Preprint: OmniSyn created a virtual library of 2,7 billion protein–molecule pairs with synthesis plans for 21 thousand human proteins
Preprint: OmniSyn created a virtual library of 2,7 billion protein–molecule pairs with synthesis plans for 21 thousand human proteins
On September 6, a preprint was posted on bioRxiv, a server for preliminary scientific papers, describing OmniSyn. This computational model takes a protein's amino acid sequence and proposes a candidate molecule together with a sequence of chemical steps to assemble it.
Many models that design molecules for a protein require the three-dimensional structure of the region where the molecule is expected to bind. OmniSyn uses the amino acid sequence, the order of the amino acids that make up the protein. At each step, it selects a commercially available chemical building block and a reaction type to connect it to the rest of the molecule. The model was trained using more than 220 thousand such building blocks and 115 common reaction types.
The model thus produces three linked outputs: a target protein, a candidate molecule, and a plan for assembling it. It can design a new structure from scratch, modify a known molecular scaffold, or propose a close analogue of a candidate for which a synthesis route is difficult to devise.
The authors evaluated the model on 35 proteins outside the training data. Among the structures generated for these proteins, OmniSyn produced exact matches to known molecules active against the assigned target more often: 0,255% compared with 0,028% for the next-best model. This comparison measures the generator's ability to select molecules for a specific protein.
When known molecules were modified, AiZynthFinder, a program that searches backward from a completed molecular structure to starting materials, found a computational synthesis route for 71,92% of the OmniSyn analogues. For 21 306 human protein sequences, the model generated 2,7 billion target–molecule records. Each record includes a model-generated synthesis plan and scores for candidate selection.
For 200 proteins, the authors used two computational assessments to test whether the scores distinguished the assigned target from randomly selected proteins. Both assessments gave better results for the assigned targets: 7,64 versus 4,47 for the interaction score and −6,88 versus −6,20 kcal/mol for computational docking of the molecule to the protein. For each protein sequence, the library stores a virtual candidate, its assembly plan, and scores for candidate selection.