ProtoPathway: Pathway-Centric Computational Models

• ProtoPathway is a pathway-centric framework that treats pathways as primary analytical units, integrating biological constraints for enhanced interpretability.
• It fuses multimodal data from imaging and genomics in cancer survival prediction, achieving high performance through biologically meaningful tokens.
• The approach extends to signaling reconstruction, metabolic design, and chemical reaction network analysis by enforcing explicit, constraint-based pathway representations.

Searching arXiv for the specified ProtoPathway-related papers and directly relevant context. Searching arXiv for ProtoPathway and the cited pathway-method papers. ProtoPathway denotes a pathway-centric style of computational modeling in which pathways, path segments, or path-like tokens are treated as primary analytical objects rather than as post hoc summaries. In its most specific recent usage, it is an interpretable-by-design multimodal framework for cancer survival prediction that fuses morphological prototype tokens from whole-slide imaging with pathway embeddings derived from a bipartite gene–pathway graph (Gallagher-Syed et al., 20 May 2026). In the broader literature summarized under the same label, ProtoPathway also refers to technical syntheses for signaling pathway reconstruction in yeast (Inostroza et al., 2020), protocell-inspired adaptation circuits (Palo et al., 2013), metabolic trunk and retrosynthetic pathway design (Court et al., 2014, Zhang et al., 15 Apr 2026), automated chemical reaction-network exploration (Simm et al., 2017), energy-based pathway decomposition (Gawthrop et al., 2016), interpretable pathway annotation rules (Boudellioua et al., 2015), threshold-based path formation (Proverbio et al., 16 Jul 2025), and allosteric route inference in proteins (Wu et al., 2022).

1. Conceptual scope and recurring structure

Across these usages, ProtoPathway is consistently associated with explicit structural representations of pathways and with ranking mechanisms that are native to the representation itself. In the cancer setting, the fundamental objects are stable prototype and pathway tokens; in yeast signaling, they are length-bounded directed paths filtered by causal timing and protein-complex cohesion; in allostery, they are residue-residue paths of optimised propensity; in metabolism and chemistry, they are enumerated reaction trunks or multistep retrosynthetic routes; in rule mining, they are antecedent-to-pathway implication models; and in threshold routing, they are minimum-threshold paths in a graph (Gallagher-Syed et al., 20 May 2026, Inostroza et al., 2020, Wu et al., 2022, Zhang et al., 15 Apr 2026, Boudellioua et al., 2015, Proverbio et al., 16 Jul 2025).

Taken together, these formulations suggest a common ProtoPathway principle: pathway entities are made semantically stable before downstream inference. In the multimodal survival framework, stable identity is achieved by learned morphological prototypes and curated Reactome/Hallmark pathways. In signaling reconstruction, stability comes from explicit source–target paths of length at most $h=5$. In bond-graph and threshold models, it comes from conservation laws and convex or shortest-path objectives. This suggests a general methodological preference for pathway-level objects whose scores, attentions, or flows can be interpreted directly rather than reconstructed indirectly after prediction.

A second recurring feature is the use of biological or physical constraints to suppress combinatorial ambiguity. The yeast signaling method imposes cell-cycle ordering and protein-complex co-membership. The protocell adaptation models use diffusion and matched opposing actions. The glycolysis study imposes metabolite concentration bounds and enzyme-budget constraints. The chemical-reaction exploration protocol restricts candidate atom pairs by opposite reactivity and uses verified transition states and IRCs. The threshold-sensing model activates links only when local potential differences exceed link-specific thresholds. A plausible implication is that ProtoPathway, across domains, denotes not merely pathway prediction but constrained pathway prediction.

2. Interpretable multimodal cancer survival prediction

"ProtoPathway: Biologically Structured Prototype-Pathway Fusion for Multimodal Cancer Survival Prediction" defines a compact multimodal architecture in which both modalities are represented by biologically meaningful tokens before fusion (Gallagher-Syed et al., 20 May 2026). On the histopathology side, non-overlapping WSI patches are embedded by the frozen UNI-2h foundation model into $D=1536$-dimensional features, then compressed into $K$ learnable morphological prototype tokens by soft assignment,

$$s_{nk} = \operatorname{softmax}_k\!\left(\frac{1}{\tau}\cos(f(h_n), f(c_k))\right),$$

with $\tau=0.1$. Prototype tokens are aggregated into a slide representation by a learned gate $w_k$, while patch-to-prototype assignments provide categorical or soft spatial maps. On the genomic side, bulk RNA-seq is encoded on a curated bipartite graph with $P=662$ pathways, $G=4574$ genes, and $|E_{GP}|=17{,}275$ bidirectional membership edges. GraphSAGE-style message passing and a final GATv2 layer yield pathway embeddings $Z \in \mathbb{R}^{P \times d}$ together with interpretable gene-to-pathway coefficients $D=1536$0.

