What AlphaFold2 Actually Solved
It is worth beginning with an honest account of what the AlphaFold2 release in 2021 achieved, because the field has moved through several phases of reaction — initial astonishment, then overclaiming, then a more measured settling into what the tool is and is not. AlphaFold2 solved the protein structure prediction problem as defined by CASP: given a protein sequence, predict the three-dimensional fold of the monomeric protein to near-experimental accuracy for most protein families with detectable evolutionary information. That is a genuine scientific achievement, and the implications for structural biology are deep and ongoing.
For PPI drug discovery specifically, AlphaFold2 — and the subsequent AlphaFold2-Multimer extension — provided something the field has needed for a long time: structural models of PPI complexes for targets where experimental structure determination was intractable or had simply not yet been done. The PDB contains structures for a subset of oncology-relevant PPI complexes, but coverage is uneven. Highly studied targets like MDM2-p53 have dozens of PDB entries including multiple small-molecule complex structures. Targets like certain STAT3 interface conformations, or specific BCL-2 family heterodimers in their physiologically relevant states, are less well represented. AlphaFold2-Multimer fills some of those gaps — it provides a starting structural hypothesis for targets where no experimental structure existed.
That is a meaningful contribution. It is not, however, the same as predicting which interfaces are druggable, how hot-spot residues are arranged, or what compound properties will achieve productive engagement with the interface. Those questions require additional layers of analysis that AlphaFold2 was not designed to answer.
The Gap Between Structure Prediction and Interface Druggability
Interface druggability assessment requires information that is not encoded in a static structural model. The key questions are: Does a hot-spot sub-pocket exist at this interface with geometry suitable for a small molecule to occupy? How rigid is the interface geometry under physiological conditions — does it present a stable binding site or a highly dynamic ensemble of conformations? What is the energetic contribution of individual interface residues to the binding free energy, and are the highest-contributing residues geometrically accessible to a small-molecule disruptor?
None of these questions can be answered from a single static coordinate set, whether that coordinate set is experimental or predicted. Static structures capture one conformation from an ensemble. PPI interfaces are often among the more dynamic regions of protein structures — they have evolved to mediate transient, regulated interactions, not to be maximally stable. A hot-spot residue identified by alanine scanning in the crystal may be partially occluded or structurally rearranged in solution. The sub-pocket geometry that looks druggable in a PDB entry may be a crystal-packing artifact.
AlphaFold2's confidence metric (pLDDT) is informative about local structure quality, and the predicted aligned error (PAE) metric provides some information about interdomain confidence, including at interface regions. But neither metric maps directly onto druggability. High pLDDT at an interface region indicates confident structural prediction, not a stable binding site for a small molecule. Low pLDDT at an interface region may indicate genuine structural flexibility — or it may reflect limited evolutionary information in that region — but the drug discovery implication differs substantially between those two interpretations.
Where AlphaFold2-Multimer Helps and Where It Falls Short
For PPI interface modeling specifically, AlphaFold2-Multimer performs best on well-conserved obligate heterodimers — cases where the two protein chains co-evolve and have co-evolutionary signal available in the multiple sequence alignment. For transient, regulated interactions — many of the oncology PPI targets we care most about — co-evolutionary signal is weaker, and the model confidence at the interface drops correspondingly.
A practical example: consider a target where the PPI interface involves a short peptide motif from one partner engaging a structured domain on the other. The p53 transactivation domain peptide interacting with MDM2 is a canonical case. The structured partner (MDM2) is modeled accurately by AlphaFold2. The short peptide interaction — the N-terminal transactivation helix of p53 — is less well-constrained because the peptide itself is intrinsically disordered outside the complex context. AlphaFold2-Multimer can produce a plausible complex model, but the exact conformation of the disordered partner and the precise geometry of hot-spot contacts will have higher uncertainty than the confidence scores might suggest. For drug discovery purposes, we treat these as starting hypotheses to be refined by molecular dynamics ensemble generation, not as ground-truth binding geometries.
There is also the question of what structure to use as the input for interface characterization. For targets where multiple PDB structures exist — including apo, holo, and small-molecule-bound forms — the selection of the "right" input structure for hot-spot analysis is itself a non-trivial decision. The apo form may have a more open sub-pocket. The holo form may show interface contraction that makes the pocket less accessible. A compound-bound form shows the interface in a specific induced-fit state. For targets with abundant structural data, we use an ensemble-weighted approach rather than a single reference structure. For targets where only AlphaFold2-predicted structures are available, we are working from a single conformational hypothesis with correspondingly higher uncertainty.
What Needs to Be Built on Top of AlphaFold2
The computational PPI modeling field has identified several specific needs that structure prediction alone does not address, and that are the focus of active development:
Interface dynamics characterization. Molecular dynamics simulations seeded from AlphaFold2-predicted or experimentally determined complex structures reveal the conformational ensemble of the interface under physiological conditions. Hot-spot sub-pocket volumes and shapes fluctuate; the "druggable pocket" identified in a static snapshot may be present only a fraction of the simulation time. For PPI modeling to be useful, it needs to characterize the ensemble of accessible pocket geometries, not just the single lowest-energy structure.
Hot-spot prediction calibrated to PPI geometry. AlphaFold2 provides coordinates. Identifying which residues are hot spots requires separate computational analysis — alanine scanning ΔΔG estimation, residue conservation analysis relative to the binding interface, solvent accessibility calculation in the complex versus the unbound state. These analyses are well-established methodologically but need to be run explicitly on the predicted interface; they are not an output of AlphaFold2 itself.
Confidence-aware druggability assessment. When using AlphaFold2-predicted complex structures, the pLDDT and PAE metrics should feed directly into the confidence weighting of downstream druggability predictions. Interface regions with low PAE confidence should be flagged as higher-uncertainty in the hot-spot characterization. We are building confidence propagation into our workflow so that output reports reflect the underlying structural uncertainty rather than presenting all interfaces with equivalent authority.
The Practical Implication for PPI Discovery Teams
The key takeaway is not that AlphaFold2 is insufficient for PPI discovery — it is that AlphaFold2 is a starting point for PPI discovery, not an endpoint. Teams that have treated AlphaFold2 structure predictions as equivalent to experimental structures for docking and virtual screening campaigns have generally been disappointed by the hit rates, and the disappointment is predictable from the analysis above.
We're not saying AlphaFold2 structures are less useful than experimental structures for all purposes — for many targets, the overall fold quality is comparable to X-ray structures at moderate resolution, and the structures are genuinely useful for hot-spot analysis when handled appropriately. We're saying that the uncertainty at PPI interfaces needs to be explicitly characterized and propagated through the downstream analysis, and that static structures of any provenance need to be treated as one conformation in an ensemble rather than as definitive binding geometry.
The structural coverage of oncology PPI complexes has genuinely expanded since AlphaFold2's release. The druggability knowledge has grown more slowly, because druggability assessment requires the additional analytical layers described above. That gap — between structural availability and actionable interface characterization — is precisely the problem that purpose-built PPI modeling workflows are designed to close.