TFC-PACT-Net: Compressed Topological Molecular GNN

• TFC-PACT-Net is a topology-augmented molecular GNN that extracts compressed higher-order features from cell complexes to enhance prediction accuracy.
• It employs an Efficient Cellular Compression module with Principal Neighborhood Aggregation to integrate global topological signals with standard RDKit features.
• The model delivers competitive performance on various regression tasks while remaining computationally tractable and parameter-efficient.

Searching arXiv for the primary paper and closely related references to ground the article. TFC-PACT-Net denotes a molecular graph neural network introduced as PACTNet with ECC, in which ordinary 2D molecular graphs are augmented with compressed higher-order topological features extracted from a cell complex representation of each molecule (Khorana, 11 Aug 2025). Its purpose is to retain some of the global geometric and topological information usually associated with higher-order or 3D molecular models, while remaining computationally tractable and parameter-efficient. The designation should be distinguished from the unrelated robotics model “PACT: Perception-Action Causal Transformer for Autoregressive Robotics Pre-Training” (Bonatti et al., 2022), which uses the same lexical stem but addresses perception–action sequence modeling rather than molecular representation learning.

1. Name, scope, and conceptual position

In the molecular-learning setting, TFC-PACT-Net is presented as a response to a three-way trade-off among string/fingerprint models, standard molecular GNNs on 2D graphs, and geometric or higher-order models (Khorana, 11 Aug 2025). String and fingerprint methods such as SMILES, SELFIES, and ECFP are described as fast and cheap to compute, but as losing explicit 3D geometry and global topology. Standard graph models such as GCN (Kipf et al., 2016), GIN (Xu et al., 2018), GraphSAGE (Hamilton et al., 2017), and GAT (Veličković et al., 2017) respect local graph structure and perform well on many tasks, but remain limited by 1-WL expressivity and local message passing. Higher-order and geometric models are described as rich in geometric or topological fidelity, but often expensive and complex to train.

The model’s central thesis is that a substantial fraction of the benefit of higher-order or topology-aware molecular representations can be obtained without operating directly on a full higher-order neural architecture. Instead, the method first constructs a cell complex from the molecular graph, computes a collection of global and higher-order topological summaries, compresses them into a fixed-size representation, and then feeds that representation into a comparatively simple but expressive GNN based on Principal Neighborhood Aggregation (PNA) (Corso et al., 2020).

This leads to a division of labor between two components. ECC, termed Efficient Cellular Compression, performs the topological feature extraction and compression. PACTNet is the downstream GNN that consumes standard atom and bond features, degree statistics, and the compressed topological representation. A plausible implication is that the model is best understood not as a replacement for graph message passing, but as an augmentation of graph message passing with precomputed higher-order structural signals.

2. Molecular lifting and topological feature compression

The topological component begins from a molecular graph $G_M = (V_M, E_M)$, built from SMILES using RDKit, and applies a Molecular-Lifting Transformation $\mathcal{F}$ that maps the graph to a cell complex $X_M$ (Khorana, 11 Aug 2025):

$\mathcal{F}: G_M \mapsto X_M,$

where $X_M$ has skeleta $\{X_M^0, X_M^1, X_M^2, X_M^3\}$. The construction is described dimension by dimension. 0-cells correspond to atoms. 1-cells attach base-points within each atom, modeling internal substructure. 2-cells encode bonds, and the 2-skeleton $X_M^2$ is stated to be isomorphic to the original molecular graph. 3-cells encode induced cycles, chemical rings, and k-hop interactions.

The lifting is explicitly framed as a graph-to-cell-complex transformation with the property that two graphs $G_1, G_2$ are isomorphic iff their cell complexes $f(G_1), f(G_2)$ are isomorphic. In the paper’s interpretation, this ensures no information loss in moving from the molecular graph to the cell complex while enriching the representation with higher-order structure.

