ProLIF interaction fingerprints — what your poses are actually doing

A docking score is a single number trying to summarize a three-dimensional binding event. It throws away almost everything that medicinal chemists actually care about: which residue made the key hydrogen bond, whether the critical π-stack is there, whether the basic amine is sitting where the conserved aspartate is. Interaction fingerprints fix that. They keep the per-residue contact information as a sparse bitvector, which you can compare, cluster, and feed into downstream models. ProLIF is the open-source tool that turned this into a few lines of Python, and it has become the standard way to do pose-level analysis on top of any docking program.

What an interaction fingerprint actually is

An interaction fingerprint (IFP) is a binary vector where each bit answers a yes/no question about a specific interaction. The simplest scheme assigns one bit per (residue, interaction-type) pair, where the interaction types are the canonical set medicinal chemists already think in: hydrogen-bond donor and acceptor, hydrophobic contact, π-stacking, π-cation, halogen bond, salt bridge, and metal coordination. A pose with 30 contacts to 12 active-site residues collapses to a fingerprint a few thousand bits long, where most bits are zero.

That representation is useful for three reasons. First, two poses that score similarly but engage different residues are easy to tell apart — their Tanimoto similarity in fingerprint space will be low even when the energies are equal. Second, a reference fingerprint (from a crystal structure or a validated pose) becomes a hard target you can rank screening hits against. Third, the fingerprint is a clean featurization for any downstream ML model — pose-quality classifier, affinity predictor, selectivity model.

Where ProLIF fits

ProLIF, published by Bouysset and Fiorucci in 2021, is a Python library that consumes a docked pose or an MD trajectory and emits the fingerprint. The implementation works on top of RDKit and MDAnalysis, which means it can read essentially every file format the field uses: PDB, SDF, MOL2, DCD, XTC, parquet pose stacks. The default interaction set is the standard eight types (hydrophobic, π-stacking, π-cation, cation-π, anionic, cationic, H-bond donor, H-bond acceptor). Geometric criteria are parameterized — angle and distance cutoffs per type — and can be reparameterized for an unusual case like a halogen-rich screening set or a metal cofactor.

For a docking output, the typical workflow is three function calls: load the protein and the pose set, instantiate a Fingerprint object, callrun(). The result is a pandas DataFrame indexed by pose ID with a multi-index column per (residue, interaction-type) pair. From there it is the usual data-science routine: similarity to a reference pose with jaccard distance, agglomerative clustering with scipy, or feature feed into sklearn for a structure-activity model. ProLIF also ships a network-style HTML viewer for human inspection, which is useful when you are trying to communicate “these 200 hits all engage the same hinge residue” to a med chem audience.

The four jobs an IFP is good at

1. Cherry-picking the right pose. A Vina-class scoring function will frequently rank a pose in the top three that misses the key catalytic contact. If you know the key contact from the co-crystal — say, the H-bond to the hinge methionine in a kinase — you can filter your top-N to only poses where that bit is set. This is the difference between “the top score” and “the top pose that is doing what we know it has to do.” In a published benchmark by Bietz et al. on the PDBbind test set, IFP-based pose selection recovered the near-native pose more often than energy ranking alone.

2. Diversity assessment in virtual screening. Two hits that share a high IFP-Tanimoto are probably binding the same way and exploring the same SAR. Two hits with a low IFP-Tanimoto but similar scores are engaging different parts of the pocket and represent genuinely independent chemical series. A library triage that picks the top-100 by score will tend to cluster — IFP-clustering across the top 1000 and picking centroids gives a more chemically diverse shortlist.

3. Selectivity rationalization. When a compound shows differential potency between WT and a mutant, the IFP tells you why. The bits that flip between WT and mutant poses point at the residues that are doing different work. This is how you turn a retrospective ΔΔ into a forward design hypothesis.

4. ML featurization. An IFP is a compact, interpretable feature set. Models trained on IFPs are smaller and more transparent than the equivalent voxel-grid or graph-neural-network model, and the feature importances point directly at residue-level chemistry. Several recent papers use ProLIF outputs as the input layer for pose-quality classifiers and for selectivity models in the kinase space.

The limits

An IFP is only as good as the geometry it is reading. If your pose is wrong, your fingerprint is wrong. So IFP-based workflows should sit downstream of a pose validator — PoseBusters for geometric sanity, re-scoring with GNINA or a CNN for the energy side — not replace one. Second, IFPs do not capture energetics: a strong hydrogen bond and a weak one flip the same bit. If your downstream question is affinity, you want the bit pattern plus a force-field readout, not the bit pattern alone. Third, the default interaction geometry cutoffs were tuned on small-molecule protein complexes; for nucleic-acid targets or macrocycles you probably want to re-parameterize.

Where this lands in the Liganx workflow

The pose viewer in Liganx already highlights the interaction set ProLIF tracks — same eight types, same geometric criteria. The result is that when you scroll through the top poses on a docking run, the colored contact arcs are functionally the same information ProLIF would emit as a fingerprint row. That is the right granularity for the question medicinal chemists actually ask about a docking result: not “what is the score?” but “is the molecule engaging the residues we know matter?”

For a pure scoring workflow with no interaction analysis, you are over-fitting to a single number. For a workflow where the score is supplemented by the IFP — either from ProLIF in a downstream notebook or from the Liganx pose viewer at decision time — the failure mode of “great score, wrong pose” is the one that gets caught.

Try the interaction view yourself

Open Studio and dock any candidate against a target. After the run completes, click any pose in the results table to open the 3D viewer. The interaction sidebar lists each contact by residue and type, color-coded the same way ProLIF colors its plots. Compare two poses by switching between them — the bits that change tell you what the molecule is doing differently, exactly the way an IFP Tanimoto would.

Liganx puts molecular docking online and free in the browser, with the interaction view built in. Using molecular docking with the contact set surfaced at decision time is the cheapest way to avoid the “great score, wrong pose” failure mode.

Primary sources

• Bouysset C, Fiorucci S. ProLIF: a library to encode molecular interactions as fingerprints. J Cheminform 13, 72 (2021). doi:10.1186/s13321-021-00548-6
• Deng Z, Chuaqui C, Singh J. Structural interaction fingerprint (SIFt): a novel method for analyzing three-dimensional protein-ligand binding interactions. J Med Chem 47, 337-344 (2004). doi:10.1021/jm030331x
• Da C, Kireev D. Structural protein-ligand interaction fingerprints (SPLIF) for structure-based virtual screening: method and benchmark study. J Chem Inf Model 54, 2555-2561 (2014). doi:10.1021/ci500319f
• ProLIF GitHub repository, chemosim-lab/ProLIF. github.com/chemosim-lab/ProLIF