The Enzyme-Inhibitor Monopoly on Drug Discovery
For most of the history of small-molecule oncology drug discovery, the dominant paradigm has been enzyme inhibition. Kinases, proteases, phosphatases — these targets offer deep, well-defined binding pockets that have been optimized over millions of compound-years of medicinal chemistry. The chemistry is understood. The ADMET space is partially mapped. The scoring functions correlate reasonably well with binding affinity.
That paradigm has produced real drugs. Imatinib, erlotinib, ibrutinib — the oncology kinase inhibitor roster is long and genuinely impactful. But it has also produced a landscape where we keep returning to the same structural families, the same druggable genome fraction, and increasingly, the same resistance mechanisms. When a tumor evolves around a gatekeeper mutation in an ATP-binding site, the biology hasn't changed — we've just run out of novel surface to exploit.
Protein-protein interaction surfaces offer a different kind of opportunity. The biology is, in many cases, more fundamentally upstream. And the reason the field hasn't moved faster on PPI targets isn't a failure of biological validation — it's a failure of the computational tooling that was brought to bear.
What Makes PPI Surfaces Structurally Distinctive
The classic characterization of PPI interfaces as "undruggable" was never really about the biology — it was about geometry. Enzyme active sites typically present a concave, enclosed binding pocket with volumes in the range of 300–800 ų, high shape complementarity with small-molecule ligands, and well-defined hydrogen bond donors and acceptors. The interface is, in a geometric sense, designed to interact with something small and specific.
PPI interfaces are different in almost every relevant dimension. The contact surface area between two interacting proteins is typically 1,000–3,000 Ų, compared to the 300–500 Ų that a typical drug molecule can span. The surface is often flat or convex rather than concave. The interaction energy is distributed across many residues rather than concentrated in a tight binding site. These characteristics make standard docking approaches perform poorly — the scoring functions reward deep pocket insertion rather than flat-surface complementarity, and the library design reflects that bias.
The critical insight that has shifted the field — and that guides our work at Genolux — is that PPI binding energy is not actually distributed uniformly across thousands of square angstroms of interface. It is concentrated in a small number of hot-spot residues, typically 3–5 side chains that contribute disproportionately to the binding free energy. Studies using computational and experimental alanine scanning across well-characterized PPI complexes consistently find that 20–25% of interface residues account for 80% or more of the binding ΔG. That is a druggable geometry — if you know where to look.
Oncology Biology Strongly Validates PPI Targets
The argument for PPI targets in oncology is not merely structural — it is biological. The PPI interfaces that have attracted the most attention are connected to pathway nodes that are genetically validated in human cancer at rates that match or exceed the most successful kinase targets.
The MDM2-p53 interaction is the canonical example. p53 is mutated or silenced in roughly half of all human cancers. In tumors where p53 remains wild-type, MDM2 overexpression is a major resistance mechanism — the tumor effectively sequesters functional p53 by promoting its ubiquitination and degradation through the MDM2 interaction. Disrupting MDM2-p53 restores p53 transcriptional activity without requiring direct p53 engagement. The interface is reasonably well-characterized structurally (multiple PDB entries for MDM2-peptide and small-molecule complexes are in the public domain), and the hot-spot architecture — particularly the deep sub-pocket occupied by the F19, W23, and L26 residues of p53's transactivation domain — has been mapped in detail.
The BCL-2 family tells a similar story. BCL-2, BCL-xL, and MCL-1 each mediate apoptotic resistance through protein-protein interactions with pro-apoptotic BH3 domain proteins. The venetoclax approval for BCL-2 demonstrated, definitively, that PPI disruption at a BH3-recognition interface is both achievable and clinically relevant. But venetoclax is a single narrow case — the MCL-1 and BCL-xL interfaces have distinct structural geometries and require different approaches.
The Wnt/beta-catenin pathway, the STAT3 SH2 dimerization interface, the KRAS/SOS1 interaction, the BRD4 bromodomain-MED1 contact — these are all biologically validated oncology targets with PPI interfaces at their mechanistic center. The bottleneck is not biological justification. It is computational tractability.
