I2Mole: Turning Molecular Pairs into Intervenable Relational Graphs

0. Link verification table

All verification dates are 2026-09-27. “Visited” means a page or PDF was actually opened during this run. “Not found” does not assert nonexistence; it means no verifiable entry point was located in the paper, official conference page, OpenReview, arXiv, author pages, or code searches.

1. Candidate pool, deduplication, and random selection

Rotation follows2026-09-27 (Sunday) in the user’s timezone, selecting AI4Science and interdisciplinary research. The local index shows recent deep reads of PemNN (single-molecule force spectroscopy), MolJO (structure-based molecular optimization), and others. These and their direct topics are excluded, as are recently covered out-of-field topics such as generative models, embodiment, and KV compression. Candidates primarily come from official ICLR2026 proceedings, cross-checked through OpenReview, arXiv, author/code pages, and public search. All candidates fall within the preferred past12-month window.

The candidate pool contains 24 papers：

1. AntigenLM: Structure-Aware DNA Language Modeling for Influenza
2. SubDyve: Subgraph-Driven Dynamic Propagation for Virtual Screening Enhancement Controlling False Positive
3. CryoNet.Refine: A One-step Diffusion Model for Rapid Refinement of Structural Models with Cryo-EM Density Map Restraints
4. Histopathology-Genomics Multi-modal Structural Representation Learning for Data-Efficient Precision Oncology
5. GeoFAR: Geography-Informed Frequency-Aware Super-Resolution for Climate Data
6. I2Mole: Interaction-aware Invariant Molecular Learning for Generalizable Property Prediction
7. Glance and Focus Reinforcement for Pan-cancer Screening
8. SC-Arena: A Natural Language Benchmark for Single-Cell Reasoning with Knowledge-Augmented Evaluation
9. BioTamperNet: Affinity-Guided State-Space Model Detecting Tampered Biomedical Images
10. PoinnCARE: Hyperbolic Multi-Modal Learning for Enzyme Classification
11. CMPhysBench: A Benchmark for Evaluating LLMs in Condensed Matter Physics
12. Learning residue level protein dynamics with multiscale Gaussians
13. Efficient Prediction of Large Protein Complexes via Subunit-Guided Hierarchical Refinement
14. Riemannian Variational Flow Matching for Material and Protein Design
15. ProTDyn: A Foundation Protein Language Model for Thermodynamics and Dynamics Generation
16. SAIR: Enabling Deep Learning for Protein-Ligand Interactions with a Synthetic Structural Dataset
17. Fast and Interpretable Protein Substructure Alignment via Optimal Transport
18. Protein Structure Tokenization via Geometric Byte Pair Encoding
19. Rigidity-Aware Geometric Pretraining for Protein Design and Conformational Ensembles
20. Fast Proteome-Scale Protein Interaction Retrieval via Residue-Level Factorization
21. SigmaDock: Untwisting Molecular Docking with Fragment-Based SE(3) Diffusion
22. MoMa: A Simple Modular Learning Framework for Material Property Prediction
23. GeomMotif: A Benchmark for Arbitrary Geometric Preservation in Protein Generation
24. One protein is all you need

After fixing candidate order, actually executed shuf -i 1-24 -n 1returned 6, selecting I2Mole. Its complete official PDF, main results, inductive/OOD tests, ablations, significance analysis, complexity, and additional regression tasks provide sufficient material, sono redraw occurred (0 times). Random selection is not popularity ranking; availability was checked after selection.

2. Motivation: beyond improving in-distribution scores

A common hidden shortcut in DDI prediction is encoding each drug into an identity-like vector, then memorizing which drugs or scaffolds co-occur with labels. Randomly splitting drug pairs can leave the same drug in training and testing, enabling reliance on identity and scaffold frequency rather than transferable interaction mechanisms. The paper defines three levels: transductive, where both drugs were seen; inductive Type 1, where only one was seen; and Type 2, where neither was seen. The difficulty increase is clear: I2Mole ZhangDDI ACC falls from transductive 88.64 to Type 1 69.12and then Type 2 61.35, a total drop of 27.29 pp(Tables 1–2, PDF p.6). Random drug-pair splits therefore poorly represent deployment on new drugs.

