Atom-Wise Allocation of Message-Passing Scale for Molecular Property Prediction

Regina Qianru Zhang, Honggang Wen, Zongru Li, Jinsong Ouyang, Zhiqin Chu, Pietro Liò, Xiaofeng Liu, Siu-Ming YiuarXiv 2026 · arXiv

TL;DR

PRISM (Per-atom Routing over Intermediate States of Message Passing) is a novel layer-combination rule for Graph Neural Networks (GNNs) that allows each atom in a molecule to dynamically choose its own optimal message-passing scale. By learning to position a unimodal distribution over intermediate states, PRISM improves prediction accuracy across diverse molecular tasks and backbones, effectively acting as a learned, task-dependent version of classical "atom typing."

What Problem Does This Paper Address?

In molecular property prediction, different chemical features matter at different scales. For example, a molecule's solubility might be dominated by local polar groups, while its bioactivity might depend on extended aromatic substructures. Standard GNNs, however, typically fix a single propagation scale (number of layers) for every atom in a molecule.

While deeper networks can capture more context, they often suffer from "over-smoothing," where atom representations become indistinguishable. Existing solutions like Jumping Knowledge (JK) networks allow for node-dependent layer selection, but they are often unconstrained and difficult to analyze. The authors identify a need for a "middle ground": a mechanism that is more expressive than a shared global profile but more constrained and interpretable than free per-layer weights.

Methods

The core of the paper is PRISM, a routing mechanism that sits on top of standard GNN operators (like GCN, GIN, or GAT).

  1. Atom-Wise Routing: For every atom $i$, the model computes a bounded scalar $\kappa_i \in (-1, 1)$ based on the atom's initial features.
  2. Unimodal Allocation: This scalar $\kappa_i$ acts as a coordinate that positions a unimodal (Gaussian-like) distribution over a "bank" of retained message-passing states from every layer. A low $\kappa_i$ weights the atom toward shallow, local information; a high $\kappa_i$ weights it toward deep, global information.
  3. Representation Bank: The bank stores the post-update states of $L$ layers. In a specialized "fingerprint-resolved" version, these states are augmented with circular-fingerprint (ECFP) embeddings at matching radii.
  4. Local Branch & Readout: PRISM includes a parallel "local branch" to preserve immediate bonded information. The final representation is a mixture of the allocated bank states and the local branch, which is then pooled (using both mean and sum pooling) for the final prediction.
The PRISM architecture: (a) Routing atom features to a scalar, (b) Allocating a unimodal distribution over a bank of layer states, and (c) The final readout process.

The authors highlight that because a single coordinate translates the entire profile, the allocated depth responds to the routing scalar in a closed-form, non-decreasing way, making the model's behavior mathematically analyzable.

Datasets

The researchers evaluated PRISM on several standard MoleculeNet benchmarks using scaffold partitions (a more challenging and realistic split than random partitioning):

  • Regression: ESOL (solubility), FreeSolv (hydration free energy), and Lipophilicity.
  • Classification: BACE, BBBP, and SIDER.
  • Long-Range: Peptides-struct (from the Long Range Graph Benchmark).
  • Synthetic: A custom "Atom-Dependent-Radius" (ADR) benchmark designed to reward models that can distinguish between local and global information requirements for different atom types.

Key Insights and Ideas

  • Learned Atom Typing: The model learns to treat atoms differently based on their chemical identity. For instance, in hydration energy tasks, the model consistently allocates deeper information to Nitrogen and Oxygen atoms compared to the model mean, while allocating shallower information to halogens in SIDER tasks.
  • Constraints Matter: Surprisingly, the constrained one-parameter PRISM profile performs as well as, or significantly better than, unconstrained node-wise gates or Jumping Knowledge attention. This suggests that the "unimodal" assumption (that an atom needs a specific range of context) aligns well with chemical intuition.
  • Spatial Heterogeneity: By permuting the learned profiles among a molecule's atoms (shuffling which atom gets which depth), the authors observed a significant spike in prediction error. This proves that the model isn't just learning a better global average, but is successfully assigning specific scales to specific atoms.
  • Backbone Agnostic: PRISM isn't a new GNN layer; it's a readout rule. It successfully improved performance when applied to GCN, GIN, and GAT backbones, and even when added to a large-scale pre-trained GIN model.

Key Results

  • State-of-the-Art Performance: PRISM achieved the lowest mean error on ESOL and FreeSolv among nine competitive baselines (including geometric models like EquiformerV2 that use 3D coordinates, which PRISM does not).
  • Significant Error Reduction: When transferred to a GCN backbone, PRISM reduced hydration free energy error by 29% relative to a standard last-layer readout.
  • Pre-training Boost: Applied to a pre-trained GIN, PRISM improved performance on four out of six endpoints and achieved the lowest overall error on the Lipophilicity task.
  • Long-Range Efficiency: On the Peptides-struct benchmark, PRISM reduced error by up to 6.9% compared to a standard GIN at matched depths.

Limitations and Future Work

  • Computational Cost: PRISM does not offer "early exit" savings. Because the model must weight all layers, it must compute all $L$ layers for every atom. This results in a substantial overhead: 41%–52% more time per epoch and 74%–176% more peak memory compared to a standard GIN.
  • Interpretability Limits: While the routing scalar is mathematically elegant, its direct physical interpretation (e.g., "this atom needs exactly 3.5 Å of context") remains difficult to prove without further structural intervention studies.
  • Future Directions: The authors suggest that pairing per-atom depth allocation with interaction modules or specialized fingerprint channels could further push the boundaries of molecular representation learning.

Summary updated

Abstract

Molecular properties depend on chemical structure at different ranges, from local polar groups governing solvation to extended substructures shaping bioactivity, yet graph neural networks usually fix one propagation scale for every atom. We present PRISM (Per-atom Routing over Intermediate States of Message Passing), in which each atom's encoded features produce a bounded scalar that positions a unimodal distribution over retained message-passing states and matching circular-fingerprint radii. Because one coordinate translates the whole profile, allocated depth responds to the routing scalar in closed form: at fixed temperature it is nondecreasing, its sensitivity scales with the profile variance, and it saturates in the clipped regions. Beyond providing closed-form analyzability, this constrained one-parameter profile matches or significantly outperforms parameter-matched, unconstrained node-wise gates and Jumping Knowledge attention. Under scaffold partitions with equal tuning budgets against eight graph, fingerprint, and geometric baselines, PRISM trained from scratch without three-dimensional coordinates achieves the lowest mean error on aqueous solubility and hydration free energy and strong accuracy on BACE and BBBP. As a general layer-combination rule, PRISM transfers directly to GCN, GIN, and GAT backbones, reducing hydration free energy error by up to 29% relative to last-layer readout. Applied to a pre-trained GIN, it improves on last-layer readout on four of six endpoints and gives the lowest lipophilicity error overall. In trained models, allocated depth differs systematically between atom types in a task-dependent way, and permuting profiles among a molecule's atoms raises prediction error on several endpoints, so the rule acts as a learned counterpart of atom typing.
BibTeX
@misc{zhang2026atom,
  author = {Regina Qianru Zhang and Honggang Wen and Zongru Li and Jinsong Ouyang and Zhiqin Chu and Pietro Liò and Xiaofeng Liu and Siu-Ming Yiu},
  title = {{Atom-Wise Allocation of Message-Passing Scale for Molecular Property Prediction}},
  year = {2026},
  eprint = {2611.02277},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2611.02277},
}
Figure