Why the Tooling Gap Has Persisted
Standard computational drug discovery workflows were not designed for PPI surfaces. Virtual screening pipelines using tools like AutoDock, Glide, or GOLD are optimized for binding pocket shape complementarity. When applied to PPI interfaces without modification, they consistently fail to distinguish compounds that engage the hot-spot architecture from compounds that dock to adjacent, non-functional surface regions. The enrichment factors at PPI interfaces, using unmodified standard docking, are frequently at or near random — which is why PPI programs that rely on general-purpose virtual screening produce poor hit rates.
The fragment-based lead discovery community made earlier progress on PPIs than the high-throughput screening community, precisely because fragment campaigns tend to reveal hot-spot sub-pockets that larger compounds miss. But fragment-to-lead optimization for PPI interfaces is slow and requires iterative structural information that is expensive to generate experimentally.
What has changed is the combination of structural data availability and computational power. AlphaFold2 and related structure prediction tools have substantially reduced the barrier to obtaining PPI complex structural models, though the gap between structure prediction accuracy and interface druggability prediction is real and non-trivial (a topic worth a separate treatment). Improved Rosetta interface scoring functions, molecular dynamics force fields calibrated for protein-protein contacts, and better pharmacophore methods for extended surfaces have all matured over the past decade. The tools to do this computationally are now good enough to be useful — the question is whether they are used in a workflow designed for PPI geometry, not adapted from enzyme-pocket workflows.
The Therapeutic Window Argument
We're not saying enzyme inhibitors have run their course — kinase inhibitor programs continue to produce clinically relevant compounds, and there is no shortage of genetically validated kinase targets in oncology. The argument for PPI surfaces is about the therapeutic window that opens up when you target a different mechanism.
Many of the most impactful PPI interfaces in oncology connect to transcription factor complexes, scaffolding proteins, or signaling hubs that are not enzymes and therefore have no classical active-site vulnerability. MDM2 has no enzymatic function relevant to the p53 interaction. c-MYC dimerization with MAX does not require catalytic activity. STAT3 SH2 domain-mediated dimerization is a protein-protein event, not an enzyme-substrate one. These targets are structurally off-limits to enzyme-inhibitor design, and they represent some of the most important oncogenic drivers in the most common cancers.
There is also a selectivity argument. Many kinase inhibitors show off-target activity at related kinase family members — this is well-documented and a major driver of kinase inhibitor toxicity profiles. PPI interfaces tend to be more structurally unique: the hot-spot residue architecture at MDM2-p53 is not reproduced at BCL-2/BH3 or KRAS/SOS1. Achieving interface selectivity at the molecular level may be more tractable for certain PPI targets than kinase selectivity is for certain kinase classes.
What Computational PPI Modeling Needs to Deliver
For PPI-focused drug discovery to benefit from computational pre-screening, the workflow has to do several things that standard virtual screening does not. It needs to identify hot-spot residue positions computationally and with enough accuracy to distinguish true hot spots from adjacent non-critical contacts. It needs scoring functions calibrated to PPI interface geometry — rewarding hot-spot engagement and surface complementarity rather than deep-pocket insertion. It needs library design biased toward PPI-relevant chemical space: higher molecular weight, higher lipophilicity, more extended shape, fewer ring systems than typical fragment screening collections.
And it needs to produce output that a medicinal chemist can act on — not just a ranked list of SMILES strings, but structural rationale for why a given compound is ranked where it is. Which hot-spot residue does it engage? What is the predicted contribution to interface disruption? Where is the SAR vector for optimization? Those are the questions that compress the timeline from computational output to synthesis decision.
PPI drug discovery is no longer a conceptual frontier. The biology has been validated for more than two decades. What the field is working through now is an execution gap — the need for computational methods purpose-built for interface surfaces rather than repurposed from enzyme-pocket discovery. That is the problem Genolux is built to address, and it is the thread that connects all the methodological work we publish here.