# Target fishing and reverse docking: finding what a compound binds

> Target fishing predicts which proteins a compound binds, either from its similarity to known ligands or by docking it into a panel of protein structures. In a 12-pocket test with AutoDock Vina, one pocket ranked first for 9 of 12 compounds. Learn how both methods work, why reverse docking scores need calibration, and how to check the hits.

## What is target fishing?

Target fishing is the computational prediction of which proteins a small molecule binds, working from the compound toward its targets. It reverses the usual docking question: instead of screening many compounds against one protein, you take one compound, often a hit from a cell assay with no known mechanism, and rank many proteins by how likely each is to bind it. Reverse docking, also called inverse docking, does the ranking with [AutoDock Vina](/app/autodock-vina) or a similar program. The other main route compares the compound with the known ligands of each target.

A reverse docking ranking looks decisive. Every protein gets a number in kcal/mol, the list is sorted, and the top reads like a prediction. We tested that by docking 12 compounds, each the crystal ligand of a different protein, into the pockets of all 12 proteins with Vina. A compound's own protein ranked first in only 4 of 12 cases. One pocket, the acetylcholinesterase gorge, ranked first for 9 of the 12 compounds, among them progesterone, erlotinib and oseltamivir carboxylate. An even spread would put each pocket first once.

Three questions decide whether a fishing run is worth the effort. Does the compound resemble anything with known targets? Is the real target in your panel, with a usable structure? Can the scores from different pockets be compared? The first two choose between similarity and docking. The third is the weak point of reverse docking, and most of this page is about it.

## Ligand-based target fishing

The older route needs no protein structure at all. Compounds that look alike tend to bind the same proteins, so a query compound is compared with the known ligands of every target in a database, and each target is scored by its closest ligands. Keiser and colleagues compared 3,665 approved and investigational drugs with hundreds of targets, each defined by its ligands, and tested 30 of the predicted pairs in the lab. Twenty-three were confirmed, and five of those were potent, below 100 nM.

SwissTargetPrediction is a widely used server built on the idea. It draws on 280,381 small molecules and 2,686 targets from ChEMBL and mixes 2D fingerprint similarity with [3D shape similarity](/use-cases/shape-based-virtual-screening). On 213 compounds that were not in its training data, at least one known target appeared among the top 15 predictions for 70% of them, and the single best-ranked target was correct for 31%.

The catch is that similarity needs neighbors. A compound with no relatives among known ligands has nothing to be matched against, and a protein with no known ligands has nothing to be scored on. The SwissTargetPrediction authors also note that it always returns some target, so for a compound with no real activity every prediction is a false positive. You can run the same comparison yourself with the [molecular similarity search](/app/molecular-similarity-search) against ligands pulled with [ChEMBL download](/app/chembl-download).

## Reverse docking, step by step

A reverse docking run has four parts: a panel of protein structures, a defined pocket in each, one docking program and a ranking of the proteins by the score of the best pose. Chen and Zhi described the first such program, INVDOCK, in 2001 and tried it on 4-hydroxytamoxifen and vitamin E. Half of the targets it proposed were already implicated or confirmed in the literature.

TarFisDock followed in 2006 with a panel of 698 structures. For vitamin E and 4-hydroxytamoxifen its top 2% contained 30% of the reported targets, and the top 5 to 10% contained half. Two test compounds is a thin base, and a major review still calls docking-based screening an enrichment tool, not an accurate predictor.

The panel sets the maximum claim. Approved drugs act on about 800 distinct targets, by [our count in ChEMBL 37](/guides/drug-targets), and only 35% of human proteins map to a Protein Data Bank entry, which covers 17% of all residues. A protein that is not in the panel, or sits there in the wrong conformation, cannot come back as a hit, and a failed docking is a gap in coverage, not a negative result. Predicted structures widen the panel at the price of less certain pockets, and the [structure prediction confidence guide](/guides/structure-prediction-confidence) covers how far to trust them.

The pocket is the second decision. Most published panels dock into a known site taken from a co-crystal ligand, which keeps the search small. [Blind docking](/use-cases/blind-docking) over the whole surface answers a different question and gives noisier scores. Before any of it runs, every structure goes through the same preparation, because differences in protonation, missing atoms or leftover ligands show up in the score as if they were chemistry. The [inverse virtual screening](/use-cases/inverse-virtual-screening) page lists the steps in order.

