Atom-Wise Allocation of Message-Passing Scale for Molecular Property Prediction
Organizations: Biomedical Imaging Institute and Department of Radiology and Biomedical Imaging, Yale University, New Haven, CT 06520, USA · Department of Computing and Data Science, The University of Hong Kong, Pokfulam Road, Hong Kong SAR, China · Department of Electrical and Electronic Engineering, The University of Hong Kong, Pokfulam Road, Hong Kong SAR, China · Department of Computer Science and Technology, University of Cambridge, Cambridge CB3 0FD, United Kingdom
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.
Figures & tables
| Model | ESOL | FreeSolv | Lipophilicity |
|---|---|---|---|
| RF (Morgan, ) | |||
| D-MPNN (Chemprop) | |||
| GCN | |||
| GAT | |||
| GIN | |||
| EquiformerV2 |
| Model | BACE | BBBP | SIDER |
|---|---|---|---|
| RF (Morgan, ) | |||
| D-MPNN (Chemprop) | |||
| GCN | |||
| GAT | |||
| GIN | |||
| EquiformerV2 |
| Model | Params (M) | MFLOPs per molecule | Time per epoch (s) | Trials | ESOL | Lipophilicity |
| RF (Morgan, ) | N/A | N/A | 0.72 a | 10 | ||
| D-MPNN (Chemprop) | 1.51 | 129.8 | 0.61 | 10 | ||
| GCN | 1.46 | 16.5 | 0.31 | 10 | ||
| GAT | 1.06 | 47.8 | 0.27 | 10 | ||
| GIN | 2.05 | 70.3 | 0.29 | 10 | ||
| EquiformerV2 | 26.66 | 16993.8 | 3.83 | 10 |
| PRISM (full) | Without | With | Mean pooling only | Sum pooling only | No adaptive allocation | |
|---|---|---|---|---|---|---|
| SIDER ROC-AUC |
| Readout of the pre-trained GIN | ||||
|---|---|---|---|---|
| Endpoint | Last layer | JK attention | PRISM a | PRISM + fp bank |
| ESOL (RMSE ) | ||||
| FreeSolv (RMSE ) | ||||
| Lipophilicity (RMSE ) | ||||
| BACE (ROC-AUC ) | ||||
| BBBP (ROC-AUC ) | ||||
| Backbone | Layer combination | ESOL | FreeSolv | Lipo | BACE |
|---|---|---|---|---|---|
| GCN | last layer | ||||
| GCN | JK attention | ||||
| GCN | PRISM | ||||
| GIN | last layer | ||||
| GIN | JK attention | ||||
| GIN | PRISM |
| Perturbation | ESOL | FreeSolv | Lipo | BACE | BBBP | SIDER |
|---|---|---|---|---|---|---|
| Permuted | ||||||
| Molecule mean | ||||||
| PRISM allocation on the pre-trained bank (Table 5 ) | ||||||
| Pre-trained, permuted | ||||||
| Pre-trained, mean | ||||||
| On plateau (%, full model) | 13.2 | 39.6 | 2.5 | 31.5 | 76.2 | 3.5 |
| Model | |||
|---|---|---|---|
| Same six-layer GIN bank, readout differs | |||
| GIN, last layer | |||
| GIN, uniform average | |||
| GIN, shared layer attention | |||
| Node-wise layer gate | |||
| JK, node-wise attention | |||
Appendix figures & tables10 assets
Supplementary material from the paper’s appendix.
Appendix
| Dataset | Hidden | Layers / | Dropout | LR | Batch | Params (M) | |||
|---|---|---|---|---|---|---|---|---|---|
| ESOL | 256 | 2 / 5 | 0.3 | 32 | 0.1 | 2.01 | 0.47 | 11.43 | |
| FreeSolv a | 128 | 5 / 5 | 0.1 | 32 | 0.01 | 2.00 | 0.49 | 4.79 | |
| Lipophilicity b | 192 | 5 / 5 | 0.1 | 128 | 0.001 | 2.04 | 0.18 | 7.88 | |
| BACE c | 64 | 3 / 5 | 0.1 | 32 | 0.1 | 2.03 | 0.31 | 2.30 | |
| BBBP a,d | 300 | 4 / 5 | 0.1 | 32 | 0.1 | 2.04 | 0.53 | 14.14 | |
| SIDER a,d | 64 | 6 / 6 | 0.1 | 32 | 0.05 | 2.52 | 0.24 | 2.19 |
| Layers | Model | MAE | MiB | s/epoch | Total min |
|---|---|---|---|---|---|
| 5 | GIN | 0.2663 | 1250 | 2.73 | 4.54 |
| 5 | PRISM | 0.2523 | 2270 | 4.16 | 6.94 |
| 7 | GIN | 0.2676 | 1250 | 2.90 | 4.83 |
| 7 | PRISM | 0.2549 | 2272 | 4.36 | 7.26 |
