← Projects  ·  Accepted at LoG 2026 · Learning on Graphs Conference

EquiSite

Equivariant graph neural networks for end-to-end prediction of multiple ligand binding sites

Animesh Animesh, Plaban Kumar Bhowmick, Pralay Mitra

Indian Institute of Technology Kharagpur

Real model outputs for UniProt Q939U1 (260 residues, 2 sites), decoded in your browser. Drag to rotate.

In one paragraph

A protein can bind ligands at several sites, so predicting them means predicting a set of residue masks. EquiSite does this with no clustering, no learned queries and no Hungarian matching. Every labelled site names one of its own residues as its anchor; every residue scores itself as a candidate anchor; each selected residue paints its own mask.

0.467 vs 0.384residue-IoU AP@0.5 on UniSite-DS, against UniSite-3D
27× fewertrainable parameters: 1.2M against 32.8M

Part 1 · The task

What is being predicted?

A ligand binding site is the set of residues lining a pocket where a small molecule docks. Finding these sites is an early step in structure-based drug design.

Many proteins have more than one. In UniSite-DS, the benchmark used throughout this page, 31.9% of the 11,510 proteins have two or more labelled sites (1.58 on average, up to 38). So the output is not one label per residue. It is a set: an unknown number of sites, each a subset of the protein's residues with a confidence score.

Three renderings of the same protein: the folded chain; the chain with two ligands bound; and the two binding sites, each highlighted as the set of residues lining its pocket.
From structure to sites. The input is the folded chain. The target is the set of residues around each bound ligand: here two sites, which touch.

Formally, the model maps a protein graph \(\mathcal G\) with \(N\) residues to

\[ f_\theta:\ \mathcal G \;\longmapsto\; \big\{(s_k,\ \hat{\mathbf z}_k,\ \hat{\mathbf m}_k)\big\}_{k=1}^{\widehat K}, \qquad s_k\in[0,1],\quad \hat{\mathbf z}_k\in\mathbb R^{3},\quad \hat{\mathbf m}_k\in\{0,1\}^{N} \]

where each element is one predicted site: a score, a 3D center and a residue mask. The number of sites \(\widehat K\) is itself an output.

Two things make this harder than classifying residues one by one.

  • Sites touch and overlap. In UniSite-DS, 6.9% of binding residues belong to more than one site. A pipeline that first labels residues as binding or not, and then clusters them, has to pull apart pockets that share residues.
  • A set has no order. Before the loss can compare a prediction with a labelled site, something has to decide which prediction is responsible for which site.

We score predictions as UniSite does. Predictions from all test proteins are ranked by confidence; a prediction counts if its residue-IoU with a labelled site that has not been claimed yet reaches a threshold (0.5 or 0.3); and the ranked list is summarised by average precision (AP).

Part 2 · The idea

Who is responsible for which site?

Every set-prediction model needs a rule that pairs predictions with labelled sites during training. UniSite, the current state of the art, follows DETR: a fixed set of learned queries each propose a site, and Hungarian matching pairs queries with labelled sites by solving an assignment problem on the current predictions, at every training step. The rule is general. It assumes nothing about what the objects are.

EquiSite explores a more task-specific alternative. A binding site is not an arbitrary region of space; it is a set of residues. So each site can designate one of its own residues as its representative: the member residue nearest the site's centroid. We call it the site's anchor. That residue is responsible for the site, and the pairing is read off the labels once, before training starts.

Two ways to decide who is responsibleInteractive schematic

Learned queries + Hungarian matching

The pairing is solved again from the current predictions at each step.

Residue anchors (EquiSite)

The pairing is a property of the labels. Nothing is solved.

step 0 ·
A schematic, not a simulation of either model. On the left, costs between queries and sites drift from step to step and the optimal one-to-one assignment is recomputed each time, so the query responsible for a site can change. On the right, each site is tied to its anchor residue for the whole of training.

Key idea

Hungarian matching is unnecessary once a deterministic, residue-based canonicalization of the sites is imposed. Detection then becomes a question that every residue answers about itself: am I the anchor of a site?

How is the anchor chosen?

The rule is two lines. Compute the site's centroid, then take the member residue closest to it. If two overlapping sites would choose the same residue, the larger site chooses first and the other takes its next-nearest member, so anchors stay distinct.

