Fig. 1

The comprehensive framework for training TopoProGenerator.A Framework for model fine-tuning: The first step involves an generative model producing fake sequences with specified topology to train a discriminative model that distinguishes between real and fake sequences. The second step entails the generative model generating sequences, and the discriminative and predictive models providing scores, which are aggregated into a reward for the sequence. Finally, in the third step, the model backpropagates using the reward to optimize the parameters of the generative model.B Pretraining process for the generative model: For each sequence, the model constructs features by considering pairwise residue relationships. It utilizes the features of previously generated residues to determine the subsequent residue, maximizing the probability of generating a complete sequence. The first token of the sequence indicates its topology. C The predictive model comprises ProtBert and a multilayer perceptron. The sequence input is passed through ProtBert to extract features, which are subsequently fed into the multilayer perceptron to generate stability scores