## Why the score of one pocket cannot be ranked against another

[Scoring functions](/guides/docking-scoring-functions) were built to rank many ligands in one pocket, and ranking one ligand across many pockets asks them for something else. Luo and colleagues measured the damage by docking 5,000 randomly chosen purchasable compounds into 100 proteins with Vina, DOCK and Glide. None of those compounds should favor any of the 100 proteins, so each protein should have come first about 50 times. With Vina, one protein came first for 1,267 compounds, and the median Vina score of a protein correlated with the contact area its pocket makes with ligands at r = −0.74.

The bias follows pocket size and, for some programs, hydrophobicity. Pan and colleagues showed the matching pull on the ligand side: energy scores favor larger compounds, because every extra atom in contact adds favorable terms. In target fishing the ligand is fixed, so its size cancels out of the ranking. The pocket's size does not, and a large pocket scores almost everything well.

Wang and colleagues, reviewed by Xu and colleagues, found "interprotein noise" in Glide scores and recovered part of it with a correction for each pocket's balance of hydrophobic and polar character, which raised target prediction accuracy from 57% to 72%. Schomburg and colleagues built a separate inverse screening method, iRAISE, around the same problem, with a score cascade and a score cutoff for each pocket.

Benchmarks agree that this is not one program's quirk. Lapillo and colleagues ran 13 docking procedures on X-ray structures with 10 active ligands per target and found that features of the target proteins affected how reliably the correct one was picked. Li and colleagues built a benchmark with similar and dissimilar decoys and saw performance biased by binding site properties in AutoDock, Vina, Glide and GOLD, with Glide the best of the four.

The usual fix is to normalize each pocket by how the same program scores it for other compounds. Luo and colleagues subtracted each pocket's median score and divided by its standard deviation across all ligands. It flattened the bias, but it helped Glide far more than Vina.

On the Astex Diverse Set, restricted to the complexes each program could redock within 2 Å, the top-ranked protein was correct for 38.5% of Glide cases before and 60.0% after, and for 33.3% of Vina cases before and 40.0% after. Normalization removes the pocket's pull, and whatever the scoring function cannot tell apart stays.

## A 12-pocket test of reverse docking

We built the smallest panel that still contains the problem. Table 1 lists 12 co-crystal structures from the Protein Data Bank that span kinases, nuclear receptors, a bromodomain, two hydrolases and a chaperone, each with a drug-like ligand. Every ligand was docked into every pocket, 144 dockings in all, with AutoDock Vina 1.2.7 at exhaustiveness 8 and seed 42. The search box was a 22 Å cube centered on each structure's crystal ligand, so the search volume is identical everywhere.

  <caption><strong>Table 1. The 12 targets and the compound docked from each.</strong> Each compound is the ligand in its own crystal structure. Heavy atoms are counted in the ligand model used for docking. Source: RCSB Protein Data Bank entries listed.</caption>
  <thead>
    <tr>
      <th scope="col">Target</th>
      <th scope="col">PDB entry</th>
      <th scope="col">Compound docked</th>
      <th scope="col">Heavy atoms</th>
    </tr>
  </thead>
  <tbody>
    <tr><td>ABL1 kinase</td><td>1IEP</td><td>Imatinib</td><td>37</td></tr>
    <tr><td>EGFR kinase</td><td>1M17</td><td>Erlotinib</td><td>29</td></tr>
    <tr><td>p38α kinase</td><td>1A9U</td><td>SB203580</td><td>27</td></tr>
    <tr><td>BRD4 first bromodomain</td><td>3MXF</td><td>JQ1</td><td>31</td></tr>
    <tr><td>Estrogen receptor α (ERα)</td><td>3ERT</td><td>4-Hydroxytamoxifen</td><td>29</td></tr>
    <tr><td>Progesterone receptor (PR)</td><td>1A28</td><td>Progesterone</td><td>23</td></tr>
    <tr><td>Glucocorticoid receptor (GR)</td><td>1M2Z</td><td>Dexamethasone</td><td>28</td></tr>
    <tr><td>Androgen receptor (AR)</td><td>1T65</td><td>Dihydrotestosterone</td><td>21</td></tr>
    <tr><td>PPARγ</td><td>2PRG</td><td>Rosiglitazone</td><td>25</td></tr>
    <tr><td>Acetylcholinesterase (AChE)</td><td>1EVE</td><td>Donepezil</td><td>28</td></tr>
    <tr><td>Neuraminidase N1 (NA)</td><td>2HU4</td><td>Oseltamivir carboxylate</td><td>20</td></tr>
    <tr><td>HSP90α</td><td>1UYG</td><td>Purine PU2</td><td>22</td></tr>
  </tbody>

