Hashcodes

Clearly, in this case, checks have to be made whether a specific isomer has already been produced by a different pathway. This is achieved by the calculation of hashcodes (see Section 2.7.4). 

The comparison For identity of chemical structures is peiTormed on the basis of a hashcode algorithm, Hashcode are unique integer numbers which can be calculated for each chemical structure within the WODCA system [55], Further exam-... 

Figure 10.3-16. The principle of similarity searches. The query (target, precursor) as well as the catalog compound are transformed by the criterion maximum oxidation state". Since the transformation for both compounds results in the samie transformed structure, the catalog compound is presented to the user as a suitable starting material. The comparison of the structure is performed by a hashcode algorithm.

The ring hashcodes used to create the clusters have been described previously [37]. Briefly, each hashcode consists of a seven-letter string and is derived from atom pair... 

Fig. 6.6 liiustration of the four different types of ring hashcodes for a sampie moiecuie. 

The final selection of compounds was compared with the set that would have been selected using the Top 4000 method on the entire dataset or on the filtered dataset, in terms of diversity and physical properties. Our selection gave a 30.3 or 41.2% increase in the number of uniquely represented ring hashcodes (2681) as compared with the Top X selection (most active 4000 compounds) from either the filtered (2058) or unfiltered... 

The 5 k actives with percent inhibition of 25 to 40% and that feU into clusters of less than five compounds were treated separately using BCUT diversity analysis, as described in Section 6.2.5. A cell-based selection biased by primary activity from six bins per each of six axes yielded 1258 compounds. The combined selection from filtering, clustering, and diversity totaled 6986 compounds representing 3337 ring scaffolds and was submitted for confirmation assays. Note that the full set of 16 k filtered actives contained 9254 ring hashcodes, so the selected set covers 36.1% of the represented scaffolds. Because of the presence of duplicate samples in the corporate screening collection, 7275 samples were pulled and assayed. 

The confirmed actives (4296 of 7275) represented 1762 unique ring hashcodes or 52.8% of those present in the selected set. Of the confirmed actives, 3916 (91.2%) and 4097 (95.4%) would have been included in a Top X selection of8000 from the unfiltered or filtered datasets respectively. 

Actives with 175 > MW > 500, clogP > 7, donors > 5, acceptors > 10, rotatable bonds > lOorTPSA > 140 A wereflaggedand subjecttovisualassessmentin Spotfrre by a chemist. After removal of duplicate samples, 3337 compounds remained and were submitted for confirmation assays. These represented 2290 unique ring hashcodes. 

The results showed a 46.6% confirmation rate (53.4% false positive rate), giving 1555 confirmed actives (1,096 unique ring hashcodes, 47.9%). A second round of filtering was applied using tighter physiochemical property criteria. Molecules with 200 > MW > 500, clogP > 5.0, rotors > 8 were flagged and visually assessed. The final selection for secondary assays was 1087 compounds (70%). 

However, the current method does have drawbacks. The first is related to the limitations of the ring hashcode descriptor for clustering, as discussed in Section 6.2.4.1. However, this is currently being addressed through the development of a new molecular descriptor. The second is the relatively resource-intensive nature of the method it requires significantly more time to apply than a Top X method. 

An alternative methodology based on the ringcontent of a database, using precalculated structure-based hash codes has been proposed (110). The comparison of the hashcode tables can be used to compare two databases and the number of distinct ring-system combinations can be used as an indicator of database diversity. A method for diversity assessment called the saturation diversity approach, based on picking as many mutually dissimilar compounds as possible from a database was also proposed. The methods were used to compare a number of public databases and gave similar results. 

Identity matching is based on a comparison of hashcodes. We have developed a hashcode algorithm that converts the connection table of a structure into a signed 32 bit integer number. Care was taken to avoid collisions, i.e., as of now we have not found an example where two different structures lead to the same hashcode. 

Initial comparison of the hashcode of the two structures shows that they are different. Then, all carbon atoms bearing oxygen atoms are put into the maximum oxidation state. The hashcode of the resulting two structures is found to be identical. Thus, we can conclude that the two initial structures are similar. 

Figure 3. Storing the hashcodes for similarity comparison in the memory...

---