AI (ARTIFICIAL INTELLIGENCE) applied to enhance the DMPK experimental efficacy

The goal of this presentation is to show multiple ways to increase DMPK experimental efficacy. An enhancement of efficacy in this context is understood as how to get more knowledge out of the data that is produced. This process is translated on how to decrease the time from data to compound design and/or how to increase the knowledge learned from the data. To achieve this objective automation of data processing, maximizing the extraction of information out of the data and the translation the newly acquired information into practical design is necessary.

Methods

To increase the automation a new methodology will be shown where the input experimental files acquired on a Thermo Q-Exactive instrument for a clearance workflow are processed without human intervention to obtain the clearance and metabolite identification results. The procedure will be exemplified in a collection of 20 PROTACS structures where the Clearance measurement in Hepatocytes has been performed. Clearance end point data computation will be shown together with the automatic fragment assignment for each of the metabolites.

Preliminary data

For example, the fragmentation observed and assigned for a particular compound was used to develop a model to predict the compound fragmentation. In this context several published spectra interpretation from the MoNA database have been used to determine a Machine Learning model to predict the probability of a bond to be broken under certain Mass Spectrometry condition. That publicly available information is used to derive a model that was validated and monitored by using the own experimental data generated in the automatic procedure mentioned before. Considering the result from the comparison between the experimental and the prediction values the new data was incorporated into an automatically learning process to improve model quality. In this way a self-learning system has been developed for the improvement of the virtual fragmentation of a compound.

The above self learning procedure was also applied for other Mass Spectrometry derived property like for example the Cross Collision section (CCS) value to assess the reliability of a structural assignment to a certain peak. In this case a regression model with a regression coefficient between the observed and the calculated CCS values of 0.98 was obtained after developing a model based on published data and enriched with our own dataset.

Finally, an example of how the metabolite identification information extracted from the previously described automation system is used to design compound with a reduction in compound clearance will be shown.