PNAS · 2026 WANG & DOKHOLYAN OPEN ACCESS
Structure prediction is close to solved. Finding the pocket on that structure is not. It is a separate problem, it sits upstream of every docking run and every screen, and getting it wrong cannot be recovered downstream. YuelPocket's answer is one extra node in the graph.
THE QUESTION
The introduction, unpacked
Two hundred million protein structures are now a database lookup. That changed what the hard part is. If you want to drug a protein you still have to answer a question the structure does not contain: out of a surface with thousands of atoms, which patch actually binds a small molecule, and which small molecule.
This paper is unusually direct about why that matters. Its significance statement argues that failures in virtual screening and lead optimisation are often caused by identifying the wrong site in the first place, rather than by any weakness in the docking algorithm or the scoring function that runs afterwards. That is a claim about where in the pipeline the error lives, and it is worth taking seriously, because errors at the front of a pipeline do not average out. They propagate.
Why the first step is the expensive one
Docking searches poses inside a box you give it. Scoring ranks what docking found. Screening ranks what scoring scored. None of those stages can tell you that the box was in the wrong place. They will return a confident, well-ranked, entirely useless answer.
But AlphaFold 3 places ligands now
AlphaFold 3 predicts protein and small molecule together, which looks like it should retire this whole problem. The paper cites an assessment by Shen and colleagues finding that while the receptor structure comes out reliably, the position of the small molecule is variable and frequently inaccurate. The gap is widest exactly where it hurts most: allosteric modulators, the sites you go looking for when the orthosteric pocket is undruggable.
So the two problems separate again. Predicting the fold and predicting the functional site are not the same task, and solving the first did not solve the second.
Receptor backbone comes out reliably
Ligand placement variable, worst on allosteric modulators
Schematic. The bars stand for the qualitative gap the paper describes, not measured values.
What came before
Select a generation to see how it looks at a protein, and where that view runs out.
The other half of the problem
THE MOVE
One node, added on purpose
The obvious way to let a protein and a ligand talk to each other in a graph is to connect every residue to every atom. It works, and it is unusable: with m residues and n atoms you have written down m × n edges, and that number grows faster than any real protein lets you afford.
YuelPocket puts a single virtual node between the two sides. Every residue connects to it, every ligand atom connects to it, and nothing connects directly across. Information still flows both ways, in two hops instead of one, and the edge count collapses from m × n to m + n.
The whole argument, in one slider
Both designs carry the same structural edges: backbone links, residue contacts, chemical bonds. That shared count is written as C in the paper. What differs is how the two sides are joined.
The whole design on one page. Panels A and B are the two molecules as graphs; panel C is the entire contribution of the paper.
Reproduced from Wang & Dokholyan, PNAS 2026, figure 1, under CC BY 4.0.
Each residue becomes a backbone node at the alpha carbon and a side-chain node at the geometric centre of its side chain, in the spirit of the UNRES coarse-grained force field. Coarse enough to stay robust to small errors, detailed enough to keep side-chain geometry in play.
Each edge carries the Euclidean distance plus four binary flags: is this a protein contact, a chemical bond, a protein-to-virtual link, or a ligand-to-virtual link. Nodes carry a one-hot identity and a three-way mask marking them protein, ligand or virtual.
One head asks whether this protein and this ligand belong together at all, trained contrastively against 50 decoy ligands. The other predicts per-residue pocket probability from the element-wise product of the residue embedding and the ligand's virtual node. Dice loss switches on only once the F1 score passes 0.2, so it refines rather than destabilises.
TWO MODES
Two different questions, two different answers
These are not the same output, and the paper is right to separate them. A list of contact residues tells you about mechanism. A set of coordinates tells docking where to put its box. YuelPocket runs both from the same trained network.
The coordinate mode reformulates pocket finding as a ranking problem. Shrake-Rupley rolls a 1.4 Å solvent probe over the protein and every surviving surface point becomes a candidate pocket centre. Each candidate gets its own small graph: nearby residues within 10 Å, all ligand atoms, and the probe itself acting as a local hub, exactly like the virtual joint node does globally.
