End-to-End Recommendation with MolRec

Exploring the vast and growing landscape of chemical compounds is a significant challenge for researchers, therefore making the development of an effective chemical recommender system not just interesting but crucial. Building such a system has the potential to revolutionize how scientists discover and interact with chemical data, enhancing their research productivity and sparking innovation across chemistry and allied sciences. Previous works on recommendation of chemical compounds has been limited and not widely explored, indicating a significant opportunity for developing systems that can effectively navigate, predict and recommend novel and existing chemical compounds. The gap in the literature underscores the necessity for methods that integrate diverse algorithmic approaches, aiming to improve the discovery and identification process within extensive chemical datasets. 

​

Towards this goal, we introduce a hybrid chemical recommender system which leverages a combination of algorithms, including collaborative filtering, content-based approaches, and Graph Neural Network variational autoencoders to effectively identify and recommend compounds of interest. It focuses on introducing scientific researchers to potentially unknown chemical compounds within large-scale chemical datasets, enhancing discovery and research efficiency in chemistry and related fields.  

​

Our end-to-end recommendation model is outlined in the figure above. In the first stage, we use classical IR methods such as collaborative filtering and content-based approaches to recommend multiple reference molecules to each researcher based on their previous research interests. This is done using a dataset of our own construction which lists molecules researchers have interacted with in their previous papers. The methods in the first stage and MolRec data are described in greater detail at Stage I: Classical Recommendation. In the second stage, we use these reference molecules and a variational autoencoder GNN to recommend novel molecules to the researcher beyond the molecules in the dataset, described in greater detail at Stage II: Deep Generative Model.

​

We next show several end-to-end recommendations for selected researchers. Please visit the end-to-end recommendation visualization section of our GitHub or the gallery below for more. In these visualizations, the top row shows reference molecules that the researcher has previously interacted with in their work. The Stage I Classical Recommendation methods use these reference molecules to recommend several molecules from the MolRec data, shown in the middle row. Finally, given the Stage I recommendations, the Stage II VAE generates a novel molecule, which is shown in the third and final row.