A second problem is conditional interactions. The same functional group may contribute differently through hydrogen bonding, hydrophobic effects, or van der Waals forces depending on its partner molecule. CGIB proposed partner-dependent core subgraphs; CMRL reframed this through conditional interventions; IGIB-ISE iteratively refines interacting substructures with EM. I2Mole argues that these still fail to turncross-molecule atom–atom relationshipsinto propagatable edges at graph construction. Earlier methods refine core extraction within separate graphs without first constructing a sufficiently concrete relational space.

A third problem is environmental change. Extracting a compressed core does not automatically ensure invariance: the remainder may encode spurious correlations involving scaffold, size, or drug identity. I2Mole quantizes non-core representations into a discrete VQ codebook rather than discarding them, then randomly replaces/recombines environment codes and requires consistent predictions for the same core. This turns avoiding environmental dependence into trainable augmentation.

Quantitative evidence also shows the task remains unresolved: on the strictest ZhangDDI Type 2, I2Mole F1 is only 39.95(Table 2, p.6); even though it improves over IGIB-ISE’s38.62 by +1.33 pp / +3.44%, absolute performance remains low. The paper offers small but stable generalization gains, not evidence that DDI prediction is clinically ready.

3. Narrative, core assumptions, and tradeoffs

The paper’s narrative has four steps.

1. First allow every relationship to exist. For molecular graphs(G_1,G_2), connect every cross-molecule atom pair with relational edges to form a merged graph(G_m). Hydrogen bonds and short-range interactions thereby gain a representational channel rather than being inferred only after independent graph vectors are combined.
2. Then learn which relationships to retain. Attention/gating weights relational edges, pruning low-weight ones and retaining about 20% by default. This shrinks the fully connected relational graph into a candidate interaction skeleton.
3. Separate core and environment within that skeleton. The GIB module samples a task-sufficient, input-compressed rationale(G_sub); the remainder is(G_env).
4. Train invariance through discrete environment recombination. Encode environments into a VQ codebook, sample environment codes, and concatenate them with core representations, forcing stable labels under environmental changes.

The central assumption is conditional independence between label and environment given the relational core: (Y \perp G_{env}\mid G_{sub}). This is more plausible than assuming a fixed substructure determines a drug’s properties, because the core is extracted for a specific pair. Nevertheless, it remains a modeling assumption, not an experimentally proven causal fact.

Tradeoffs are explicit. All cross-molecule edges introduce quadratic complexity; VQ recombination expands environmental coverage but may create chemically unrealized core–environment combinations; the information bottleneck improves robustness but loses task information when too strong. The paper’s phase-transition evidence shows that raising β from the optimum(10^{-4}) to 0.1 collapses ZhangDDI ACC from 88.64 to 56.46, a drop of 32.18 pp（Table 5，p.8）。

4. Formalization and procedure

Let molecular graphs be(G_i=(V_i,E_i,X_i)) and labels(Y\in{0,1}). The merged graph adds candidate cross-molecule edges

[ E_{12}={(u,v):u\in V_1,v\in V_2},\qquad G_m=(V_1\cup V_2,E_1\cup E_2\cup E_{12}). ]

GatedConv propagates local within-bond information, while GAT updates cross-molecule relationships. Relational edge(e_uv) receives importance(a_uv), and a fractionρ is retained. The main configuration usesρ=20%, but Table 9(p.22) shows ACC varies only within 88.52–88.64 for 5%–30%, indicating no sharp threshold.

The theoretical GIB objective(Eq.1, p.2) is

[ \min_{G_{sub}} -I(Y;G_{sub})+\beta I(G_m;G_{sub}), ]

which retains label-relevant information while compressing dependence on the original graph. The paper further uses

[ Y\perp G_{env}\mid G_{sub} ]

to express invariance(Eq.2–3, p.2). Environment encoding(z_e=f_env(G_env)) is assigned to its nearest codeword

[ q(z_e)=e_{k^},\quad k^=\arg\min_k|z_e-e_k|_2^2, ]

An environment code(e_j) is then sampled from the codebook, concatenated with core representation(z_c), and used to predictŷ. Total loss(Eq.24, p.6) is

[ \mathcal L_{total}=\mathcal L_{inv}+\mathcal L_{pre}+\beta\mathcal L_{MI}+\gamma\mathcal L_{vq}. ]

where ℒpre is original-example BCE, ℒinv constrains predictions on recombined environments, ℒMI implements the bottleneck, and ℒvq updates the discrete codebook.

