PANIP: Protein Interaction Energy Predictor

• PANIP is a specialized machine learning potential that directly predicts pairwise non-covalent interaction energies between protein fragment dimers using an ensemble of NequIP models.
• It leverages a multi-fidelity active learning pipeline on PDB-derived fragment dimers to achieve near-DFT accuracy with mean absolute errors below 0.2 kcal/mol.
• PANIP supports fragment-based protein-ligand scoring and global NCI analysis, outperforming models like ANI-2x in handling protein-specific charged and geometry-sensitive motifs.

to=arxiv_search.search 彩神争霸大发json code these words: "PANIP pairwise non-covalent interaction potential protein fragments NequIP", "page": 1, "max_results": 5} to=arxiv_search.search qq天天中彩票json code {"query":"PANIP pairwise non-covalent interaction potential protein fragments NequIP","page":1,"max_results":5} to=arxiv.search 天天中彩票网站json code {"query":"PANIP pairwise non-covalent interaction potential protein fragments NequIP","max_results":5} PANIP, short for PAirwise Non-covalent Interaction Potential, is a protein-specific machine-learning interatomic potential designed to predict pairwise non-covalent interaction energies between biomolecular fragments. It is implemented as an ensemble built on the NequIP framework and trained on protein-derived fragment dimers extracted from the Protein Data Bank, with the explicit goal of modeling hydrogen bonds, electrostatics, dispersion, cation–$\pi$, sulfur–aromatic, ionic, backbone-sidechain, and protein-water contacts at near-DFT accuracy and near force-field cost. In its published form, PANIP is not presented as a universal molecular potential; it is a specialized short-range interaction model whose principal uses are fragment-fragment energy prediction, large-scale analysis of interaction motifs across the PDB, and fragmentation-based protein-ligand scoring (Zeng et al., 14 Jan 2026).

1. Definition, scope, and scientific motivation

PANIP was developed to address a mismatch between general-purpose ML interatomic potentials and the interaction motifs that dominate proteins. The underlying argument is that broad-coverage MLIPs often undersample the geometry-sensitive and biologically consequential NCIs that are central to protein structure and recognition, especially rare but important motifs involving charged fragments, cation–$\pi$ contacts, sulfur-containing interactions, and protein-water interactions. PANIP therefore adopts a deliberately narrow scope: it predicts the interaction energy of a dimer rather than the total energy of an arbitrary biomolecular system, and it is trained directly on protein-derived fragment dimers rather than on generic small-molecule data (Zeng et al., 14 Jan 2026).

A defining distinction from models such as ANI-2x is the training target. PANIP predicts the interaction energy directly,

$\Delta E_{AB} = E(AB) - [E(A) + E(B)],$

instead of inferring it by subtracting separately predicted monomer and dimer total energies. This design is intended to avoid compounded subtraction error and is especially important for charged systems, which the PANIP study identifies as a major weakness of ANI-2x. The model is also specialized to fragments and motifs common in proteins, rather than attempting to span arbitrary organic chemistry (Zeng et al., 14 Jan 2026).

The reported quantitative claim in the abstract is that PANIP retains $\omega$B97X-D3BJ/def2-TZVPP-level accuracy and achieves mean absolute errors below 0.2 kcal/mol on out-of-distribution systems. A plausible implication is that PANIP occupies a middle ground between fragment quantum chemistry and docking-force-field heuristics: it is narrow in domain, but optimized for a scientifically important one (Zeng et al., 14 Jan 2026).

2. Dataset construction and the PDB-FRAGID workflow

The PANIP dataset pipeline begins with 29,204 proteins selected from the PDB using resolution $\le 2.5$ Å and unique UniProt IDs. Missing hydrogens were added using HTMD, original PDB coordinates were preserved, and no further structural optimization was performed. Protonation states were assigned according to typical physiological conditions: histidine in all three common protonation states, lysine and arginine protonated, and aspartate deprotonated (Zeng et al., 14 Jan 2026).

