Skip to main content
Back to timeline
bioRxivSource publication:

Pallatom2 designs binders atom by atom: 112 of 210 candidates bind, and six targets reach picomolar affinity

Synopsis

Protein binder design has long been limited by the gap between plausible structures and wet-lab binding. Pallatom2 puts sequence, backbone, side chains and target-interface contacts into one atom-level generative model, using residue-guided sparse all-atom attention to keep the computation tractable in a model of just 65.9 million parameters. Across ten targets, 210 candidates were tested by biolayer interferometry and 112 were active, 53.3%, with six targets reaching picomolar apparent affinity, including 3.96 pM for SC2RBD and 33.8 pM for PD-L1, while the harder TNFα reached 1.166 nM.

AI-generated editorial illustration: Pallatom2: Navigating the De Novo Protein Universe with Atom-to-Atom Modeling

Interpretation

A protein binder can now have its backbone, side chains and target-interface contacts settled together during generation, instead of generating a backbone first and filling in sequence and side chains afterward. Earlier routes either generated only the backbone and left side-chain conformations and specific atomic contacts to later steps, or fixed the backbone before sequence design, so backbone conformation, residue identity and side-chain arrangement could not be optimized together. Across ten targets and 210 candidates that completed biolayer interferometry, 112 were active, 53.3%, and binders were obtained for all ten targets; six targets reached picomolar apparent dissociation constants, with 3.96 pM for SC2RBD, 33.8 pM for PD-L1, 110 pM for VEGF-A, 182 pM for IL-7RA, 187 pM for BCMA and 828 pM for H1.

The model describes protein structure directly with atom-level single and pair representations, so backbone geometry, side-chain conformation and interface contacts are updated within one representation. Dense all-atom attention must process a number of atom pairs that grows quickly with protein size, and that computational and memory cost is what has kept atom-level modeling from scaling to larger structures. Residue-level correlations first identify strongly related residue pairs, and those residues define the local groups within which sparse all-atom attention is computed, in a model of 65.9 million parameters; on the nine targets with cross-method data, mainly α-helix sampling gave the highest screening success rate, ranging from 35.6% for TNFα to 98.5% for BHRF1.

The same generative framework also performs interface transfer, affinity maturation and rescue of an inactive molecule, not only de novo discovery. Beyond de novo design, reworking an existing binder usually means another design round or reliance on experimental screening, and whether the starting interface's kinetics survive is uncertain. In BCMA design conditioned on side-chain geometry at eight PR-001 interface sites, five of six designs were active and all kept association rates at the starting order despite only 12.5%–37.3% sequence identity; partial diffusion on TNFα improved the best variant PB-153 by 7.57-fold; and CDR interface remodeling of nanobody PB-272, which showed no binding at 2000 nM, produced PB-276 and PB-277 with apparent dissociation constants of 4.31 and 5.93 nM.

Perspective

The demonstrated scope is minibinder design on ten targets, BCMA interface transfer, TNFα affinity maturation and nanobody rescue, all ending at apparent dissociation constants measured by biolayer interferometry; the 53.3% active fraction applies only to the 210 candidates that completed testing, and 74 of the broader 284-candidate cohort were not expressed.

The apparent affinities still need complementary characterization of specificity, stability, aggregation and target-dependent function, and experimental structures have not yet tested whether the intended atomic contacts are realized; separately, the contributions of training-data preparation, architecture and sampling strategy to design success were not isolated and call for controlled ablations.