From $X_M$, the method computes several classes of topological and spectral descriptors. These include Betti numbers $\mathcal{F}$0, spectral chains derived from the eigen-decomposition of chain matrices or boundary operators, top-$\mathcal{F}$1 eigenvalues of cell-complex Laplacians, degree centrality over skeleta, and all-pairs shortest path distances over $\mathcal{F}$2. The $\mathcal{F}$3-Laplacian is given as

$\mathcal{F}$4

These quantities are then compressed to a fixed-size vector. The compression stage keeps top-$\mathcal{F}$5 eigenvalues, computes summary statistics, applies mean aggregation over spectral chains, forms histograms over distances or centralities, concatenates the resulting blocks, and pads them to a uniform dimension. The resulting tensor is the ECC representation of molecule $\mathcal{F}$6.

A concise way to view ECC is as a graph-level descriptor that summarizes ring structure, higher-order adjacency, global connectivity shape, and centrality patterns in a single fixed-size vector. The paper emphasizes that this vector is topologically rich yet compact.

3. PACTNet architecture and learning pipeline

The neural component operates on the ordinary molecular graph while incorporating ECC as an additional source of information (Khorana, 11 Aug 2025). The graph $\mathcal{F}$7 supplies standard node features and edge features derived from RDKit. Node features include atom type, degree, formal charge, aromaticity indicator, and possibly other standard RDKit-derived descriptors. Edge features include bond type, rotatable bond indicator, aromatic bond flag, and possibly other RDKit bond properties. The method also computes degree histograms of the graph and embeds them as additional structural signals.

The architectural core is PNA, described in the paper as the chosen convolution scheme. PNA combines multiple aggregators and scales them using degree statistics to increase expressivity beyond 1-WL. A standard formulation consistent with the reported design is

$\mathcal{F}$8

$\mathcal{F}$9

with $X_M$0 and $X_M$1 implemented by small MLPs. In the reported system, PACTNet combines three kinds of information: local neighborhood structure through PNA, global higher-order topology through ECC spectral features, and node-level connectivity statistics through degree histograms.

The forward pipeline can be summarized as follows. SMILES is converted to a molecular graph. RDKit atom and bond features are computed. Degree histograms are extracted. The graph is lifted to a cell complex, from which the ECC vector is derived. Node features are embedded, PNA layers with batch normalization are applied, and node embeddings are pooled into a graph-level embedding. The pooled embedding is then concatenated with ECC and other global features and passed to a prediction MLP.

The training objective for the reported tasks is Mean Absolute Error (MAE):

$X_M$2

Optimization uses Adam with default $X_M$3. Hyperparameter optimization is performed with Optuna TPE over a learning rate range $X_M$4, hidden dimensionalities $X_M$5, and batch sizes $X_M$6. Training runs for at most 200 epochs with early stopping of patience 20.

4. Benchmarks, datasets, and empirical behavior

The empirical evaluation covers seven regression datasets, largely from MoleculeNet together with QuanDB and BindingDB (Khorana, 11 Aug 2025). If a dataset contains more than 2,000 molecules, the reported experiments use a random subset of 2,000. IC50 and BindingDB targets are converted to pIC50 or p-affinity when appropriate. Evaluation uses nested 5-fold cross-validation, with an outer 5-fold split for performance estimation and an inner 5-fold loop for hyperparameter optimization. Validation reports mean ± standard deviation over outer folds; test performance is reported with 95% confidence intervals from non-parametric bootstrap.

Dataset	Size	Target
ESOL	1,128 molecules	water solubility (log mol/L)
FreeSolv	642 molecules	hydration free energy (kcal/mol)
Lipophilicity	4,200 molecules	logD (octanol/water distribution)
Boiling Point	2,983 molecules	boiling point (°C) in range $X_M$7
QM9	134k molecules; sampled up to 2,000	heat capacity $X_M$8 (cal/mol/K)
IC50	2,822 molecules	pIC50
BindingDB	4,614 protein–ligand complexes	binding affinity

Against GCN, GAT, GraphSAGE, and GIN baselines using ECFP, SELFIES, or SMILES-derived inputs, PACTNet+ECC is reported to achieve the best validation and test RMSE/MAE on QM9, ESOL, BoilingPoint, Lipophilicity, and FreeSolv, while remaining within 5% of the best model and statistically tied on BindingDB and IC50. Statistical testing uses Nadeau–Bengio corrected paired t-tests with Holm FWER correction; for QM9, ESOL, BoilingPoint, Lipophilicity, and FreeSolv, the reported $X_M$9-values are below $\mathcal{F}: G_M \mapsto X_M,$0 after Holm correction.