The anchor is a function of the labelInteractive · drag a residue
A toy chain in two dimensions. Two sites (teal and amber) share one residue. Drag residues around: each centroid (✕) follows its members, and the anchor (ringed) is the member nearest to it. The thin arrows are the vote targets used later in training: each site residue points at its site's centroid.

On UniSite-DS the anchor sits a median 4.0 Å from its site's centroid (90th percentile 5.9 Å), roughly a third of the way from the centroid to the edge of a typical site. The tie-break rarely matters: it changes the chosen residue in fewer than 1% of cases.

Part 4 · Supervision

Training without matching

Because every labelled site has named its anchor \(c_k\), every training target is fixed before the first gradient step. From the site masks \(\mathbf m_k\) and the anchors we build four targets, once, offline:

  • Centerness \(q^\star_i\): 1 on the \(K\) anchors and 0 elsewhere. One positive per site among hundreds of residues, so its focal loss carries the largest weight.
  • Objectness \(y_i\): 1 if residue \(i\) lies in any site.
  • Vote \(\mathbf v^\star_i\): the offset from a site residue to its site's centroid.
  • Masks: for each labelled site, the mask head is conditioned on that site's true anchor, and its output is compared with that site's mask alone.
\[ \begin{aligned} \mathcal L \;=\;& \underbrace{w_q\,\mathcal L_{\text{focal}}(q,q^\star)}_{\text{centerness}} \;+\; \underbrace{w_{\text{obj}}\,\mathcal L_{\text{focal}}(p,y)+w_{\text{dice}}\,\mathcal L_{\text{dice}}(p,y)}_{\text{objectness}} \;+\; \underbrace{w_v\,\mathcal L_{\text{sL}_1}(\mathbf v,\mathbf v^\star)_{\text{fg}}}_{\text{vote}} \\[0.4em] &+\; \underbrace{\frac1K\sum_{k=1}^{K}\Big[2\,\mathcal L_{\text{dice}}(\mathbf u_{c_k},\mathbf m_k)+\mathcal L_{\text{focal}}(\mathbf u_{c_k},\mathbf m_k)\Big]}_{\text{one mask per labelled site, conditioned on its anchor } c_k} \end{aligned} \]

The first three terms are dense, per-residue losses. The last is per-site, and it needs no matching because the anchor says which mask output belongs to which site. At inference the mask head is conditioned on the predicted centers instead. The Labels tab of the explorer above shows these targets on real proteins.

Are the auxiliary signals needed?

Objectness and votes never enter the decode, so it is fair to ask what they are for. We retrained without each.

Switch an auxiliary loss offInteractive
Centerness and per-site masksalways on: these define the decode
AP@0.5 change
AP@0.3 change

Predictions per protein:

Change in AP relative to the full model on the UniSite-DS test set. One training run per setting, same seed, same decode. The vote result shows the value of vote supervision as a training signal; it does not by itself show that the vote has to be equivariant.

The per-site mask term matters far more than either auxiliary. Without the mask head, sites have to be formed by assigning each binding residue to the nearest center through its vote, and AP@0.5 drops from 0.464 ± 0.007 to 0.201 ± 0.012 over three seeds.

Part 5 · Symmetry

What does equivariance buy?

Rotate or translate a protein and its binding sites move with it. Which residues form a site does not change. EquiSite has this built in. For a fixed neighbourhood graph, every scalar it produces (objectness, centerness, mask logits, and therefore the masks and their scores) is invariant to rigid motions; votes rotate with the structure, and predicted site centers move with it. This is Proposition 1 of the paper:

\[ f_\theta(g\cdot\mathcal G)=\big\{(s_k,\ \mathbf R\,\hat{\mathbf z}_k+\mathbf t,\ \hat{\mathbf m}_k)\big\}_{k=1}^{\widehat K}\qquad\text{for every rigid motion } g=(\mathbf R,\mathbf t)\in \mathrm{SE}(3) \]
Move the input: scores stay, votes turnInteractive
Pose Acoordinates as deposited
Pose B = R · Adrag to rotate the input
\(q_i\) along the chain · pose A
\(q_i\) along the chain · pose B
Top-ranked center · q: in A, in B · vote: in A, in B
of test proteins return exactly the same ranked list of sites after a random rotation and translation
largest change in any centerness score when the graph is unchanged
Ålargest gap between a vote and the rotated original vote
What the two panels show. The axes in the corner are fixed; it is the protein that turns. Colours (centerness), rings (the predicted centers) and the two strips are the same in both poses, while every vote arrow turns with the structure: its coordinates change, its meaning does not. The network is not re-run in your browser. The three numbers underneath are what we measured when we did re-run it, on every test protein.

