Interaction Fingerprints Online: Comparing Hsp90 Binding Modes

Introduction

Interaction fingerprints turn a docked protein-ligand complex into a compact table of contacts: hydrogen bonds, hydrophobic contacts, van der Waals contacts, pi interactions, ionic contacts, halogen bonds, and related interaction classes. Instead of looking at one pose at a time, you can compare many complexes as fingerprints and ask whether ligands use the same pocket features, split into binding-mode families, or depend on different residues [1].

On Neurosnap, the Interaction Fingerprint service is useful after docking, structure prediction, or molecular dynamics. You provide protein-ligand complex structures, and the service returns CSV files that are easy to filter, cluster, plot, or feed into downstream machine-learning workflows.

This walkthrough uses a deliberately paired example: human Hsp90-alpha and Candida albicans Hsp90, both prepared from radicicol-bound crystal structures, with the same 30 DUDE-Z HSP90A actives docked into each pocket using GNINA [2,3]. The final Interaction Fingerprint jobs succeeded for 29 matched ligands in each receptor.

The Hsp90 workflow used here: two matched receptors, the same DUDE-Z active set, GNINA docking, matched protein-ligand complexes, Interaction Fingerprint jobs, and live CSV outputs for Plotly comparisons.

What Interaction Fingerprints Add

Docking scores are useful for ranking, but they compress a pose into one or a few numbers. Interaction fingerprints preserve the contact pattern behind the pose.

That makes them good for questions like:

• Do two ligands with similar scores make the same contacts?
• Which pocket residues appear repeatedly across a ligand set?
• Are apparent analogs actually using the same binding mode?
• Does the same ligand family contact different residues in two homologous proteins?
• Which docked poses are outliers that should be inspected before prioritization?

For a single complex, an interaction table is mostly annotation. For 20-50 complexes, it becomes a comparison layer.

Example: Human and Candida Hsp90

Hsp90 is a conserved molecular chaperone and a long-running drug-discovery target. The selectivity problem is the interesting part here: human Hsp90 inhibition can be toxic, while fungal Hsp90 remains attractive for antifungal strategies. Structural work on Candida albicans Hsp90 showed that species-selective inhibition is possible, partly because ligand-induced pocket rearrangements can differ between fungal and human Hsp90 nucleotide-binding domains [4].

For an apples-to-apples fingerprint example, we used two radicicol-bound ATP-pocket structures:

The docking boxes were centered on each crystallographic radicicol pose and set to 24 x 24 x 24 A. The ligand set came from the DUDE-Z HSP90A actives. We docked the first 30 actives into both receptors with GNINA, converted the top pose for each receptor-ligand pair into a protein-ligand complex, then submitted one Interaction Fingerprint job per receptor.

One ligand, CHEMBL453386, failed fingerprint generation in both receptors, so the matched comparison below uses the 29 successful complexes in each job.

Configure the Workflow

This example is a two-service workflow:

The completed shared Interaction Fingerprint jobs are:

Because these jobs are shared, the plots below fetch the result CSV files directly from the Neurosnap API instead of relying on a static asset copied into the blog.

Reading the Results

The Interaction Fingerprint service reports 17 interaction classes for these jobs, using explicit-hydrogen mode and a 6.0 A vicinity cutoff. The human Hsp90 job detected a mean of 10.72 interactions per ligand, while the C. albicans job detected a mean of 9.55 interactions per ligand. Those are not binding affinities; they are counts of detected ligand-residue contacts under the fingerprint rules.

The useful signal is the pattern. In this run, both receptors are dominated by van der Waals contacts, implicit hydrogen-bond donor contacts, and hydrophobic contacts. The most frequent human pocket residues include ASP93.A, ASN51.A, PHE138.A, and THR184.A. The aligned fungal pocket shows the analogous repeated-contact pattern around ASP82.A, ASN40.A, PHE127.A, and THR174.A.

Live Interaction Fingerprint Plots

The plots fetch live shared Interaction Fingerprint outputs from Neurosnap: contact counts, interaction classes, pairwise Jaccard similarity matrices, and recurrent pocket residues.

What to Look For

Start with the count plot only as a triage view. For example, CHEMBL404630 produced 21 detected interactions in human Hsp90 and 16 in C. albicans Hsp90, while CHEMBL365617 produced 18 in both receptors. That does not mean the first ligand is better or the second ligand is equipotent across species. It means those poses deserve a closer look because the fingerprint density is high in both pockets.

The type-frequency plot is a quality check. If one receptor suddenly showed mostly exotic interactions while the other showed mostly van der Waals and hydrophobic contacts, that would suggest a preparation, protonation, or pose problem. Here the broad interaction-type profile is consistent across receptors.

The similarity heatmaps are the most useful view. A high off-diagonal Jaccard value means two docked complexes share many of the same ligand-residue interaction bits. In this run, both receptors have sparse similarity matrices with a few compact clusters rather than one universal Hsp90 binding mode. The median off-diagonal similarity was 0.083 for human Hsp90 and 0.118 for C. albicans Hsp90, so most pairs do not share many exact fingerprint bits.

Finally, the residue-frequency plot shows which residues dominate the docked ensemble. This is the view to use before mutational interpretation or selectivity claims: frequent contacts can reflect real pocket chemistry, but they can also reflect receptor preparation, the selected receptor conformation, and the chemical bias of the ligand set.

Practical Tips

Use interaction fingerprints as a second-pass analysis, not as a replacement for docking quality control.

1. Start from chemically plausible poses. If a ligand is strained, outside the pocket, or missing hydrogens, the fingerprint can be precise but meaningless.
2. Keep receptor preparation consistent across comparisons. Here both receptors were prepared from radicicol-bound crystal structures, using chain A and the same box size.
3. Compare matched ligand sets when possible. The same 29 successful DUDE-Z actives were used for both receptors in the final plots.
4. Inspect outliers. Very low-count or very high-count poses can reveal failed docking, unusual chemistry, or a genuinely different binding mode.
5. Do not overinterpret a single receptor conformation. Hsp90 pockets are flexible, and the fungal selectivity literature points directly at ligand-induced rearrangements [4].

Limitations

This post is a workflow example, not a selectivity study. The ligands were known HSP90A actives, not a balanced fungal-vs-human selectivity panel. The receptors were single crystal conformations with ligands removed, and GNINA was used in a fixed-pocket docking setup. DUDE-Z is useful for benchmarking-style work, but benchmark actives and decoys are not the same thing as prospective medicinal chemistry data [3].

The right conclusion is narrower: Interaction Fingerprint gives a fast, inspectable way to compare contact patterns across a docked ligand set. For this Hsp90 example, it exposes repeated ATP-pocket residues, shows which ligands share contact patterns, and makes receptor-to-receptor differences easier to audit before making stronger claims.

Sources

[1] C. Bouysset and S. Fiorucci, "ProLIF: a library to encode molecular interactions as fingerprints," Journal of Cheminformatics, 2021.

[2] A. T. McNutt et al., "GNINA 1.0: molecular docking with deep learning," Journal of Cheminformatics, 2021.

[3] R. M. Stein et al., "Property-Unmatched Decoys in Docking Benchmarks," Journal of Chemical Information and Modeling, 2021.

[4] L. Whitesell et al., "Structural basis for species-selective targeting of Hsp90 in a pathogenic fungus," Nature Communications, 2019.

[5] RCSB PDB, "4EGK: Human Hsp90-alpha ATPase domain bound to Radicicol."

[6] RCSB PDB, "6CJL: Candida albicans Hsp90 nucleotide binding domain in complex with radicicol."