Fig. 2From: A comparison of embedding aggregation strategies in drug–target interaction predictionUMAP visualization of the test compound-protein embeddings extracted from the fully-connected layers of the best MLP configuration trained on the DAVIS dataset. The configuration uses CNN compound and protein branches. The points are colored using the binarized affinities (orange for active and green for inactive). From the figure, we observe that starting from the first layer (after the concatenation of the compound and protein branch embeddings) to the third, the separation of the two classes becomes clearerBack to article page