Fusion is performed by prototype-to-pathway cross-attention rather than patch-level co-attention. With prototype tokens $D=1536$1 and pathway embeddings $D=1536$2, the attention matrix

$D=1536$3

is itself an inference-time attribution because both axes have stable identities. The model explicitly interprets this query direction as matching “the biological direction of molecular programs shaping tissue morphology.” A fused representation concatenates three streams, $D=1536$4, $D=1536$5, and $D=1536$6, and a discrete-time survival head predicts logits for $D=1536$7 quantile bins, with patient-level risk defined by $D=1536$8.

The reported evaluation uses five TCGA cohorts with unified preprocessing and five-fold cross-validation: BRCA ($D=1536$9), BLCA ($K$0), COADREAD ($K$1), HNSC ($K$2), and STAD ($K$3). ProtoPathway attains the highest overall C-index, $K$4, ranking first on BRCA, BLCA, HNSC, and STAD, and second on COADREAD ($K$5 versus $K$6 for SurvPath). It exceeds MCAT ($K$7), SurvPath ($K$8), and MMP ($K$9). Computationally, the model uses $$s_{nk} = \operatorname{softmax}_k\!\left(\frac{1}{\tau}\cos(f(h_n), f(c_k))\right),$$0K parameters, $$s_{nk} = \operatorname{softmax}_k\!\left(\frac{1}{\tau}\cos(f(h_n), f(c_k))\right),$$1G FLOPs per forward pass, $$s_{nk} = \operatorname{softmax}_k\!\left(\frac{1}{\tau}\cos(f(h_n), f(c_k))\right),$$2 MB VRAM, $$s_{nk} = \operatorname{softmax}_k\!\left(\frac{1}{\tau}\cos(f(h_n), f(c_k))\right),$$3 ms/patient for training, and $$s_{nk} = \operatorname{softmax}_k\!\left(\frac{1}{\tau}\cos(f(h_n), f(c_k))\right),$$4 ms/patient for inference, while patch-level cross-attention baselines are substantially slower.

Interpretability is native but not unrestricted. The model provides per-patch pathway heatmaps,

$$s_{nk} = \operatorname{softmax}_k\!\left(\frac{1}{\tau}\cos(f(h_n), f(c_k))\right),$$5

and per-patch gene heatmaps,

$$s_{nk} = \operatorname{softmax}_k\!\left(\frac{1}{\tau}\cos(f(h_n), f(c_k))\right),$$6

yet the paper explicitly notes that attention magnitudes are not expression direction. It also identifies limitations: bulk RNA averages cell populations, pathway coverage depends on Reactome/Hallmark, and domain shifts across scanners or laboratories remain relevant. These caveats are central to understanding the model’s interpretability claims.

3. Signal-transduction and allosteric route reconstruction

In yeast signaling reconstruction, ProtoPathway is formulated as a two-stage inference problem over a weighted undirected PPI graph $$s_{nk} = \operatorname{softmax}_k\!\left(\frac{1}{\tau}\cos(f(h_n), f(c_k))\right),$$7 with $$s_{nk} = \operatorname{softmax}_k\!\left(\frac{1}{\tau}\cos(f(h_n), f(c_k))\right),$$8 proteins and $$s_{nk} = \operatorname{softmax}_k\!\left(\frac{1}{\tau}\cos(f(h_n), f(c_k))\right),$$9 interactions (Inostroza et al., 2020). Candidate pathways are first generated by a directed-edge-based heuristic derived from Gitter et al. for the NP-hard Maximum Edge Orientation problem, with path reliability defined by

$\tau=0.1$0

Sources and targets are receptor-to-effector pairs, and path length is bounded by $\tau=0.1$1, chosen from curated pathway analyses. Candidate paths are then filtered by a cell-cycle consistency model defined on phases

$\tau=0.1$2

with successor relation $\tau=0.1$3, peak-expression and phenotype mappings $\tau=0.1$4 from Cyclebase 3.0, and complex memberships $\tau=0.1$5 from DAPG or CYC2008. A consecutive protein pair is accepted if temporal ordering is consistent under any combination of $\tau=0.1$6 and $\tau=0.1$7, or if the proteins share at least one complex. Finally, complex coverage is summarized by

$\tau=0.1$8

and the path is retained if $\tau=0.1$9.

This construction improves ranked retrieval relative to Gitter alone and to PathLinker-RWR. Using predicted complexes, the best GCC-20/30 configurations achieve $w_k$0–$w_k$1 top-100 true positives versus Gitter’s $w_k$2; with curated CYC2008 complexes, the best result is $w_k$3 versus $w_k$4. The reported MAP is $w_k$5 for GCC, compared with $w_k$6 for Gitter and $w_k$7 for PathLinker-RWR. The method is explicitly limited to activation-like cascades in evaluation: inhibition interactions were removed from the gold standard, and no explicit sign assignment $w_k$8 is learned. A common misconception is therefore excluded by design: the framework orients candidate paths but does not infer inhibitory signs.

