The Pulse
Pallatom2 designs show 53.3% hit rate in lab tests, preprint reports
Researchers report that 112 of 210 protein-binder candidates generated with Pallatom2 were active in laboratory tests. The preprint also describes computational screening and attempts to improve or recover binders.

Laboratory tests found active binders among 112 of 210 candidates designed with the Pallatom2 generative model, a 53.3% hit rate, according to a bioRxiv preprint. The authors, all affiliated with Hangzhou Levinthal Biotechnology, report that they obtained binders for all ten targets evaluated experimentally.
The preprint was posted on November 12. It describes both the experimental binding results and computational tests of the model's ability to generate proteins and screen potential binders.
One model for sequence and structure
Pallatom2 jointly generates protein sequence, backbone conformation, side-chain geometry and contacts at the target interface, rather than treating them as disconnected design stages. Its architecture uses an Atom14 representation and residue-guided sparse all-atom attention to exchange information between residue-level and atom-level representations.
The model has 65.9 million parameters. Its conditioning system controls binder-target engagement and protein conformations at both global and local scales, the authors say.
In computational evaluation, the researchers report that the model generated long-chain monomers, multi-domain proteins and homomultimers, with monomer sampling extending to 1,600 residues. Selected symmetric homomultimer designs spanned C3 to C10 symmetry and assembly lengths of 600 to 2,000 residues, according to Levinthal's accompanying research repository.
Binding results and benchmarks
The experimental measurements used biolayer interferometry to identify active designs and measure apparent binding affinities. Six targets produced picomolar-scale apparent affinities, the preprint reports. The reported apparent K_D values included 3.96 pM for SC2RBD and 33.8 pM for PD-L1; TNFα reached 1.166 nM.
The ten evaluated targets were BCMA, PD-L1, VEGF-A, SC2RBD, IL-7RA, BHRF1, TNFα, TrkA, IL-17A and H1. Across six shared targets, Pallatom2 had lower median apparent K_D values for five targets and lower minimum values for four, according to the repository.
In computational binder screening, the researchers say helix-rich Pallatom2 sampling achieved the highest reported screening success across all nine targets for which the report had cross-method data, under the protocols used. The repository presents those computational results separately from the experimental binding measurements.
Reworking existing designs
The work also tested ways to modify designs beyond generating binders from scratch. For BCMA, the team reports that a design process guided by a starting interface produced five active designs among six tested. Those active designs had 12.5% to 37.3% sequence identity with BC40 while retaining association rates near the starting value; one, PB-006, reached an apparent K_D of 252.1 pM.
In a separate TNFα affinity-maturation experiment, partial diffusion improved affinity by up to 7.57-fold, with 21 active variants among 23 tested, according to the repository. The authors also report that remodeling the CDR-associated interface of an inactive de novo nanobody produced two active variants with apparent K_D values of 4.31 and 5.93 nM.
The supporting materials include biolayer interferometry plots and disclosed active minibinder sequences. IL-7RA, BCMA and nanobody design sequences are not disclosed in those materials.