Inputs: drug graphs G1,G2; label y; retention ratioρ; environment sample countθ
1  Run3 GatedConv layers separately within G1 and G2
2  Add all cross-molecule atom-pair edges and run3 GAT layers
3  Estimate cross-molecule edge weights; retain only top-ρ edges
4  GIB samples core subgraph G_sub; treat the remainder as G_env
5  Encode G_env and quantize into the VQ codebook
6  Sampleθ environment codes and recombine with the same core
7  Jointly minimize original prediction, recombination invariance, mutual-information, and VQ losses
Outputs: DDI probability and an interpretable core relational subgraph

5. Key mechanisms and engineering details

6. Datasets, splits, and metrics

Transductive evaluation uses 60/20/20 splits. Inductive evaluation distinguishes Type 1(one unseen drug) and Type 2(both unseen). Each setting repeats 8 timeswith means and standard deviations reported(Sec.5, p.6; Appendix J, p.22). ACC depends on class proportions; AUROC measures ranking; F1 emphasizes positive-class precision/recall balance. PR-AUC, calibration error, sensitivity/specificity, and decision curves are absent, so these tables cannot establish clinical alert quality.

7. Main results: complete data matrices

8. Component ablations: what actually drives the gains?

ZhangDDI ablations(ACC/AUROC/F1, Table 4, p.8):

First, the largest contribution comes from the VQ environment module, not the cross-molecule edges emphasized in the title. Removing VQ causes an ACC loss 3.42 timesthat of removing cross-interactions. Second, cross-molecule interactions and GIB have similar marginal magnitudes. Third, these are leave-one-module-out ablations of a combined system; modules may complement each other, so 14.12,4.13, and 3.92 cannot be added as independent contributions.

Table 11(p.23) replaces VQ environment recombination with random noise. Random Distribution ACC on Zhang/ChCh/Deep is 87.21/93.47/92.39, in-distribution noise gives 88.02/94.27/94.56, and full VQ gives 88.64/95.34/96.51. Generic noise regularization therefore cannot explain all gains; on DeepDDI, VQ exceeds random noise by +4.12 pp. However, no head-to-head comparison uses an equal-parameter learnable continuous environment model.

9. Hyperparameter sensitivity, scale transitions, and statistical significance

Bottleneck strength β: For(0,10^{-5},10^{-4},10^{-3},0.1), Zhang ACC is respectively 85.71、85.84、88.64、86.83、56.46(Table 5, p.8).(10^{-4}) is a narrow peak, while 0.1 causes catastrophic compression. This is an information-retention versus spurious-correlation-removal transition, not monotonically stronger invariance.

VQ weight γ: From(2\times 10^{-5}) to(10^{-3}), ACC stays within 88.21–88.64 (Table 5, p.8), indicating relative flatness with respect to γ. Exact codebook-size curve values are absent; the paper states only that 10 codewords is best. This report does not estimate values from pixels.

Relational-edge retention: At 5%,10%,20%,30%,50%, ACC is 88.52、88.60、88.64、88.54、88.34(Table 9, p.22).20% is best, but5% is only 0.12 pp lower, suggesting ranking edges matters more than the exact retained fraction.

Environment sample count θ: Table 10(p.23) compares 2–6 samples, withθ=4 best. Gains are not monotonic; excessive recombination may introduce implausible combinations. Means/standard deviations are reported, but no multiple-comparison correction is applied to this sweep.

Model scale: From23.4M to 49.0M parameters, ACC is respectively 83.61、85.94、87.43、88.27、88.64、88.64、88.35、87.91、86.87(Tables 12–13, p.24). Performance plateaus at 35.4–38.4M, then declines with further growth; training time changes from 7.4h without JVP to 22.3h. This is a clear overparameterization turning point rather than a scaling law.

