AdaptiveFlow selects molecules for laboratory testing from 69 billion candidates

AdaptiveFlow selects molecules for laboratory testing from 69 billion candidates

On September 1, Nature Biotechnology published a paper describing an open system for virtual molecular screening. The system prepares a library of 69 billion chemical candidates for computational analysis, samples different regions of the library, and directs the search toward promising regions for detailed screening. For FSP1 and PARP1, the authors synthesized and tested a subset of the candidates. They also determined protein complex structures for selected molecules.

Virtual screening identifies molecules that may bind to a specific site on a protein. The software tests their possible poses within the three-dimensional protein structure and assigns a computational score to each candidate. When the search space contains tens of billions of candidates, evaluating every structure exhaustively can consume substantial computational resources before the first compound reaches the laboratory.

AdaptiveFlow converts the original compound records into three-dimensional variants suitable for these calculations. This preparation produced 68,7 billion records from 31,5 billion structures in the Enamine REAL Space library because a single source structure can yield several three-dimensional variants. The system then distributes these records across a space defined by 18 properties, including molecular weight, charge, and hydrogen-bonding capacity. The 12 million occupied cells contained an average of approximately 5 600 molecules each.

For a given protein, the system tests one to ten representatives from each cell. It then examines every molecule in cells that produced favorable computational scores. In a comparison across ten proteins, this approach found better computational scores with fewer runs than random sampling. After the first round, the authors can add machine learning, although the initial selection relies on molecular properties and computational results.

The authors selected FSP1 as a target because few inhibitors of this protein have a known protein complex structure. This enzyme protects cells from ferroptosis, an iron-dependent form of cell death triggered by lipid oxidation in cell membranes. The authors reconstructed an ancestral variant of human FSP1. Its reaction site remained conserved, and the protein crystallized readily. These properties allowed the authors to define a site for screening and then determine where a molecule bound. Following a preliminary screen of 12 million molecules and a main screen of 10 million molecules, 33 of the 42 ordered compounds were synthesized and tested. Two inhibited FSP1 activity by at least 50% at a concentration of 10 μM. For afi-FSP1-1, the inhibition constant, which measures the strength of enzyme inhibition, was 0,283 μM. The crystal structure showed how this molecule occupied the intended protein site.

PARP1, an enzyme involved in repairing single-strand DNA breaks, provided a second test of the method against a different target. After the main screen of 100 million molecules, the authors synthesized 160 candidates. Seven inhibited PARP1, and iParp1 had a half-maximal inhibitory concentration of 8,8 nM. Nuclear magnetic resonance measurements and an X-ray structure confirmed binding to the protein. iParp1 was less active in cell experiments because hydrolysis reduced its ability to cross the cell membrane.

The map of chemical properties determines where the program searches next. A subset of the candidates is then synthesized and evaluated in a biochemical assay. For selected molecules, the researchers also determine the structure of the protein complex.