TorchCraft designs molecular binders, with laboratory evidence in two protein formats

The September 17 preprint applies this approach to small protein binders, single-domain antibodies, cyclic peptides and pockets intended to bind small molecules. The public TorchX implementation exposes the design process. [1] [2]

Selected minibinders and antibody domains showed measured binding, with campaigns covering four targets in each format. The cyclic-peptide and small-molecule-pocket demonstrations remain computational. The distinction matters: a predicted fit and a binding experiment support different conclusions, and neither alone establishes a useful medicine. [1]

Some computational comparisons reuse the pretrained prediction weights that guided design. The authors caution against reading these scores as experimental success rates: a useful selection score is not an independent test of the selected molecule. [1]

TorchX, the public implementation, gives researchers a route to inspect the design, prediction and scoring components. Its code is Apache-licensed, while model parameters have their own access arrangements. The repository inspected on September 28 distinguishes released inference from a separate training environment. [2]

For a first campaign, fix the target and selection budget, retain rejected candidates and measure binding independently. Track both the yield of useful binders and the cost of producing a measurable candidate; better computational scores do not settle either question.