Each receptor is the protein chain or chains around the ligand, with waters, ions and cofactors removed, missing side-chain atoms added with PDBFixer and the file prepared with Meeko. Each compound starts from the wwPDB ideal coordinates, never from its crystal pose. A compound's rank for its own protein is its place among the 12 pockets by best-pose score. We began with 16 structures and dropped four whose flexible ligands docked too slowly for our time budget, so the panel is a convenience set and not a random draw.

A [redocking](/guides/redocking-docking-validation) check shows whether the setup works at all. The top pose of each compound in its own pocket fell within 2 Å of the crystal pose for 6 of the 11 we could measure: imatinib 0.76 Å, erlotinib 1.46, SB203580 0.60, 4-hydroxytamoxifen 1.20, progesterone 0.84 and dihydrotestosterone 0.63.

It missed for the purine (3.1 Å), oseltamivir carboxylate (4.7), JQ1 (5.0), rosiglitazone (6.9) and donepezil (8.9 Å). Dexamethasone's crystal ligand would not parse against its template, so that case is not measured. The scores, ranks and scripts are in the [scores table](/data/guides/target-fishing/reverse-docking-scores.csv) and the [data folder](/data/guides/target-fishing/README.md).

## What the 144 scores show

The pocket mattered more than the compound. Pocket averages ran from −10.06 kcal/mol for AChE to −5.13 for AR, a spread of 4.9 kcal/mol that Vina's scale would read as a factor of about 4,100 in Kd. Which pocket a compound was docked into explained 19% of the variance in all 144 scores, and which compound explained 10%. The AChE pocket scored best for 9 of the 12 compounds, including donepezil, whose pose in it sat 8.9 Å from the crystal pose, so the first place there did not mean the right pose.

![Horizontal bars showing for each of 12 pockets how many of the 12 compounds it ranked first for under raw Vina scores and after calibration; the acetylcholinesterase pocket ranked first for 9 raw and 1 calibrated](/images/charts/reverse-docking-pocket-first-counts.webp "**Figure 1. One pocket wins most raw rankings.** For each pocket, the number of the 12 compounds for which it had the best score, raw and after calibration by the pocket's scores for the other 11 compounds. An even spread is 1 each. Source: ProteinIQ 12 x 12 AutoDock Vina test, November 2026.")

We did not measure why. Luo and colleagues tied such interference proteins to pockets that make a large contact area with ligands, and the AChE active site is a deep, narrow gorge, which fits that pattern without proving it. The steroid receptors show the opposite problem. Imatinib does not fit them and scores +8.6, +5.2 and +1.7 kcal/mol in PR, AR and GR, so those pockets have a score spread of 3.7 to 5.7 kcal/mol, against 0.8 to 1.5 for the other nine. A pocket's whole score distribution depends on which compounds were docked into it.

## Calibrating each pocket

4-Hydroxytamoxifen is the clearest case, and it is the compound that Chen and Zhi and the TarFisDock authors used to test their programs. Raw Vina puts its own receptor second, at −9.62 kcal/mol, behind AChE at −10.03, a gap well inside docking error. Table 2 shows what a pocket-by-pocket z-score does. For each pocket we took the mean and standard deviation of the scores that the other 11 compounds got there, and asked how far this compound's score sits from them. The estrogen receptor moves to first place, and AChE falls to fifth.

  <caption><strong>Table 2. 4-Hydroxytamoxifen across six of the 12 pockets, raw and calibrated.</strong> The z-score is the compound's score minus the mean of the other 11 compounds in that pocket, divided by their standard deviation; lower is better. We calculated all values from our 144 Vina scores.</caption>
  <thead>
    <tr>
      <th scope="col">Pocket</th>
      <th scope="col">Raw score (kcal/mol)</th>
      <th scope="col">Raw rank</th>
      <th scope="col">z-score</th>
      <th scope="col">Calibrated rank</th>
    </tr>
  </thead>
  <tbody>
    <tr><td>AChE</td><td>−10.03</td><td>1</td><td>+0.03</td><td>5</td></tr>
    <tr><td>ERα (true target)</td><td>−9.62</td><td>2</td><td>−1.21</td><td>1</td></tr>
    <tr><td>PPARγ</td><td>−8.52</td><td>3</td><td>−0.09</td><td>4</td></tr>
    <tr><td>BRD4</td><td>−8.33</td><td>4</td><td>−0.53</td><td>2</td></tr>
    <tr><td>PR</td><td>−6.96</td><td>10</td><td>−0.21</td><td>3</td></tr>
    <tr><td>AR</td><td>−3.95</td><td>12</td><td>+0.24</td><td>7</td></tr>
  </tbody>