Fragments were extracted by in-house scripting using SMARTS patterns and RDKit substructure matching, with coordinate mapping to fragment templates. When fragment isolation required cleavage, a C–C or C–N bond was cut, the terminal atom was marked as a dummy, and the fragment was capped with hydrogen. The resulting chemical coverage comprised 17 chemically distinct fragment types: Water (HOH), Ethylammonium (ETAM), Ethanol (ETOH), Ethanethiol (ETSH), Propane (PRPA), Dimethyl sulfide (MSM), o-Methylimidazole (MIMD), $\varepsilon$-Methylimidazole (MIME), Methylimidazolium (MIMM), Methylbenzene (MBZ), p-Methylphenol (PMPO), Methylindole (MIND), Acetamide (ACEM), N-methylacetamide (NMA), N-(1-pyrrolidinyl)acetamide (N1PA), Acetate (ACET), and Methylguanidinium (MGDM) (Zeng et al., 14 Jan 2026).

Candidate dimers were retained when the shortest heavy-atom distance was between 2 and 4 Å and the two fragments were separated by at least two residues in sequence. These filters produced 36.3 million raw dimers spanning 153 unique dimer combinations. Because high-level labeling of the full pool was computationally prohibitive, the authors introduced an automated multi-fidelity active learning (MFAL) pipeline. Low-fidelity r2SCAN-3c interaction energies were computed for all 36.3 million dimers, while high-fidelity labels were reserved for a distilled subset called PDB-FRAGID (Zeng et al., 14 Jan 2026).

The MFAL loop started from a 2% random bootstrap set. For unscreened dimers, the acquisition rule flagged cases satisfying

$|E_{\mathrm{pred} - E_{\mathrm{ref}}|/\sqrt{N} > 0.04 \ \text{kcal/mol},$

with $E_{\mathrm{ref}}$ taken from r2SCAN-3c and $N$ the atom count. At each iteration, 2% of the flagged high-error cases were added, and the loop terminated when fewer than 5% of the unscreened dimers remained high-error, after which all remaining flagged cases were included. This reduced the candidate pool to approximately 3.15 million dimers, or 8.7% of the original set, yielding PDB-FRAGID (Zeng et al., 14 Jan 2026).

The low-fidelity/high-fidelity correlation itself was reported as MAE = 0.272 kcal/mol, RMSE = 0.423 kcal/mol, and $R^2 = 0.999$. This suggests that r2SCAN-3c was sufficiently correlated with the target level of theory to support an error-based screening strategy, while still leaving room for systematic correction at high fidelity (Zeng et al., 14 Jan 2026).

3. Model architecture, reference labeling, and training formulation

PANIP is implemented as a 5-model ensemble of NequIP models trained by five-fold cross-validation, with predictions reported as ensemble averages unless otherwise noted. NequIP is characterized in the paper as an E(3)-equivariant graph neural network that explicitly preserves rotational, translational, and permutational symmetries. In PANIP, this equivariant architecture is used to output a scalar dimer interaction energy, with a cutoff radius of 7.5 Å (Zeng et al., 14 Jan 2026).

The reported implementation settings include batch size = 512, early stopping patience = 100 epochs on validation loss, and global energy rescaling = dataset_total_energy_std. The authors also note a modified chemical-symbol configuration, specifically the inclusion of O, N, and S in the chemical symbols list. Before the final unified model was trained, MFAL-selected subsets from each fragment type were used to train fragment-specific NequIP models; these were then merged into the unified PDB-FRAGID training set used for the final PANIP ensemble (Zeng et al., 14 Jan 2026).

The high-fidelity target level was $\pi$0B97X-D3BJ/def2-TZVPP, computed in ORCA 5.0.3 with RIJCOSX, auxiliary basis sets def2/J and def2-TZVPP/C, and counterpoise BSSE correction. The study further states that ma-def2-TZVPP diffuse augmentation was added only for negatively charged oxygen atoms, such as acetate oxygens. The rationale given is that diffuse functions improved anionic oxygen treatment, whereas neutral systems did not benefit and could suffer numerical instability (Zeng et al., 14 Jan 2026).

