AMPTCR: Topology-Enriched Molecular Surface

• AMPTCR is a structured representation that models a molecule's surface as a canonically aligned point cloud enriched with quantum-derived scalars and topology descriptors.
• It constructs the point cloud from molecular meshes using quantum calculations and geodesic methods to capture local curvature and electronic properties.
• Evaluated with a DGCNN-based pipeline, AMPTCR shows promise in molecular weight prediction and E. coli growth inhibition, outperforming conventional descriptors.

Searching arXiv for the AMPTCR paper and closely related point-cloud / molecular-surface representation work. arxiv_search(query="(Mihalcea, 22 Jul 2025)", max_results=5) arxiv_search: {"query":"(Mihalcea, 22 Jul 2025)","max_results":5} AMPTCR, short for Aligned Manifold Property and Topology Cloud Representation, is a molecular surface representation introduced for molecular property prediction that encodes a molecule as an aligned surface point cloud whose points carry geometric, quantum-derived, and topology-aware information (Mihalcea, 22 Jul 2025). Rather than treating a molecule primarily as a SMILES string, atom-bond graph, or voxelized 3D object, AMPTCR represents sampled points on the molecular surface in a canonical reference frame and augments each point with a chemically meaningful scalar and geodesically derived topology vectors. In the reported formulation, this representation is evaluated with a DGCNN-based learning pipeline on molecular weight prediction and E. coli growth inhibition, where it is presented as a compact, expressive, and architecture-agnostic representation for surface-mediated molecular properties (Mihalcea, 22 Jul 2025).

1. Definition and representational scope

AMPTCR is a structured point cloud representation of a molecular surface manifold in which each sampled surface point is annotated with three feature classes: aligned Cartesian coordinates, one local quantum-derived scalar value, and one or more topology channels. A faithful formalization given in the technical description is

$$\mathcal{X}=\{x_i\}_{i=1}^N,\qquad x_i=[\,p_i,\;q_i,\;t_i\,],$$

where $p_i\in\mathbb{R}^3$ denotes aligned coordinates, $q_i\in\mathbb{R}$ a scalar surface property, and $t_i\in\mathbb{R}^d$ a topology descriptor block (Mihalcea, 22 Jul 2025).

The representation is motivated by a surface-centric view of molecular learning. SMILES and molecular graphs are described as atom-centric and connectivity-centric, and therefore as not directly representing molecular surface shape, local curvature, or spatially distributed electrostatic and reactivity patterns. The paper also contrasts AMPTCR with voxelization and 2D manifold projection, arguing that such approaches may blur fine local geometry or distort topology and geometric relations. A central claim is that many intermolecular and surface-mediated properties are better characterized at the level of the electron-accessible surface and its local chemical variation than at the level of graph connectivity alone (Mihalcea, 22 Jul 2025).

AMPTCR is therefore not merely a point-cloud encoding of 3D structure. Its intended contribution is the combination of three signals within one surface manifold representation: explicit geometry through aligned coordinates, local electronic context through a scalar field, and intrinsic manifold structure through topology vectors. This suggests that AMPTCR is best understood as a representation-level proposal rather than as a new neural architecture.

2. Construction pipeline and data objects

The reported pipeline begins from a molecular structure in .PDB format and generates intermediate surface data including a .PLY file before producing a compact .NPZ output containing the final point cloud and associated channels (Mihalcea, 22 Jul 2025). Quantum calculations are used to derive local electron-density-related scalar information over the molecular surface; the tools explicitly named for the workflow are Psi4 for quantum calculations, scikit-image for surface extraction, and Trimesh for mesh manipulation (Mihalcea, 22 Jul 2025).

Surface generation is described as proceeding from “charge-perturbed density grids,” from which a molecular surface mesh is extracted and assigned a scalar field reflecting local chemical reactivity or electrostatic character. The mesh is then sampled to a fixed number of evenly distributed surface points. The experiments report at least two point counts: 256 points for an initial molecular-weight experiment and 1024 points for fuller experiments (Mihalcea, 22 Jul 2025).

Each sampled point receives a single floating-point scalar value, normalized per molecule. The scalar may be ESP or the Dual Fukui function, depending on the task. The exact interpolation or projection rule used to map scalar information to sampled points is not disclosed in the manuscript body. After scalar assignment, the point coordinates are transformed into a canonical aligned frame, topology vectors are computed from the surrounding surface geometry using an intrinsic geodesic framework, and the resulting fixed-size point cloud is saved (Mihalcea, 22 Jul 2025).

The overall pointwise representation can therefore be summarized as a molecule-level set

$$\mathcal{X}=\{x_i\}_{i=1}^{N},\qquad x_i=[\,p_i,\;q_i,\;t_i\,],$$

with total feature dimension

$F=3+1+d=4+d.$

The exact value of $d$ is not numerically specified (Mihalcea, 22 Jul 2025).

