSTPAGWY MILVF  

2DPropensity 3 2D Structure Propensity AEQHKMLR CTIVFYW SGPDN  

Instead of the discrete membership described above (Table 6.2), sometimes it may be desirable to define overlapping memberships. This may apply to properties such as 

---

Table 6.2 Alphabet sets for feature representation some examples.

The descriptor pool can be reduced in size by using standard statistical procedures. For example, descriptors that contain little or no useful information or are similar in value and content to other descriptors are rejected. Therefore descriptors that have the same value for nearly every compound are rejected, as are ones that are highly correlated (r>0.95) with other descriptors. When pairwise correlation is encountered, the descriptor that has the more obvious feature representation or is easier to calculate is retained. 

Figure 7 Illustration of a decision tree generated with C4.5 obtained for the classification of chemical carcinogens in rodents [61]. The classification of 122 chemicals was based on a molecular feature representation that included structural alerts, biological activities in different assays and pathological indicators. The classification leaves at the right hand side of the decision tree are visualized graphically by filled or open boxes. A filled box indicates that the chemicals in that box are classified as rodent carcinogenic open boxes indicate chemicals without rodent carcinogenicity. Note the strong imbalance of the decision tree.

Finally, the real value in many papers is derived from fhe features developed in the search for a meaningful representation of profein sequences and structure. However, too many papers focus on mechanistic feature representations such as... 

The form feature representation records information about the means of shape control. Sweeping is often preferred. Figure 4-15 shows three applications of axisymmetric sweep for creating... 

The present study proposed to extract six affine moment invariants for feature representation. These features are invariant under rotation, scale and translation, thus useful to represent the bacilli. Detailed discussion of affine moment invariants can be found in [11]. 

A water pump, taken from [6], is used as an example in this presentation. It is composed of 19 parts (Please refer [6] for the pictorial view of the water pump.). The feature representation of a package ring is shown in Figure 3. 

Figure 3 Feature Representation of a Package Ring The connection of a bolt to the flange is described in Figure 4.

---