| 9 | GIN | 0.2670 | 1250 | 3.15 | 5.24 |
| 9 | PRISM | 0.2537 | 3252 | 4.53 | 7.55 |
| Layers | ( ) | MAE ( ) | ( ) | MAE ( ) | |
|---|---|---|---|---|---|
| 5 | 3.00 | 0.2529 | 2.00 | 2 | 0.2538 |
| 7 | 4.20 | 0.2534 | 3.00 | 3 | 0.2549 |
| 9 | 5.40 | 0.2527 | 4.00 | 4 | 0.2530 |
| 11 | 6.60 | 0.2513 | 5.00 | 5 | 0.2531 |
| 15 | 9.00 | 0.2529 | 7.00 | 7 | 0.2552 |
| Dataset | 0.00 | 0.05 | 0.10 | 0.15 | 0.20 |
|---|---|---|---|---|---|
| BACE | |||||
| BBBP | |||||
| ESOL | |||||
| FreeSolv |
| Dataset | mean | SD | 5th to 95th pct. | % | % | SD | |||
|---|---|---|---|---|---|---|---|---|---|
| ESOL | 5 | 2.02 | 0.46 | 1.16 | 1.15 | 0.10 to 3.88 | 0.0 | 0.9 | 1.181 |
| FreeSolv | 5 | 2.01 | 0.49 | 0.90 | 0.86 | 0.09 to 3.97 | 36.4 | 11.2 | 0.889 |
| Lipophilicity | 5 | 2.04 | 0.14 | 2.18 | 1.01 | 0.01 to 4.00 | 0.0 | 17.7 | 1.004 |
| BACE | 5 | 2.02 | 0.37 | 2.98 | 0.82 | 0.99 to 3.96 | 0.0 | 25.6 | 0.831 |
| BBBP | 5 | 1.99 | 0.57 | 0.35 | 0.65 | 0.08 to 3.30 | 73.5 | 0.7 | 0.695 |
| SIDER | 6 | 2.53 | 0.24 | 3.73 | 1.01 | 1.04 to 5.00 | 0.0 | 6.5 | 1.011 |
| mean | 5th to 95th pct. | % | % | ||||
|---|---|---|---|---|---|---|---|
| 5 | 2 / 2 | 3.00 / 2.00 | 2.41 / 2.29 | 2.92 / 2.26 | 2.90 to 2.93 / 2.24 to 2.29 | 0.0 / 0.0 | 0.0 / 0.0 |
| 7 | 2 / 3 | 4.20 / 3.00 | 2.13 / 2.18 | 4.54 / 3.56 | 4.53 to 4.56 / 3.54 to 3.58 | 0.0 / 0.0 | 0.0 / 0.0 |
| 9 | 2 / 4 | 5.40 / 4.00 | 2.24 / 2.06 | 5.72 / 5.21 | 5.69 to 5.74 / 5.18 to 5.22 | 0.0 / 0.0 | 0.0 / 0.0 |
| 11 | 2 / 5 | 6.60 / 5.00 | 2.12 / 2.15 | 6.68 / 5.36 | 6.66 to 6.70 / 5.31 to 5.41 | 0.0 / 0.0 | 0.0 / 0.0 |
| 15 | 2 / 7 | 9.00 / 7.00 | 2.07 / 1.93 | 8.91 / 8.54 | 8.89 to 8.91 / 8.46 to 8.62 | 0.0 / 0.0 | 0.0 / 0.0 |
| Model | M params | ms/batch | MiB | RMSE |
|---|---|---|---|---|
| SchNet Schütt et al. (2017) | 0.39 | 16.64 | 5102 | 2.353 0.054 |
| DimeNet++ Gasteiger et al. (2020) | 2.28 | 91.84 | 43,188 | 2.455 0.030 |
| MACE Batatia et al. (2022) | 1.44 | 114.88 | 15,918 | 1.727 0.032 |
| EquiformerV2 Liao et al. (2024) | 1.80 | 32.96 | 6368 | 1.815 0.040 |
| PRISM | 1.90 | 20.48 | 4774 | 1.531 0.027 |
| Endpoint | Baseline | Baseline mean | 95% CI | ||
|---|---|---|---|---|---|
| ESOL | RF (Morgan, ) | 1.340 | 0.014 | ||
| ESOL | D-MPNN (Chemprop) | 0.914 | 0.009 | ||
| ESOL | GCN | 0.789 | 0.494 | ||
| ESOL | GAT | 0.794 | 0.189 | ||
| ESOL | GIN | 0.746 | 0.880 | ||
| ESOL | EquiformerV2 | 0.759 | 0.649 |
| Endpoint | Model | P0 | P1 | P2 | P3 | P4 | Better | Best of all |
|---|---|---|---|---|---|---|---|---|
| ESOL (RMSE) | PRISM | 0.646 | 0.757 | 0.755 | 0.748 | 0.798 | 3/5 | 1/5 |
| GIN | 0.667 | 0.726 | 0.867 | 0.649 | 0.821 | |||
| FreeSolv (RMSE) | PRISM | 0.491 | 0.460 | 0.359 | 0.583 | 0.399 | 3/5 | 0/5 |
| EquiformerV2 | 0.610 | 0.499 | 0.337 | 0.409 | 0.438 | |||
| Lipophilicity (RMSE) | PRISM | 0.601 | 0.659 | 0.634 | 0.655 | 0.589 | 0/5 | 0/5 |
| GAT | 0.547 | 0.591 | 0.605 | 0.635 | 0.579 |
| Variant | BACE | BBBP | SIDER | ESOL | FreeSolv | Lipo |
|---|---|---|---|---|---|---|
| PRISM (full) | ||||||
| Component removals | ||||||
| Without | ||||||
| With | ||||||
| Mean pooling only | ||||||
| Sum pooling only | ||||||
Message Passing Neural Networks
Message passing neural networks update graph representations by repeatedly combining information from neighboring nodes and edges.
See all 72 papers