Representative results reported in the paper include the following.

Dataset	Validation	Test
QM9	RMSE $\mathcal{F}: G_M \mapsto X_M,$1, MAE $\mathcal{F}: G_M \mapsto X_M,$2	RMSE $\mathcal{F}: G_M \mapsto X_M,$3, MAE $\mathcal{F}: G_M \mapsto X_M,$4
ESOL	RMSE $\mathcal{F}: G_M \mapsto X_M,$5, MAE $\mathcal{F}: G_M \mapsto X_M,$6	RMSE $\mathcal{F}: G_M \mapsto X_M,$7, MAE $\mathcal{F}: G_M \mapsto X_M,$8
Lipophilicity	RMSE $\mathcal{F}: G_M \mapsto X_M,$9, MAE $X_M$0	RMSE $X_M$1, MAE $X_M$2

The reported pattern suggests that the model’s gains are strongest on tasks where ring topology, global connectivity, and higher-order structural organization plausibly contribute to the target. The paper does not provide an explicit ablation table isolating ECC from PNA and degree histograms, so any finer decomposition of those gains would remain an inference rather than a reported result.

5. Interpretability, efficiency, and relation to other molecular models

A defining property of TFC-PACT-Net is that its additional features are intended to remain human-interpretable (Khorana, 11 Aug 2025). Betti numbers have direct topological meanings: $X_M$3 measures connected components, $X_M$4 counts independent rings or cycles, and higher $X_M$5 capture higher-dimensional cavities in the adopted molecular-complex construction. The explicit inclusion of chemical rings and k-hop interactions as 3-cells connects the representation to recognizable chemical motifs such as aromatic rings, fused-ring systems, and scaffold-level connectivity. Spectral features of Laplacians are described as encoding global connectivity and shape, while degree centrality statistics emphasize highly connected atoms or substructures.

The model is also framed as parameter-efficient. The learned component is a relatively standard PNA-based GNN with hidden sizes in $X_M$6, while ECC itself is computed through algebraic and combinatorial procedures rather than a separate deep network. The ECC vector is fixed-size, can be computed once per molecule, and can be cached across runs. The paper further states that the method remains orders of magnitude more computationally efficient than first-principles methods like DFT, and that all experiments were run on an AWS m8g.4xlarge CPU-only instance.

This positioning places TFC-PACT-Net between conventional 2D graph learning and more expensive geometric or higher-order neural architectures. Relative to ordinary 2D GNN baselines, it adds global topological structure. Relative to direct higher-order or equivariant models, it substitutes precomputed compressed descriptors for end-to-end higher-order message passing. A plausible implication is that its practical advantage is strongest when explicit 3D equivariance is unnecessary, but global scaffold topology still matters.

6. Limitations and prospective developments

The paper explicitly identifies several limitations (Khorana, 11 Aug 2025). Benchmark coverage is limited to seven datasets, some of which are relatively small, and several larger datasets are sub-sampled to at most 2,000 molecules. The representation is not designed for ultra-fine physical detail, including tasks requiring quantum-level interaction modeling or very precise 3D conformer behavior such as force-field learning or potential-energy-surface fitting. Although ECC is intended to encode topology and coarse geometry, the method does not consume explicit 3D coordinates with equivariant operations. Scalability is also noted as a concern for large biomolecules and complexes, since cell-complex construction and spectral analysis may become more costly as system size increases.

The future directions proposed in the paper follow directly from these limitations. They include scaling to larger biomolecules and protein–ligand systems, incorporating richer topological invariants such as persistent homology, integrating equivariant layers to better capture 3D physics, and studying the training dynamics of models that rely on ECC embeddings. This suggests a broader research program in which compressed topological summaries serve as a bridge between classical molecular graph learning and more physically detailed geometric modeling.

Within current molecular representation learning, TFC-PACT-Net is therefore best characterized as a topology-augmented, parameter-efficient graph neural network: it preserves the operational simplicity of graph-based pipelines derived from SMILES and RDKit, while injecting higher-order structural information through a fixed-size cellular compression stage.