Protein designers showed what their tools can build — and where they still fail

Ligases that pick their spot

A preprint describes enzymes designed from scratch to attach ubiquitin, a small protein tag, to a selected lysine on a target protein. The designs were built with RFdiffusion3, a computational design tool, and use a cysteine-based catalytic triad to do the joining. In the tests reported, the designed ligases converted 90% of unactivated ubiquitin to PCNA K164, a specific site on a specific protein, and the paper reports that this outperformed natural promiscuous E2 enzymes. m:…6113

Selectivity is the point. The paper reports that the designed ligases kept high site-selectivity even where numerous off-target lysines were present, which is the hard part of the job: a tag attached at the wrong lysine changes a different protein, or the same protein in the wrong way. m:…6113

Another preprint reports a related platform: de novo amide ligases, also built around cysteine catalytic triads, computationally programmed for high selectivity across diverse substrates. The designed ligases selectively functionalized the N-terminus or a single lysine residue in a polypeptide containing 36 competing lysines. The same platform enabled the stereoselective synthesis of lenacapavir, an HIV therapeutic, which is the kind of concrete output that moves a design method from a demonstration to a tool. m:…2837

Rules built into the generator

A ChemRxiv preprint takes a different route to the same problem: instead of designing a protein at a time, it builds the design rules into the generator itself. The method, called IH-LM, generates 773 valid, novel cyclic peptides per 1,000 samples, and embedding design rules directly into the decoder guarantees validity by construction for cyclic peptide generation. At a matched sampling budget, it produces three times as many distinct scaffolds as tuned baselines. m:…0993

The claim that matters here is not the count but the guarantee. A generator that can only produce valid designs wastes no sampling budget on shapes that cannot exist, and distinct scaffolds are the raw material for finding a peptide that binds a new target. This suggests the field is shifting effort from generating more candidates to generating only candidates that pass the rules. m:…0993

The misses are part of the result

A comparison of binder-design tools BindCraft2 and RFD4-Proteina reports a split verdict. RFD4-Proteina is more cost-effective than BindCraft2 for de novo peptide binder design against small, contiguous insecticide targets. But both tools failed to produce validated binders for the ecdysone receptor, known as 1R1K, and the ryanodine receptor, known as 6KIM. m:…3852

The failure is reported plainly, and it is useful. Cost-effectiveness on small, contiguous targets does not carry over to harder receptors, and a field that publishes its misses alongside its wins gives anyone choosing a tool a truer picture of where each one works. The implication is that binder design is not one solved problem but a set of target classes, each with its own difficulty. m:…3852

Why it matters

Taken together, the ligase papers point at a practical goal: enzymes that attach a tag exactly where asked, on demand. Conversion of 90% of unactivated ubiquitin to a chosen site, and selective functionalization against 36 competing lysines, are the kind of numbers that make a method usable in a real workflow rather than a demonstration. The lenacapavir synthesis shows the same platform reaching an actual therapeutic molecule. m:…6113 m:…2837

The peptide papers add the rest of the picture: generation is getting more disciplined, with validity guaranteed by construction and three times as many distinct scaffolds per budget, while the binder comparison shows the limits are still target-dependent. For anyone deciding where to put effort in protein design, the message is to ask for the build test: what did the tool make, at what selectivity, and where did it fail. m:…0993 m:…3852