Across all 12 compounds the correction helped and fell short. The own protein ranked first in 6 of 12 cases instead of 4, in the top 3 in 9 instead of 7, and its mean rank improved from 3.08 to 2.67, against 6.5 for a random ordering. But the first-place pocket was still wrong for 6 compounds, and the correction made some cases worse: erlotinib's rank went from 7 to 9 and progesterone's from 2 to 4.

![Horizontal bars of the rank of each compound's own protein among 12 pockets, raw and calibrated, from imatinib and ABL1 at rank 1 to erlotinib and EGFR at ranks 7 and 9](/images/charts/reverse-docking-true-target-rank.webp "**Figure 2. Where each compound's own protein ranked.** Rank among the 12 pockets by best-pose Vina score (1 is best), raw and after calibration. A random ordering averages 6.5. Source: ProteinIQ 12 x 12 AutoDock Vina test, November 2026.")

Even the choice of correction moves the answer. Four other reasonable versions (median and MAD, clipping clashes at zero, centering without scaling, and using all 12 compounds as the reference) put the own protein first in 5 to 7 of 12 cases, with mean ranks of 2.50 to 2.75. With 12 compounds, one case is 8 percentage points, so those versions cannot be told apart. The reference set matters as much as the formula, because a pocket's reference is only as good as the compounds docked into it.

This test flatters reverse docking in several ways. The panel has 12 pockets, so a wrong pocket has only 11 chances to intrude, and a panel of 800 gives it far more. Every box was placed on the known site from a crystal structure that held a ligand, which is the easiest version of the problem. We used one program, one seed and one preparation. The numbers show the mechanism, and they are not an accuracy estimate. Luo and colleagues' larger test, with 33.3% first-place accuracy for Vina on 45 Astex complexes, is the better guide to what to expect.

## Choosing a method for your compound

The compound in hand decides the method more than any benchmark does. Table 3 matches the usual starting points to what each can and cannot see. [Ligand-based](/use-cases/ligand-based-virtual-screening) and structure-based fishing fail in different places, so a target that both propose is less likely to be an artifact of either.

  <caption><strong>Table 3. Where to start, by what you know about the compound.</strong> The last column names what the starting method cannot see.</caption>
  <thead>
    <tr>
      <th scope="col">What you have</th>
      <th scope="col">Start with</th>
      <th scope="col">Add</th>
      <th scope="col">Blind spot</th>
    </tr>
  </thead>
  <tbody>
    <tr>
      <td>A close analog of a known drug or tool compound</td>
      <td>Similarity to known ligands</td>
      <td>Docking into the top few targets, for a pose</td>
      <td>Targets with no known ligands</td>
    </tr>
    <tr>
      <td>A new scaffold and a short list of suspected targets with structures</td>
      <td>Reverse docking into a curated, focused panel</td>
      <td>Calibration with known ligands for each pocket</td>
      <td>Any protein outside the panel</td>
    </tr>
    <tr>
      <td>A phenotypic hit and no target idea</td>
      <td>Both methods, over the widest panel you can prepare</td>
      <td>Keep the targets that appear in both lists</td>
      <td>Targets without a structure or a known ligand</td>
    </tr>
    <tr>
      <td>A lead whose off-targets worry you</td>
      <td>Docking into a panel of known counter-targets</td>
      <td>Pose and contact checks on every hit</td>
      <td>Binding to proteins nobody has listed as a liability</td>
    </tr>
  </tbody>

## What to do with the ranked list

Treat the first few ranks as a shortlist and nothing finer. A [docking score is not a binding affinity](/guides/docking-score-binding-energy-affinity), and Vina's typical error of 2.4 kcal/mol is a factor of 54 to 57 in Kd, so neighboring ranks cannot be told apart. Check each pose with [PoseBusters](/app/posebusters), which tests that a pose is physically plausible, and compare its contacts with those of the ligand in that protein's crystal structure using [PLIP](/app/plip) or [ProLIF](/app/prolif). A hit whose pose ignores the residues that the known ligand uses deserves doubt. A clean check says nothing about whether the protein is a real target.