Then those raw scores have to become discrete sites. Each surface point climbs toward its highest-scoring neighbour, points that converge on the same peak get merged with a union-find structure, and the peaks become ranked pocket centres. It is a small piece of engineering, and it is what turns a cloud of numbers into something a docking program can accept.
THE EVIDENCE
Tested where it is hard to cheat
Training and test are separated by 30% protein sequence similarity, 30% protein-ligand interaction similarity and 50% ligand Tanimoto similarity, all at the same time. 309,140 systems for training, 832 for validation, 1,036 held out. The authors also dropped COACH420 entirely after finding most of it already sat inside PLINDER's training data, which is the kind of decision that costs you a favourable number and buys you a trustworthy one.
Residue level
The model returns a probability for every residue, so success has to be defined against two dials: how many guesses you allow, and how close counts as close. A prediction succeeds if at least one of the top K residues lands within D ångström of any ligand atom.
Which dial setting is the real number depends on what you are doing. Top-1 at 4 Å is what to quote if you get one shot at the site. Top-10 at 8 Å is what matters if you intend to dock into everything plausible and let scoring sort it out afterwards.
What the output actually looks like
Curves say how often it works. A coloured structure says what "working" looks like, and the second of these two is doing more than decoration. A large share of real binding sites form at the interface between subunits, where the pocket is made of residues contributed by two different chains.
That is the exact case a local graph handles worst, because the two halves of the site may be far apart along every path the graph knows about. Showing a dimer here is the paper quietly answering the objection it raised in its own introduction.
Two ways to be right
Coordinate predictions get scored with two different distances, and the difference matters more than it sounds. Distance to Closest Atom measures from the predicted centre to the nearest atom of the real ligand: it asks whether you landed on the site at all. Distance Centre-to-Centre measures to the ligand's centroid: it asks whether you also read the shape of the site correctly.
The second is strictly harder and the gap shows it. You can be inside the right pocket and still put its centre in the wrong place, which is precisely the error that sends a docking box off target.
Against the standard baseline
P2Rank is the sensible thing to measure against: widely used, template free, fast. On the hard split YuelPocket is ahead on both distance measures, by roughly seven points at Top-1 within 4 Å on the first and five on the second.
Note how much lower every DCC curve sits than its DCA counterpart. That gap is the same one the diagram above describes, and it is a property of the task rather than of either method.
And on the easier one
This is the panel a less careful paper would have left out. On Holo4k the curves converge and the advantage mostly disappears, and rather than bury that, the discussion explains it: PLINDER's construction discards a large slice of the archive on similarity grounds, so the Holo4k systems that survive filtering are likely drawn from that discarded pool and sit closer to training data.
Read together, the pair says something more useful than either alone. The harder the split, the bigger the gap, which is the shape you want if the claim is about generalisation rather than fit.
The same protein, two methods
Curves aggregate away the thing you most want to see, which is what a good and a bad prediction look like side by side. Step through the four states: raw surface scores first, then what survives clustering, for each method in turn. The dashed circle marks the true site.
YuelPocket produces a dense, confident cloud that collapses onto the right place. P2Rank's scores on this system are sparse enough that the clustered centre drifts off the ligand entirely.
The test that actually matters now
The reason this is useful in 2026 is that most targets arrive as predicted structures. So the authors ran AlphaFold on all 1,036 PLINDER test proteins and evaluated on those instead of on the experimental ones.
Performance holds
Top-3 residue success stays above 70% at 4 Å and Top-10 above 90%. The interesting part is the coordinate comparison: the lead over P2Rank grows to roughly ten percentage points on predicted structures, larger than on experimental ones.
That direction makes sense given the design. A coarse-grained representation that deliberately ignores fine atomic detail should degrade more gently when the atomic detail is itself a prediction.
Where the errors come from
The middle panel is the one worth pausing on. Success dips as the predicted structure drifts from the crystal, then stops falling and flattens, which is not what you would expect if bad folds were driving failure.
The authors' reading: at low deviation the damage is side-chain placement, which the model is partly sensitive to despite the coarse representation. At high deviation the error is usually how subunits sit relative to each other in an oligomer, while the individual monomers, and therefore their pockets, are still fine.