We checked this on all test proteins, applying a random rotation and translation to each and re-running the whole model. return exactly the same ranked list of sites. The two that do not are instructive.

  • An edge at the cutoff. In proteins a pair of residues sits within about 10−4 Å of the 10 Å graph cutoff, and float32 round-off in the moved coordinates flips that edge. Centerness scores then shift by up to , and in one of the two proteins Suppress keeps a different mask. This is the "fixed neighbourhood graph" condition of the proposition at work.
  • A near-tie. In one protein the sites are identical but two almost equal scores swap rank. The cause is round-off in the float32 pairwise-distance routine; with exact distances the ranking is identical too.

A property, not the source of the gain

We do not claim that equivariance is what makes EquiSite more accurate. What the equivariant encoder contributes is a lightweight, pose-consistent backbone: about 1.1M parameters, where the encoder stack of UniSite-3D has about 26M.

Part 6 · Results

How well does it work?

EquiSite and UniSite-3D are trained on UniSite-DS with its official split at 90% maximum sequence identity (9,217 training and 2,293 test proteins), and take the same inputs: the structure and frozen ESM-2 embeddings.

UniSite-DS test set, residue-IoU average precision
UniSite-3D is re-evaluated from its released checkpoint; the other baselines are quoted from the UniSite paper. All use the same evaluation protocol.

Does it hold as the number of sites grows?

Test proteins grouped by number of labelled sites
UniSite-3DEquiSite+x% = relative gain
EquiSite is ahead in every group, and its relative margin is larger where there are more sites. Absolute accuracy falls steeply for both methods: with four or more sites per protein, neither does well.

Less similar test proteins

UniSite-DS ships stricter splits in which no test protein exceeds 70% or 50% sequence identity with the training set. Each model is retrained on each split.

AP@0.5AP@0.3
Max. train/test identityUniSite-3DEquiSiteUniSite-3DEquiSite
90%0.38350.46740.56030.6237
70%0.37340.45340.55790.6221
50%0.28010.38340.46770.5575
UniSite-3D values at 70% and 50% are quoted from the UniSite paper.

Predicted structures

In practice binding sites are often sought on predicted structures. For the 2,161 test proteins with an AlphaFold model, we ran the same checkpoint on AlphaFold coordinates in place of the experimental ones. AP@0.5 moves from 0.4764 to 0.4706 and AP@0.3 from 0.6324 to 0.6249, a change under 0.008.

Mean pLDDT of the modelProteinsCα RMSD to experimentAP@0.5, experimentalAP@0.5, AlphaFoldChange
94.5 or lower5412.62 Å0.45380.4310−0.0228
94.5 to 96.25401.35 Å0.46310.4570−0.0061
96.2 to 97.45400.78 Å0.48440.4877+0.0033
97.4 or higher5400.54 Å0.52650.5330+0.0064
Scope. These are high-confidence AlphaFold models (mean pLDDT 95.4, none below 70). The result is limited to structures of that quality: the loss is concentrated in the least confident quarter.

Other benchmarks

HOLO4K-sc and COACH420 are the standard single-chain benchmarks of the binding-site literature. For this comparison EquiSite is trained on the benchmark split of UniSite-DS and applied to the two sets without further training. DCC and DCA are the classical center-distance success rates.

HOLO4K-scCOACH420
MethodAP@0.3DCCDCAAP@0.3DCCDCA
Fpocket0.27110.30760.43820.21060.27080.4107
Fpocket-rescore0.58990.51830.76540.56020.44050.7113
P2Rank0.60110.53000.81880.61880.46430.7411
DeepPocket0.54150.49250.73690.51840.39580.6756
GrASP0.66680.51310.74160.71500.48510.7620
VN-EGNN0.26060.58610.69990.26370.54460.7530
UniSite-3D0.70910.57160.78790.71960.47020.7381
EquiSite0.77860.71760.86240.69280.55560.7544
Best value per column underlined. EquiSite leads on all three HOLO4K-sc metrics and on COACH420 DCC. On COACH420 it is behind UniSite-3D and GrASP on AP@0.3 and behind GrASP on DCA.

