Results

Benchmark generalization, architectural ablation, representation quality, transfer learning, and retrieval-oriented enzyme prioritization.

Benchmark evaluation

Predictive performance across biochemical novelty

AIMS-ESI was evaluated against PMF-CPI, ECFP-ESM, ESP, ESP-PRE, and MEI using identical data sources, split definitions, and metric sets.

Split	ACC	AUROC	MCC	F1	Evaluation focus
Unseen enzyme	0.9337	0.9741	0.8671	0.9428	Enzyme-level shift
Unseen substrate	0.9671	0.9937	0.9342	0.9708	Substrate-level shift
Row random	0.9694	0.9949	0.9359	0.9748	Pair-level reference

AIMS-ESI benchmark accuracy and AUROC across evaluation splits — **Figure 2.** Benchmark performance across the individual evaluation splits and the cross-split mean. Accuracy and AUROC are shown for AIMS-ESI and reproduced deep-learning baselines.

Ablation and representation

Where the improvement comes from

Stepwise restoration and paired view-removal controls show that multiview evidence, hierarchical interaction modeling, and adaptive routing make complementary contributions.

Constructive ablation study of AIMS-ESI components — **Figure 3.** Constructive ablation analysis. Progressive restoration of the complete architecture yields monotonic gains; removing the fingerprint or protein language-model view reduces all four metrics.

Representation fidelity, calibration, challenge cohorts and CYP transfer results — **Figure 4.** Local neighborhood fidelity, calibration, held-out challenge cohorts, and CYP transfer. The analyses connect representation quality with balanced downstream decisions.

Application studies

From pairwise prediction to candidate prioritization

The same compatibility representation supports polymer-specific screening and repeated enzyme retrieval across individual reactions and pathway-style panels.

Polymer prioritization

Mean selectivity margins were 0.942 (PVA), 0.998 (PU), 0.870 (PHB), 0.923 (PHA), and 0.935 (phthalate).

Reaction queries

Across three nucleotide-like queries, AIMS-ESI achieved the lowest average mean positive-enzyme rank: 4.61.

Pathway steps

Across six related queries, the average mean positive-enzyme rank was 4.88 with top-10 recovery of 0.930.

Polymer-selective scores and selectivity margins — **Figure 5.** Polymer-selective scoring and off-target suppression on the public plastic-degradation benchmark. Scores indicate computational prioritization, not direct measurements of intact-polymer degradation.

Reaction-specific enzyme retrieval results — **Figure 6.** Reaction-specific retrieval for three phosphorylated nucleotide-like substrates. AIMS-ESI consistently moves labelled positives toward the top of compact candidate lists.

Pathway-style multi-reaction enzyme retrieval results — **Figure 7.** Pathway-style retrieval across six chemically related phosphatase-associated queries, highlighting ranking consistency as substrate chemistry changes.

Interpretation boundary. Application studies evaluate computational compatibility and prioritization. Prospective experimental validation remains necessary before inferring catalytic activity or polymer-degradation capability.