Significance: Appendix Table 8(p.22) gives independent t-tests against the runner-up: Zhang88.08±0.26 vs 88.64±0.24,(p=3.52\times 10^{-4}); ChCh94.92±0.21 vs 95.34±0.19,(p=2.13\times 10^{-3}); Deep95.85±0.19 vs 96.51±0.14,(p=8.74\times 10^{-5}). All are below 0.01, but normality, equal variance, pairing, and cross-dataset/metric corrections are unspecified. Statistical significance does not imply large deployment value: absolute ACC gains remain0.42–0.66 pp.

10. Efficiency and reproduction realities

Table 14(p.25) reports I2Mole with 35.4M parameters, training on ZhangDDI for 13.23hand using 17.0GB; CGIB uses 11M、1.50h、5.1GB. Thus I2Mole has approximately 3.22×the parameters, 8.82×the training time, and 3.33×the GPU memory, in exchange for transductive ACC +1.32 pp. On DeepDDI, I2Mole training takes 33.07h，CGIB 3.73h, likewise about 8.87×.

Table 16(p.26) replaces the main network with lighter backbones: GIN reaches ACC 91.35, training in 188.10sand testing in 19.27s; the main configuration gives 95.34,1020.62s,90.61s. GIN training is faster by 5.43×and testing by 4.70×, at a cost of −3.99 pp. The method therefore need not use a heavy backbone, but applications should plot accuracy–latency–memory Pareto frontiers instead of reporting only the main configuration.

11. Genuine contributions, reusable design, and transfer potential

Actual contribution 1: moving relational priors earlier. Earlier models interact after separately encoding molecules; I2Mole incorporates cross-molecule relationships into message passing. This transfers to solute–solvent, drug–target, and material-component pairs, but geometry/distance priors are needed to control fully connected edge counts.

Actual contribution 2: turning environment from discarded residual into a sampleable variable. The VQ codebook supplies an actionable environment-replacement interface. More generally, graph tasks with core/background decompositions can try discrete background quantization and constrained counterfactual recombination.

Actual contribution 3: a relatively complete evidence chain. Main results, Type 1/2, cross-domain tests, ablations, β/γ, relational-edge ratios, sample counts, noise replacements, scale, efficiency, additional regression tasks, and t-tests jointly reveal where gains arise and their costs.

Reusable design means three principles rather than copying every module: explicitly model inter-object relations, parameterize potentially spurious environments, and use strict entity-disjoint splits to test transfer. Protein–ligand tasks can replace full connectivity with radius graphs; material mixtures can constrain environment codes to temperature, concentration, or phase; clinical polypharmacy requires extending pairs to hypergraphs.

12. Limitations acknowledged by the paper

1. Only drug pairs are modeled. The paper notes that patients with comorbidities often take several drugs, while the framework handles only binary relationships and needs higher-order polypharmacy extensions(Limitations, PDF p.25).
2. Fully connected cross-molecule edges expand rapidly. All atom-pair edges become computationally expensive for large molecules; the authors suggest sparse interaction mechanisms(Limitations, p.25).
3. Physical environmental variables are missing. Drug function depends on temperature, pH, and other conditions not explicitly included in current inputs(Limitations, p.26).

13. Independent critique: evidence–inference chains

1. [Attribution] Because removing VQ lowers ACC by 14.12 ppwhile removing cross-interactions lowers it only by 4.13 pp(Table 4, p.8), the title’s interaction-aware component is neither the sole nor largest source of gains. More accurate attribution is discrete environment augmentation as the main driver, supplemented by relational graphs.
2. [External validity] Because F1 on strict Zhang Type 2 is only 39.95(Table 2, p.6), positive-class recognition for two entirely new drugs remains weak despite significant baseline gains. Transductive88.64 ACC cannot establish safety for actual new drugs.
3. [Efficiency] Because I2Mole takes approximately 8.82×CGIB’s ZhangDDI training time and 3.22×its parameters for only 1.32 pphigher ACC(Tables 1,14), resource-constrained screening may favor CGIB/light backbones. Cost-normalized rankings are needed.
4. [Causality] Because(Y\perp G_env\mid G_sub) is a training assumption and VQ codewords are not aligned with real temperature, pH, dosage, or metabolic pathways, environment intervention means representation-level augmentation, not actual chemical intervention.
5. [Adequacy of controls] Because VQ beats random noise but lacks equal-capacity continuous-mixture, learnable Gaussian-environment, or shuffled-but-valid-scaffold controls, gains cannot be uniquely attributed to discrete semantics rather than extra parameters or a better noise distribution.
6. [Statistics] Because many datasets, metrics, and settings are compared but only three ACC results receive independent t-tests without pairing/correction details, significance coverage is limited. Paired8-seed tests, effect sizes, confidence intervals, and Holm correction would be stronger.
7. [Data splitting] Because appendix training-ratio experiments simultaneously alter training and test counts(test share falls from 45% to 5%), the ACC increase from 53.27 to 92.25(Tables 12–13, p.24) mixes more training data with a smaller test set and cannot be interpreted purely as sample efficiency.
8. [Reproduction] Because the PDF’s anonymous.4open code link was inaccessible and the hyperparameter text suspiciously combines initial LR(10^{-4}) with “increased to 0.5,” final implementation details remain unconfirmed. A stable de-anonymized repository and commit hash are needed.