The manuscript does not explicitly provide the supervised loss formula used for NequIP training. It does, however, make clear that PANIP is trained to predict direct pairwise NCI energies, not forces and not total molecular energies. This matters for interpretation: PANIP is an energy regressor specialized to fragment dimers, not a general reactive potential or a full biomolecular force field (Zeng et al., 14 Jan 2026).

4. Accuracy, transferability, and comparison with ANI-2x

PANIP was evaluated on four principal benchmarks: a low-energy PDB-derived set, an optimized low-energy set, a CSD-derived out-of-distribution structural set, and a random-generated non-equilibrium set. The reported results are summarized below.

Benchmark	Size	Performance
Low-energy PDB-derived dimers	277	MAE 0.09, RMSE 0.163, $\pi$1 0.999
Optimized low-energy dimers	277	MAE 0.547, RMSE 1.207, $\pi$2 0.995
CSD-derived benchmark	33,274	MAE 0.171, RMSE 0.507, $\pi$3 0.999
Random-generated benchmark	15,300	MAE 0.448, RMSE 1.372, $\pi$4 0.996
Random-generated, $\pi$5	8,572	MAE 0.195, RMSE 0.363, $\pi$6 0.999

These figures show a characteristic pattern: PANIP is strongest on equilibrium and near-equilibrium attractive dimers, remains strong on structurally OOD CSD dimers, and degrades on non-equilibrium repulsive or clashing geometries. The paper attributes underestimation of very high positive interaction energies to their scarcity in training and to fragment-class sparsity for some systems such as ETSH dimers (Zeng et al., 14 Jan 2026).

The comparison with ANI-2x is particularly stark on the CSD benchmark. PANIP achieved MAE 0.171, RMSE 0.507, $\pi$7, whereas ANI-2x yielded MAE 9.261, RMSE 23.765, $\pi$8 on the full set. Even after excluding charged dimers, ANI-2x remained substantially worse, at MAE 0.882, RMSE 1.25, $\pi$9. The interpretation given in the paper is that PANIP’s advantage arises from four design choices: direct prediction of interaction energies, explicit charged-fragment coverage, protein-specific geometries, and high-fidelity BSSE-corrected reference labels (Zeng et al., 14 Jan 2026).

The random-generated repulsive subset contains an explicit reporting inconsistency. The main text gives RMSE 0.882 kcal/mol for the repulsive subset, while the SI figure gives MAE 0.882, RMSE 3.137, $\Delta E_{AB} = E(AB) - [E(A) + E(B)],$0. The paper summary notes this discrepancy without resolving it. For technical use, this suggests caution when interpreting performance on the most strongly repulsive configurations (Zeng et al., 14 Jan 2026).

The speed comparison is also central. For 15,300 random-generated samples on a single CPU core, $\Delta E_{AB} = E(AB) - [E(A) + E(B)],$1B97X-D3BJ/def2-TZVPP required 463 days, 11 h, 19 min, 6.1 s, whereas PANIP required 6 h, 17 min, 55.6 s. The manuscript describes this as more than two orders of magnitude faster than the target DFT method (Zeng et al., 14 Jan 2026).

5. Fragment-based binding scoring and global analysis of protein NCIs

PANIP was integrated into a fragmentation-based protein-ligand scoring framework through the binding-energy decomposition

$\Delta E_{AB} = E(AB) - [E(A) + E(B)],$2

Here, $\Delta E_{AB} = E(AB) - [E(A) + E(B)],$3 is the PANIP-predicted interaction energy for a protein fragment and a ligand fragment, while $\Delta E_{AB} = E(AB) - [E(A) + E(B)],$4 and $\Delta E_{AB} = E(AB) - [E(A) + E(B)],$5 correct for linking atoms and double-counting. In the reported docking workflow, receptor and ligand were fragmented, only fragments within 5 Å around the ligand were retained, poses were generated by DOCK 3.7, and PANIP was used for rescoring (Zeng et al., 14 Jan 2026).

