Fig. 6: Scheme of the FlexConsensus training workflow.
From: Merging conformational landscapes in a single consensus space with FlexConsensus algorithm
A set of independent spaces is fed to different encoders, generating an independent representation of the initial spaces on a common consensus space. All encoded consensus spaces are then forwarded to the first decoder, which is responsible for transforming the consensus spaces into a space as similar as possible to the first input space. A representation loss is evaluated and backpropagated through the network at this step. The previous decoded step is repeated sequentially for all the decoders, allowing the training of all the weights of the multi-autoencoder network.
