Inversion of protein folding neural networks, such as AlphaFold2, RosettaFold, ESMFold, and trRosetta, has been used for protein design by hallucination as well as adversarial attacks. It involves inversion of the networks and backpropagation to input sequence using a custom loss.
Mentions
- Designing proteins by hallucination can be improved by Markov chain Monte Carlo sampling at the end (1). This is in contrast to just gradient descent.
Figure 2A from (1)
Examples
- Hansen et al. (2) designed a thermostable glycoside hydrolase using hallucination. Although the proteins folded, they were inactive due to misplacement of the active site side chains.
Ref (2)
1.
Goverde CA, Wolf B, Khakzad H, Rosset S, Correia BE. De novo protein design by inversion of the AlphaFold structure prediction network. Protein Science. 2023;32(6). Available from: https://doi.org/10.1002/pro.4653
2.
Hansen AL, Theisen FF, Crehuet R, Marcos E, Aghajari N, Willemoës M. Carving out a Glycoside Hydrolase Active Site for Incorporation into a New Protein Scaffold Using Deep Network Hallucination. ACS Synthetic Biology. 2024;13(3):862–75. Available from: https://doi.org/10.1021/acssynbio.3c00674