14. Related work: problem origins and technical transfer

                                                  ┌─ MoleOOD: unknown environments → invariant substructures
Environment/OOD ─────────┤
                         └──────────────┐
                                        ├─ I2Mole: VQ environment codes + recombination invariance
Information bottleneck ─ GIB ─ CGIB: partner-conditioned core ─┤
                                        │
Causal branch ─ CMRL: conditional interventions/shortcuts ─────┤
                                        │
Interaction extraction ─ IGIB-ISE: iterative EM coupling ─────┘
                                                │
                                                └─ Explicit cross-molecule edges + joint core/environment modeling

15. Open challenges: actionable head-to-head experiment gaps

16. Recent-work radar(verified links)

1. GenRel-DDI: Rethinking Drug-Drug Interaction Modeling as Generalizable Relation Learning(2026): shifts representation from drug identity to relations themselves, directly challenging the assumption that larger molecular encoders suffice for generalization; a strong relational-representation comparator for I2Mole.
2. Cross-Domain Molecular Relational Learning(2026): addresses cross-domain MRL through structure–semantics transfer, complementing I2Mole’s predominantly single-modality molecular-graph modeling.
3. BOOM(2025): systematically compares 140+ model–task combinations and finds the best models still have average OOD error3× ID error, cautioning against broad OOD claims from one DDI split.
4. SigmaDock(ICLR2026): introduces structured chemical priors into docking through rigid-fragment SE(3) diffusion, illustrating the value of designing physical relational spaces first.
5. ProTDyn(ICLR2026): unifies protein conformational ensembles and multi-timescale dynamics, suggesting I2Mole environments could represent dynamic states rather than only a static codebook.
6. SC-Arena(ICLR2026): evaluates single-cell natural-language reasoning with knowledge augmentation, emphasizing biological evidence alongside AI4Science scores.
7. AntigenLM(ICLR2026): preserving genomic functional units in pretraining substantially outperforms shuffling/fragmentation, sharing I2Mole’s theme that correct structural units matter more than indiscriminate model scaling.

17. Conclusion

I2Mole is worth reading because it makes relational formation and environmental variation trainable, beyond raising transductive ACC by less than 1 pp. Its key evidence consists of three mutually constraining numbers: full-model ACC exceeds the strongest baseline by only 0.42–0.66 pp; removing VQ costs 14.12 pp; training cost relative to CGIB is approximately 8.82×. Together, these suggest discrete environment recombination can improve robustness, but attribution, efficiency, and chemical validity need stricter controls.

For researchers, the reusable element is the relational-graph + core/environment decomposition + environment-recombination interface. For practitioners, cautions are low absolute Type 2 F1, inaccessible code, and absent calibration/prospective validation. It is a reasonable engineering blueprint for robust molecular relational learning, but evidence is far short of a clinical DDI decision system.

Index status: This report updated ml-report-index.mdwith 2026-09-27, AI4Science and interdisciplinary research, I2Mole, ICLR2026, and a one-sentence conclusion.
Numerical audit: All quantitative values come from the official I2Mole PDF actually read or explicitly cited official related-paper pages/PDFs. Missing matrices were not invented from secondary abstract summaries. Unreported PR-AUC, calibration, clinical thresholds, confidence intervals, and exact codebook-curve points remain explicitly missing.