Size and speed

UniSite-3DEquiSiteRatio
Trainable parameters32.8M1.2M27×
Model forward pass, per protein49.3 ms9.2 ms5.4×
ESM-2 embedding, per protein (both models)29.8 ms29.8 ms1×
End to end, per protein79.1 ms39.0 ms2.0×
Batch size 1 on one NVIDIA H100. Both models consume the same frozen ESM-2 (650M) embeddings, which dominate the end-to-end time, so the honest speed-up for a user starting from a structure is 2×, not 5×. EquiSite trains in under an hour on the same GPU.

What the predictions look like

Molecular renderings comparing ground truth, EquiSite and UniSite-3D on a one-site protein (P84078) and a two-site protein (Q7YYQ9), with per-site IoU values under each panel.
Two test proteins. Left: labelled sites. Middle and right: the predictions matched to them, drawn over the labels, with their IoU.

Part 7 · Error analysis

Where does it still fail?

At the paper's settings EquiSite recovers 58% of the labelled test sites at IoU ≥ 0.5: 2,119 of 3,645. We traced each of the 1,526 misses to the first stage of the decode that lost it.

Which stage lost each missed site?
mask quality never nominated Select, other Select, close pair of sites Suppress

UniSite-DS test set, grouped by the number of labelled sites in the protein. A miss is assigned to the first stage that lost the site, in the order Suppress, Select, Nominate, mask quality.

Three quarters of the misses (1,146 of 1,526) are a learning problem rather than a decoding problem: the site was never nominated, or its mask was not good enough. The share that is never nominated grows sharply in proteins with four or more sites (271 of 625 misses), where sites are small and crowded. The two suppression steps cost less. Select accounts for 276 misses, of which 55 involve two labelled sites closer together than \(\rho\). Suppress accounts for 104, or 2.9% of all labelled sites.

Both suppression steps are a trade

Select and Suppress each remove candidates, and each occasionally removes a correct one. Loosening them on the whole test set shows what they buy.

Decode settingAP@0.5 changeAP@0.3 changeSites found at IoU ≥ 0.5Predictions per protein
Paper: ρ = 8 Å, τnms = 0.50058.1%4.47
ρ = 6 Å+0.006+0.00860.3%5.92
ρ = 4 Å+0.010+0.01162.0%7.64
Select off+0.013+0.01363.1%9.20
Suppress off−0.017−0.04961.0%5.85
Both off−0.060−0.12870.1%21.27
Same checkpoint, only the decode changes. Proteins in the test set have 1.59 labelled sites on average.

A smaller \(\rho\) finds more sites and nudges AP up, while doubling the number of predictions per protein and the number of mask-head evaluations. The paper's value of 8 Å takes the compact end of that trade. Suppress is the opposite case: without it AP falls, because near-duplicate masks crowd the ranking.

How strict should duplicate removal be?Test set
\(\tau_{\text{nms}}=0.5\) was chosen on a validation sweep. Between 0.5 and 0.8, AP@0.5 moves by less than 0.005 while AP@0.3 declines steadily. Suppress can still remove a correct mask when prediction errors make two different sites look alike; that is the 104 sites above.

Limitations

  • Many-site proteins remain hard. With four or more sites per protein, AP@0.5 is 0.121. Most misses there are sites that no residue nominated.
  • The language model is not free. Frozen ESM-2 embeddings are an input and dominate end-to-end inference time.

Reference

Cite

@inproceedings{animesh2026equisite,
  title     = {EquiSite: Equivariant Graph Neural Networks for End-to-End
               Prediction of Multiple Ligand Binding Sites},
  author    = {Animesh Animesh and Plaban Kumar Bhowmick and Pralay Mitra},
  booktitle = {The Fifth Learning on Graphs Conference},
  year      = {2026},
  url       = {https://openreview.net/forum?id=wBMHDA5q2S}
}

EquiSite builds directly on UniSite and its UniSite-DS benchmark, whose data, baselines and evaluation protocol we use throughout, and on DETR, CenterNet, HEGNN and ESM-2. Full references are in the paper.