One case, both ways
The same pocket is identified on the experimental structure and on the AlphaFold model. Subtle side-chain rotations between the two shift the predicted probabilities slightly without moving the site, which is the concrete version of the trend in the panels above.
Figure 4G and 4H · PDB 4I4V
PROBES
A consequence the authors turned into a method
Because YuelPocket conditions on a specific molecule, the same protein returns different pockets for different inputs. That could be filed as an inconvenience. Instead the paper turns it into a tool: pick a small set of molecules that between them light up as many distinct sites as possible, and use them as probes.
How the set is built
Start with nothing. Each round, score every candidate ligand by how many previously uncovered pocket residues it would newly reach across a batch of proteins, add the winner, repeat until coverage passes threshold. No learning, no optimisation, just a coverage argument.
Fifteen molecules is where it stops being worth adding more. The curves show the return diminishing steadily rather than cliff-edging, which suggests the set could be trimmed or extended depending on how much compute you want to spend per target.
What got chosen
This is the part worth staring at. The top three probes are not designed, not privileged, not chemically exotic. They are ordinary ligands lifted out of ordinary structures, and their value is purely that their pocket predictions happen to be complementary to each other.
Which tells you the property being exploited is not about the molecules at all. It is about what the network has learned to key on when it conditions a prediction, and a probe is just a handle for steering that.
1D1V, asked twice
Query this protein with H4B, the ligand it actually crystallised with, and the model confidently marks the H4B site and essentially ignores the PTU site. Query it with KI2 from the probe set instead, and both light up. The H4B site loses a little confidence in exchange, which the authors accept as a fair trade for finding a pocket you would otherwise never have looked at.
THE EDGES
Read the discussion, not just the abstract
The paper is unusually willing to explain away its own good numbers. That is worth more than the numbers.
Holo4k scores higher than PLINDER throughout. Rather than lead with the better figure, the authors explain it: PLINDER's construction throws out a large fraction of the archive on similarity grounds, so the 300 Holo4k-only complexes are likely drawn from that discarded pool and sit closer to training data. PLINDER is the lower bound and should be read as the real one.
Two nodes per residue buys robustness to imperfect structures. It also throws away atomic detail that sometimes decides which residues line a pocket. The authors note that dropping to backbone alone would simplify the model further and would probably cost accuracy, and leave the trade open.
48% on PLINDER. The headline numbers are Top-3 and Top-10, and those are legitimate for a workflow where you dock into several candidate sites. They are not the same as getting it right first time, and a reader skimming the abstract could easily come away thinking they are.
The 1DY4 example in the supplement shows the same protein returning quite different maps for different query ligands, with shallow sites producing weak signal. That is arguably correct behaviour, since shallow sites do bind more weakly. It also means a low score is ambiguous between "not a pocket" and "a poor pocket".
Conditioning on the ligand is the source of most of what is interesting here, and it is also a requirement. For a fresh target with no known binder you are choosing probes rather than reading off an answer, and the fifteen-probe set is a starting point rather than a guarantee of coverage.
THE SHORT VERSION
Strip away the benchmarks and the argument is small and structural. Everyone agreed graphs were the right representation for a protein and a ligand together. The blocker was that joining them properly costs m × n edges, so the field built local graphs instead and quietly gave up on long-range cooperation. Routing both sides through a single node makes the global graph affordable, and once it is affordable the model can be trained on three hundred thousand systems rather than a few thousand. The binding affinity correlation nobody asked for is the sign that the representation learned something real.
READ IT YOURSELF
Published open access under a Creative Commons Attribution 4.0 licence, which is also why every figure on this page can be shown here with attribution. Mirrored locally so the download works regardless of what PNAS is doing.
COLOPHON
An interactive reading of Wang & Dokholyan, PNAS 2026. All numbered figures are reproduced from the paper under its CC BY 4.0 licence. Diagrams in the opening section are drawn for this page and are schematic, not reproductions. The work was supported by the National Institutes of Health (1R35 GM134864, R01 AT012053, R01 AG071675) and the National Science Foundation (2210963).