Defining the Rules of Protein Interactions

Some protein surfaces readily engage partners, while others stay stubbornly inert. Understanding what makes one surface welcoming and another “silent” is central to both basic biology and drug discovery. In new work led by Aerin Yang (Stanford), Hanlun Jiang (Berkeley) and Kevin Jude (Stanford) published in Science, the Garcia (Stanford) and Listgarten (Berkely) groups construct protein–protein interfaces on normally nonbinding surfaces. By following this artificial evolution, the scientists were able to understand how new interfaces form on surfaces with no evolutionary binding imprint, and why natural sites differ from silent regions that appear counterselected for binding.

Turning nonbinding surfaces into binders

Many successful drugs work by latching onto natural protein binding sites, places on the protein surface that evolution has already tuned to recognize other molecules.  By contrast, flat or featureless surfaces that never evolved to bind anything are notoriously hard to target. To probe the difference between these “druggable” and “undruggable” regions, the team used a small, well‑studied protein domain called the Z domain, which normally binds antibodies through one face formed by two helices (H1–H2).  They then deliberately moved the interaction surface to a completely different region—between helices H2–H3 on one partner and H1–H3 on the other—areas that do not usually contact any ligand.  At these naïve surfaces, they built large libraries in which selected surface residues were randomized to hydrophobic amino acids, creating countless potential new interfaces.

Using yeast display and a clever protease‑cleavage assay, they could enrich for Z‑domain pairs that stuck together.  If the two proteins interacted, a fluorescent tag remained attached after cleavage; if not, the signal disappeared.  Across successive rounds of selection, fluorescence increased as better‑binding pairs took over the population, mimicking how natural evolution gradually improves affinity and specificity.

Many solutions, shallow energy wells

When the researchers sequenced the enriched libraries, they did not see a single dominant “best” sequence.  Instead, their data formed seven distinct sequence clusters, or families of variants that had each discovered a different way to make the two proteins stick.  Network analysis of these sequences showed that both binding partners coevolved within each cluster, pointing to genuine bidirectional interface formation rather than one side adapting to a fixed surface on the other.

From six of these clusters, the team solved high‑resolution crystal structures of representative complexes.  Strikingly, the docking orientations varied widely, like clock hands frozen at different hours around the same hub, even though overall polarity (which end of each protein faced which) was preserved.  This contrasted with their earlier work on natural Z‑domain interfaces, where different evolved binders converged on nearly the same docking geometry, suggestive of a deep energy minimum that evolution had “locked in.”

To quantify this idea, they trained a machine‑learning–based “Selection Probabilistic Model” on sequencing data from all selection rounds, turning enrichment into an estimated fitness (and thus energy) landscape.  Simulations of in silico coevolution across this landscape revealed a single dominant deep energy well for the natural interface, but multiple shallower wells for the synthetic, nonnatural surfaces.  Variants occupying these synthetic wells bound in the sub‑ to low‑micromolar range, in contrast to the deep‑well natural variant at nanomolar affinity and a landscape with larger energy gaps between optimal and suboptimal binders.

Epistasis, “seed” contacts, and how interfaces start

The study goes further by asking how specific residue interactions shape these landscapes. Using the learned fitness model, the authors computed epistasis, or how the effect of one mutation depends on another, and related it to the physical positions of residues in the solved structures.  Strong intrachain epistatic pairs tended to sit on the same helix, suggesting that favorable mutations first preorganize each partner’s surface before fine‑tuning the contact between them.  Over successive rounds of selection, the relative importance of interchain epistasis diminished, consistent with a progression from finding a partner to optimizing local packing and stability.

By cataloging all close (<4 Å) contacts at the interfaces, the authors classified them as interactions between variable “library” residues on both sides (LL), between library and fixed framework residues (LF), or between frameworks (FF).  In both natural and synthetic interfaces, the core is dominated by LL contacts, representing hotspots where library residues from each protein meet.  Around these hotspots, LF contacts form, and FF contacts appear only at the outskirts of the interface.  Synthetic surfaces showed a broader and generally higher ratio of LL to LF contacts than natural surfaces, consistent with more ways for library positions to find each other but with less deep optimization into a single configuration.

Importantly, machine‑learning–guided analysis of the fitness landscape highlighted specific “seed” contacts: particular residue pairs that appear early, sit at the core of the interface, and are preserved along many evolutionary trajectories.  These seeds effectively anchor the initial encounter between two previously nonbinding surfaces, while the rest of the interface evolves around them.  Different seeds can lead to different docking modes, explaining how multiple structural families arose from the same starting scaffolds.

Implications for undruggable targets

These results support a compelling view of why many protein surfaces resist drug binding. Regions that have evolved as natural interfaces sit atop deep energy wells. Once evolution finds a good solution, it is strongly favored and becomes a magnet for binders, including drugs and biologics.  In contrast, “silent” surfaces appear to have been evolutionarily counterselected to avoid unwanted interactions, resulting in shallow, fragmented energy landscapes where tight binding is disfavored and multiple modest‑affinity modes are possible.

For drug discovery, this suggests that some undruggable surfaces may not merely lack a convenient pocket; they may be fundamentally shaped by evolution to avoid stable interactions.  Overcoming this barrier could require designing molecules that either introduce new “seed” contacts or exploit transient, pivoting docking modes rather than trying to force a natural‑like deep minimum where evolution has actively prevented one from forming.