Do not count a hit from the same family as the true target as a failure before you look at it. Known drugs average about six targets, and close relatives of a protein share most of its pocket. A relative that ranks first may be a real off-target, or it may be the pocket that scores everything well. Only an assay tells the two apart.

## Target fishing tools on ProteinIQ

A run goes in this order: gather what is known about the compound, build and prepare the panel, dock, then check every hit. The workflow template on the inverse virtual screening page strings the structure-based steps together over a panel with PDBFixer, GNINA, PoseBusters and ProLIF.

  <caption><strong>Table 4. Tools for ligand-based and structure-based target fishing.</strong></caption>
  <thead>
    <tr>
      <th scope="col">Task</th>
      <th scope="col">Tool</th>
    </tr>
  </thead>
  <tbody>
    <tr>
      <td>Pull the known ligands of a target</td>
      <td>[ChEMBL download](/app/chembl-download), [PubChem download](/app/pubchem-download)</td>
    </tr>
    <tr>
      <td>Compare a compound with known ligands</td>
      <td>[Molecular similarity search](/app/molecular-similarity-search)</td>
    </tr>
    <tr>
      <td>Fetch the structures that make up a panel</td>
      <td>[PDB download](/app/pdb-download), [AlphaFold database download](/app/alphafold-database-download), [UniProt download](/app/uniprot-download)</td>
    </tr>
    <tr>
      <td>Repair every structure the same way</td>
      <td>[PDBFixer](/app/pdb-fixer)</td>
    </tr>
    <tr>
      <td>Find a pocket when no ligand is bound</td>
      <td>[fpocket](/app/fpocket)</td>
    </tr>
    <tr>
      <td>Dock one compound into each pocket</td>
      <td>[AutoDock Vina](/app/autodock-vina), [smina](/app/smina), [GNINA](/app/gnina)</td>
    </tr>
    <tr>
      <td>Dock without defining a pocket</td>
      <td>[DiffDock](/app/diff-dock)</td>
    </tr>
    <tr>
      <td>Check that a pose is physically plausible</td>
      <td>[PoseBusters](/app/posebusters)</td>
    </tr>
    <tr>
      <td>Compare a pose's contacts with the crystal ligand's</td>
      <td>[PLIP](/app/plip), [ProLIF](/app/prolif)</td>
    </tr>
    <tr>
      <td>Judge whether the compound could reach the target</td>
      <td>[ADMET-AI](/app/admet-ai)</td>
    </tr>
  </tbody>

## Frequently asked questions

### What is target fishing?

Target fishing is the computational prediction of the proteins a small molecule binds, starting from the compound. It works either by comparing the compound with the known ligands of each target or by docking it into a panel of protein structures and ranking the proteins.

### Is reverse docking the same as inverse docking?

Yes, the two names describe the same method: one compound docked into many proteins, which are then ranked. Inverse virtual screening is the wider term, because it also covers approaches that use ligand similarity or pharmacophores instead of docking.

### How accurate is reverse docking?

Modestly. Luo and colleagues found the correct protein ranked first for 33.3% of cases with Vina on 45 Astex complexes, and in the top 5 for 68.9%. Their score normalization raised those figures to 40.0% and 73.3%. In our 12-pocket test the own protein ranked first for 4 of 12 compounds raw and 6 of 12 calibrated.

### How many proteins does a target fishing panel need?

It depends on the claim. A focused safety or selectivity panel can be a few dozen proteins. A search for the target of an unexplained phenotype needs far more: approved drugs alone act on about 800 targets, and TarFisDock's panel held 698 structures.

### Can I do target fishing for free?

Yes. The ligand-based SwissTargetPrediction server is free and needs no login. A structure-based panel can be run on ProteinIQ with the tools above, and the main cost is preparing and curating the structures.

### Should I use similarity or docking?

Use similarity when the compound resembles known ligands, since it needs no structure and SwissTargetPrediction was tested on 213 new compounds against 2,686 targets. Use docking when the compound is new or its target has few known ligands. The two methods have not been benchmarked on the same compounds here, so their accuracy figures cannot be compared directly.