3. Alignment, scalar fields, and topology descriptors

A defining element of AMPTCR is its canonical alignment step. The paper states explicitly that each molecule is aligned into a consistent local frame before learning and that the alignment is derived from both molecular geometry and the spatial distribution of the surface scalar field, but the exact method is withheld for intellectual-property reasons (Mihalcea, 22 Jul 2025). A generic abstraction of this transformation is

$p_i=R(\tilde p_i-c),$

where $c$ is a molecular center and $R$ a molecule-specific rotation chosen to define a canonical orientation. This formula is presented as a standard abstraction rather than as a disclosed algorithmic detail.

The purpose of alignment is to reduce the burden of rotational variability so that conventional SE(3)-sensitive models can be used without requiring fully equivariant architectures. The paper notes that fully equivariant 3D models can slow training by roughly 10–30×, and positions AMPTCR as a preprocessing-based alternative that shifts part of the invariance problem from the model to the representation (Mihalcea, 22 Jul 2025).

The scalar channel is the local quantum-derived component of AMPTCR. Two scalar types are used in the reported experiments. Electrostatic potential (ESP) is used in molecular-weight prediction experiments, while Dual Fukui functions, denoted F2, are used in the bacterial inhibition experiments (Mihalcea, 22 Jul 2025). The paper states that the bacterial inhibition study uses Dual Fukui functions because they may be more biologically relevant than ESP in “redox-sensitive systems.” The scalar stored at each point is a single normalized float:

$p_i\in\mathbb{R}^3$0

The topology component is the second principal innovation. The paper describes the topology vectors as custom descriptors derived from an intrinsic geodesic framework and intended to encode local curvature, directional structure, and surrounding manifold context. An abstract formulation consistent with the description is

$p_i\in\mathbb{R}^3$1

where $p_i\in\mathbb{R}^3$2 denotes a geodesic neighborhood around point $p_i\in\mathbb{R}^3$3. The exact descriptor construction, dimensionality, neighborhood scale, and summary statistics are not disclosed (Mihalcea, 22 Jul 2025).

The manuscript is also explicit about alignment failure modes. Highly symmetric molecules may admit multiple plausible canonical orientations, and the reported rotation challenges show occasional sign flip along one principal axis, producing mirrored configurations. This is important for interpretation: AMPTCR is not claimed to eliminate orientation ambiguity entirely. Instead, the reported training setup partly compensates through positional jitter, optional random whole-cloud rotation jitter, and an attention mechanism said to maintain robustness to global rotations and alignment-heuristic sign flips (Mihalcea, 22 Jul 2025).

4. Integration with learning architectures

AMPTCR is evaluated primarily with a Dynamic Graph Convolutional Neural Network (DGCNN). In the reported setup, the local K-nearest-neighbor graph is built using only the 3D positional data $p_i\in\mathbb{R}^3$4 of each point, while the topology information is passed through a shared pointwise MLP to generate a learned embedding (Mihalcea, 22 Jul 2025). This design preserves a distinction between graph construction from geometry and feature enrichment from scalar and topology channels.

Two neighborhood regimes are reported. In the initial molecular-weight test, K = 1, effectively disabling graph functionality. In the larger molecular-weight experiment and the bacterial inhibition tasks, K = 20 is used (Mihalcea, 22 Jul 2025). This distinction matters because it separates a minimal pointwise sanity check from graph-enabled experiments that exploit local neighborhood structure.

For the bacterial inhibition task, the paper adds a relational attention mechanism. With per-point features $p_i\in\mathbb{R}^3$5, the input is projected into queries, keys, and values and reshaped into $p_i\in\mathbb{R}^3$6 heads of dimension

$p_i\in\mathbb{R}^3$7

The attention logits are written as

$p_i\in\mathbb{R}^3$8

Here the three additive bias terms are derived from geometry, scalar difference, and topology similarity, respectively. The geometric term uses coordinate displacement

$p_i\in\mathbb{R}^3$9

the quantum term uses scalar difference

$q_i\in\mathbb{R}$0

and the topological term is based on the dot product of the first intrinsic topology vectors, expressed in the description as $q_i\in\mathbb{R}$1 (Mihalcea, 22 Jul 2025).

The bacterial inhibition experiments also incorporate Morgan fingerprints as an auxiliary modality. The fingerprint vector is mapped to a scalar by a dedicated MLP and blended linearly with the AMPTCR branch output. The reported FP weight is 0.25 for binary classification and 0.15 for regression (Mihalcea, 22 Jul 2025). This means the empirical evaluation does not present AMPTCR only as a standalone representation, but also as one that can be fused with established cheminformatic descriptors.

5. Empirical evaluation

The first evaluation is a molecular weight benchmark on the Bergstrom dataset, reported as 269 entries after conversion and used as a sanity check for whether the representation encodes physically meaningful information (Mihalcea, 22 Jul 2025). In an initial run using 6-fold cross validation, a 95:5 train:validation split within each fold, 30 epochs, 256 points, ESP, and KNN = 1, the reported performance is validation $q_i\in\mathbb{R}$2 and validation slope = 0.88, with training $q_i\in\mathbb{R}$3 and training slope = 0.93 (Mihalcea, 22 Jul 2025).

