Antibody and drug design are learning to skip the structure step

Design without structures

Three preprints posted this week attack the same bottleneck in drug and antibody design from three directions: knowing how strongly two molecules bind, without waiting for a structure or a surface assay. The first is LangAAI, a model that conditions exclusively on antibody and antigen primary sequences to model complementarity-determining regions without 3D structures. The conditioning is not a trick of the training data, the authors argue: LangAAI exhibits genuine antigen conditioning, evidenced by reduced sequence recovery in the CDR3 region when cognate antigens are replaced with decoys. And it matches the performance of structure-based methods and inverse-folding models in affinity ranking benchmarks. If that holds, the structure step that used to gate antibody optimization becomes optional rather than mandatory. m:…3973

The second preprint upgrades the physics side. The gmx_MMPBSA suite for end-state binding free energy calculations now supports GBNSR6 and nonlinear Poisson-Boltzmann models, and includes a new amber_MMPBSA entry point for native AMBER input files. The updated Python API and GUI facilitate large-scale dataset analysis and cross-system comparisons. The practical point is throughput: free-energy calculations have historically been fiddly, single-system affairs, and the update is aimed at running them across datasets and comparing systems rather than babysitting one calculation at a time. m:…7050

The third preprint attacks measurement itself. The nanoTAK method, which measures binding in solution by nano temperature alteration kinetics, resolves binding kinetics for interactions that re-equilibrate in under 1 second. It provides thermodynamic parameters, enthalpy and entropy, alongside kinetic rates from a single ligand dilution series. And it successfully differentiates binding kinetics between binary and ternary complexes in a VHL-PROTAC-BRD4 system. That last detail matters because PROTACs work by forming ternary complexes, and a method that cannot tell binary from ternary binding cannot evaluate them properly. m:…7311

Why the combination matters

Taken together, the three tools sketch a faster loop. A sequence-only model proposes antibody variants without structures, a free-energy suite scores candidates at scale, and an in-solution method verifies the kinetics of the winners, including fast and multi-component interactions that surface-based assays struggle with. No single tool removes the wet lab, but the loop shrinks the part of discovery that used to be gated on solved structures and slow assays. For teams building design pipelines, this suggests the scarce resource shifts from structure determination to candidate ranking and verification, which are exactly the steps these preprints try to speed up. m:…3973 m:…7050 m:…7311

There is also a useful discipline in the details. LangAAI's decoy test is the kind of control that separates real antigen conditioning from memorization, and nanoTAK's ternary-complex result is the kind of demonstration that separates a general kinetics method from one that works on the hard cases. Both are worth borrowing as evaluation habits: test the model against decoys, and test the assay on the interaction type you actually care about. A benchmark win on easy cases is cheap; a win on decoys and ternary complexes is the expensive kind, and it is the one that predicts whether a tool survives contact with a real pipeline. m:…3973 m:…7311

What to keep in mind

All three results are preprints, and their headline numbers come from their own authors' benchmarks. LangAAI matching structure-based methods is a benchmark claim, not a head-to-head on a prospective design campaign. The free-energy suite's new models and interface are capabilities, and capabilities only matter if they change results on real systems. And nanoTAK's sub-second resolution was demonstrated on the systems its authors chose. The honest read is that the tooling for binding prediction is converging on a faster, structure-light loop, while the proof that the loop produces better drugs and antibodies is still to come. m:…3973 m:…7050 m:…7311