Three case studies are highlighted. For L99A–indole (PDB 185L), PANIP ranked the native pose 1st among 500, whereas DOCK ranked the native pose 160th and selected a pose with RMSD 0.34 Å; the best PANIP pose had approximately 0.10 Å RMSD. For L99A/M102H–phenol (PDB 4I7L), PANIP again ranked the native pose 1st, while DOCK ranked it 76th, with best-pose RMSDs of approximately 0.18 Å for PANIP and 0.21 Å for DOCK. For PKM2–serine (PDB 4B2D), PANIP ranked the native pose 3rd and achieved a best pose of 0.44 Å RMSD, whereas DOCK ranked the native pose 46th and its best pose had 1.29 Å RMSD (Zeng et al., 14 Jan 2026).

Across 25 model systems, PANIP achieved a 44% native pose recovery rate at rank 1. The authors argue that this is possible because accurate short-range NCI energetics dominate pose discrimination in the relatively simple systems tested, even without explicit long-range electrostatics or solvation corrections. At the same time, the PKM2–serine example reveals a failure mode: the top PANIP pose placed the serine hydroxyl near a serine NH$\Delta E_{AB} = E(AB) - [E(A) + E(B)],$6-like fragment, indicating potential intramolecular repulsion not captured by the pairwise fragment approximation (Zeng et al., 14 Jan 2026).

Because PANIP is fast enough to score all 36.3 million PDB-derived dimers, the paper also uses it as an analysis tool for whole-PDB interaction motifs. For ETAM–PMPO, comprising 28,914 dimers, the lowest representative energy was −20.41 kcal/mol, with cation–$\Delta E_{AB} = E(AB) - [E(A) + E(B)],$7 motifs accounting for 17.7% and NH···O hydrogen-bond motifs for 19.8%. For ETAM–MIND, comprising 24,098 dimers, the strongest representative energy was −24.22 kcal/mol, with cation–$\Delta E_{AB} = E(AB) - [E(A) + E(B)],$8 motifs at 27.6% and perpendicular CH···$\Delta E_{AB} = E(AB) - [E(A) + E(B)],$9 motifs at 33.7%. For MSM–MBZ, comprising 78,674 dimers, the interaction-distance range was 3.5–7 Å, the major peak was near 5.1 Å, energies spanned −4.9 to 0.2 kcal/mol, the most populated region was around −1.6 kcal/mol, and the lowest representative energy was −4.91 kcal/mol. These analyses support the claim that PANIP is useful not only as a predictor but also as a survey instrument for biomolecular NCI landscapes (Zeng et al., 14 Jan 2026).

6. Limitations, domain of applicability, and terminological context

PANIP is explicitly a pairwise short-range fragment model. Its intended domain is pairwise NCIs between the 17 defined protein-derived fragment classes, in proteins and protein-like local environments, with emphasis on short-range interaction analysis and fragmentation-based rescoring. It is not yet a universal biomolecular potential (Zeng et al., 14 Jan 2026).

The principal limitations identified in the paper are also those expected from a pairwise fragment decomposition. PANIP does not explicitly represent intramolecular strain/repulsion, many-body polarization, long-range electrostatics, or solvent effects. The authors propose broader fragment diversity, larger chemical space, long-range electrostatics, many-body effects, and solvation as future directions, with potential extensions through force-field correction terms, ML-based correction schemes, and metamodeling approaches. This suggests that PANIP should be interpreted as a high-fidelity local interaction model rather than a complete end-to-end energetic theory for proteins (Zeng et al., 14 Jan 2026).

There is also some naming ambiguity in adjacent arXiv literature. PANIS refers to Physics-Aware Neural Implicit Solvers, a probabilistic surrogate framework for parametrized PDEs rather than a biomolecular interaction potential (Chatzopoulos et al., 2024). Panini denotes an anonymous anycast protocol, again unrelated to protein NCIs (Coijanovic et al., 2023). In panoramic image generation and inpainting discussions, “PANIP” may appear as a broader task label rather than the name of the protein-interaction model itself (Wang et al., 22 May 2025). Within the specific context of atomistic modeling of biomolecular NCIs, however, PANIP denotes the PAirwise Non-covalent Interaction Potential introduced for protein-derived fragment dimers (Zeng et al., 14 Jan 2026).