Jonathan Fine

Liquid chromatography is a cornerstone analytical technique for separating, quantifying, and purifying components from complex mixtures. To accelerate method screening, an uncertainty-aware graph-based neural network that predicts retention times across multiple column chemistries and buffer pH conditions is herein introduced. Specifically, the Multiple Condition Retention time model, MC-Retention, incorporates explicit column and buffer descriptors and is trained on a newly assembled dataset produced by screening 480 analytes under eight chromatography conditions. By being able to accurately model multiple chromatographic methods simultaneously, MC-Retention can substantially reduce the time and resources required for LC method screening and development. With an aggregate R2 of 0.86 and MAE of 15.5 seconds during cross validation, MC-Retention's respectable performance is greatly enhanced for analytes made present during training for one the four column chemistries when those analytes are otherwise not associated the columns being validated. This selective training of analytes for a single column chemistry reduces error by 50%, with an R2 of 0.95 and MAE of 8.2 seconds in aggregate. Additionally, we demonstrate that MC Retention effectively identifies optimal conditions for separating peptide coupling reaction components, while also suppling calibrated uncertainty estimates for its retention time predictions.

The design-make-test cycle for drug discovery is highly dependent on the purification of synthesized compounds. Prior to evaluation of suitability, ultrahigh performance liquid chromatography is used for an initial standard analysis, where retention times of analytes are measured with a shorter standard gradient method and used to select the appropriate gradients for a final purification method. To circumvent this preliminary screening experiment for small molecule libraries, retention time prediction had been achieved previously by use of commercial modeling methods. However, these retention time prediction models can have limited applicability when built from smaller datasets and are less effective when constructed from disparate data collected under differing chromatography conditions. Having thousands of measured retention times from high-throughput physiochemical screening, we sought to leverage these data for the construction of predictive models for a standard preliminary method enabling high-throughput purification of macrocyclic peptide libraries. Utilizing 4549 analytes and their retention times from high-throughput physiochemical screening, a structure to retention time model was built using a graph invariant network, a form of artificial neural network architecture. Once fitted to high-throughput screening data, the model was re-trained with standard gradient method data, a technique known as transfer learning. Through transfer learning, a training set of 80 analytes yielded a neural network model, that when evaluated against a test set of 24 analytes, displays high performance metrics with a coefficient of determination (R2) of 0.82 and mean average error of 0.088 minutes, or 1.26% of the gradient time. This model has been deployed internally as a dash-app to help democratize the use of the developed models and is being used for selecting purification methods based on analyte structure.

The screening of chemical libraries is an essential starting point in the drug discovery process. While some researchers desire a more thorough screening of drug targets against a narrower scope of molecules, it is not uncommon for diverse screening sets to be favored during early stages of drug discovery. However, a cost burden is associated with the screening of molecules, with potential drawbacks if particular areas of chemical space are needlessly over represented. To facilitate triaged sampling of chemical libraries and other collections of molecules, we have developed Dedenser, a tool for the downsampling of chemical clusters. Dedenser functions by reducing the membership of clusters within chemical point clouds while maintaining the initial topology, or distribution, in chemical space. Dedenser is a Python package that utilizes Hierarchical Density-Based Spatial Clustering of Applications with Noise to first identify clusters present in 3D chemical point clouds, and then downsamples by applying Poisson disk sampling to clusters based on either their volume or density in chemical space. A command line interface tool is available with Dedenser, which allows for generation of chemical point clouds, using Mordred for QSAR descriptor calculations and uniform manifold approximation and projection for 3D embedding, as well as visualization. We hope that Dedenser will serve the community by enabling quick access to reduced collections of molecules that are representative of larger sets, selecting even distributions of molecules within clusters rather than single representative molecules.