Fig. 3From: A comparison of embedding aggregation strategies in drug–target interaction predictionUMAP visualization of the compound embeddings extracted from the last fully-connected layer of the compound branch (CNN) of the best MLP (left) and dot product (right) configurations trained on the KIBA dataset. The points are colored based on the affinities of the four proteins in the training set with the highest number of recorded entries. Comparing the two strategies, we observe that both create clusters of similar qualityBack to article page