A larger molecular-weight experiment uses 16-fold cross validation, a 90:10 train:validation split, 20 epochs, 1024 points, and KNN = 20. On raw predictions, the reported validation metrics are $q_i\in\mathbb{R}$4, slope = 0.25, intercept = 217.43, NRMSE = 75.84\%, and ±1SD = 29.59 (Mihalcea, 22 Jul 2025). The paper interprets this as a case of strong correlation but severe output compression. A post hoc affine calibration is then introduced by fitting

$q_i\in\mathbb{R}$5

on the training set and applying the learned mapping uniformly. After this realignment, the reported validation performance becomes $q_i\in\mathbb{R}$6, slope = 1.01, intercept = -3.24, NRMSE = 39.41\%, and ±1SD = 7.75 (Mihalcea, 22 Jul 2025).

The second evaluation studies E. coli growth inhibition using a custom-curated ChEMBL subset. Molecules were required to have at least 4 recorded entries of inhibitory activity against E. coli, 50 or fewer atoms, molecular weight less than 500, and a median molecular-weight-normalized inhibition value $q_i\in\mathbb{R}$7 of 50 or less, yielding 521 entries after conversion (Mihalcea, 22 Jul 2025). For binary classification, a hit is defined as median MIC $q_i\in\mathbb{R}$8 and a non-hit as MIC $q_i\in\mathbb{R}$9, with intermediate values excluded. This produces 405 compounds, consisting of 252 hits and 153 non-hits, with class imbalance approximately 1.65:1 (Mihalcea, 22 Jul 2025).

On the binarized inhibition task, using 6-fold cross-validation and 25 epochs, the reported performance is mean ROC AUC = 0.912 ± 0.011 SE, precision = 0.881, and recall = 0.881 (Mihalcea, 22 Jul 2025). The reported Morgan fingerprints only baseline achieves mean ROC AUC = 0.715 ± 0.021, which the paper uses as the clearest baseline comparison in support of AMPTCR’s added value (Mihalcea, 22 Jul 2025).

For inhibition regression on the full 521-compound dataset, the target is transformed as

$t_i\in\mathbb{R}^d$0

and the reported protocol uses 24-fold cross-validation, a random 90:10 split each fold, 25 epochs, and post hoc calibration (Mihalcea, 22 Jul 2025). The reported regression results are mean slope = 0.71, $t_i\in\mathbb{R}^d$1, ±1SD = 0.58 \log_{10}\mu\text{M}, and intercept = -0.01. The Morgan fingerprints only baseline yields slope = 0.53, $t_i\in\mathbb{R}^d$2, and ±1SD = 1.22 \log_{10}\mu\text{M} (Mihalcea, 22 Jul 2025).

6. Interpretation, misconceptions, and limitations

AMPTCR is presented as compact, expressive, and architecture-agnostic, but each of these terms has a specific technical meaning in context (Mihalcea, 22 Jul 2025). It is compact because each surface point carries only coordinates, one scalar, and a compact topology block rather than dense voxel features or higher-order equivariant tensors. It is expressive because it combines surface location, local electronic information, and intrinsic manifold context. It is architecture-agnostic because the alignment step allows use of conventional point-cloud or graph models rather than requiring specialized equivariant architectures.

Several misconceptions are explicitly precluded by the technical description. First, AMPTCR is not a replacement for all molecular representations; it is positioned particularly for surface-mediated molecular properties such as solubility, binding affinity, minimum inhibitory concentration or growth inhibition, reactivity, permeability-related biological effects, and other intermolecular phenomena where local surface geometry and electron distribution matter (Mihalcea, 22 Jul 2025). Second, it is not inherently rotation-invariant. The representation relies on a deterministic alignment heuristic, and the paper documents residual ambiguity for highly symmetric molecules and occasional sign-flip failures. Third, it is not fully reproducible from the preprint alone, because the exact alignment method and exact topology descriptor construction are intentionally withheld for intellectual-property protection (Mihalcea, 22 Jul 2025).

The stated limitations are substantial. The quality of AMPTCR plausibly depends on quantum calculation settings, mesh extraction quality, surface sampling quality, and scalar normalization choices. The custom topology encoding has not been benchmarked against alternatives such as SHAP-related descriptors, Zernike moments, or other surface-shape descriptors. The reported experiments focus on small molecules and one biological endpoint, E. coli inhibition, leaving broader validation for future work (Mihalcea, 22 Jul 2025). A plausible implication is that the current evidence supports AMPTCR most strongly as a promising surface-centric representation for small-molecule learning rather than as a universally validated standard.

Taken together, the reported results support a precise characterization: AMPTCR is a canonically aligned, chemically annotated, topology-enriched molecular surface point cloud designed to make surface-local physical information accessible to conventional 3D deep learning pipelines (Mihalcea, 22 Jul 2025). Its main scientific claim is that combining aligned coordinates, local quantum-derived scalar fields, and intrinsic topology descriptors yields a useful representation for predicting molecular properties driven by the molecular surface.