At protein structural scale, allosteric ProtoPathway inference uses paths of optimised propensity rather than PPI paths (Wu et al., 2022). The atomistic graph contains covalent and non-covalent interactions weighted by interatomic energies, with bond-to-bond transfer quantified by

$w_k$9

Raw bond propensity from the orthosteric source is normalized, distance-corrected by quantile regression, and converted into quantile scores $P=662$0. These bond scores are lifted to residue-residue edges, and signalling paths are selected between orthosteric and allosteric residues under the constraint of no more than one consecutive backbone hop. Path quality is measured by the geometric mean of weak-edge scores,

$P=662$1

The method identifies experimentally consistent routes in h-Ras, caspase-1, and PDK1. In h-Ras, the highest POP score is $P=662$2, with recurrent roles for Tyr96, Ile100, and HOH367. In caspase-1, the Cys285-to-Glu390 score drops from $P=662$3 in wild type to $P=662$4 in Arg286Ala and $P=662$5 in Glu390Ala. In PDK1, the highest scoring path is Arg131–Glu130–Lys111 with $P=662$6, and modulator-specific average POP scores to Sep241 range from $P=662$7 for 2A2 to $P=662$8 for the inhibitor 1F8. The significance of this line of work is that pathway inference is performed at atomistic resolution while retaining explicit route scores.

4. Adaptation circuits, thermodynamics, and energy-aware pathway physics

A distinct ProtoPathway formulation appears in protocell-inspired models of sensory adaptation that achieve perfect adaptation without internal energy consumption (Palo et al., 2013). The one-component model uses a single transmembrane receptor with one extracellular and one intracellular ligand-binding site; the same small diffusible ligand binds outside and, after passive permeation, binds inside with the opposite effect. With $P=662$9 and $G=4574$0, the free-energy difference and receptor activity are

$G=4574$1

The two-component model instead releases fast subunit $G=4574$2 and slow subunit $G=4574$3 from a receptor; a membrane effector is activated by $G=4574$4 and inhibited by $G=4574$5 with equal and opposite strengths, giving

$G=4574$6

In both cases, diffusion sets the delay structure, and activity returns exactly to baseline when the opposing arms equalize.

The thermodynamic statement is precise. For passive transport through the membrane,

$G=4574$7

which is positive only when $G=4574$8 and vanishes at steady state. The models therefore show that perfect adaptation and approximate fold-change detection can arise in equilibrium, but they also identify the drawbacks of such energy-free adaptation: limited adjustability, poor separation of sensing and signaling, modest sensitivity and dynamic range, and weak spatial control. The paper explicitly presents these drawbacks as reasons why evolution favored energy-consuming eukaryotic pathways.

An energy-aware ProtoPathway perspective is developed further by bond-graph pathway analysis of biomolecular networks (Gawthrop et al., 2016). Here the pathway object is not only a mass-flux mode but also an energy-transduction channel. For species $G=4574$9, chemical potential is

$|E_{GP}|=17{,}275$0

reaction affinity is $|E_{GP}|=17{,}275$1, and per-reaction power is

$|E_{GP}|=17{,}275$2

Given a positive pathway matrix $|E_{GP}|=17{,}275$3 satisfying $|E_{GP}|=17{,}275$4, pathway affinities are aggregated as $|E_{GP}|=17{,}275$5, and pathway power can be written as $|E_{GP}|=17{,}275$6 when $|E_{GP}|=17{,}275$7. The exemplar applications are glycolysis and the Sodium-Glucose Transport Protein 1 cycle, where the method distinguishes productive transduction from dissipation and shows that slip reactions reduce efficiency. This extends pathway analysis from stoichiometric feasibility to energetic accounting.

5. Metabolic and chemical pathway design

In metabolic pathway design, ProtoPathway is used to describe both exhaustive enumeration under biochemical rules and learned retrosynthetic ranking. The lower-glycolysis/gluconeogenesis study constructs all biochemically feasible alternative trunks between glyceraldehyde 3-phosphate and pyruvate by enumerating charged, linear 2–4C CHOP metabolites and reactions from 12 EC classes (Court et al., 2014). This yields $|E_{GP}|=17{,}275$8 internal metabolites, a network of $|E_{GP}|=17{,}275$9 reactions, $Z \in \mathbb{R}^{P \times d}$0 five-step glycolytic trunks, and $Z \in \mathbb{R}^{P \times d}$1 five-step gluconeogenic trunks. Flux comparison uses diffusion-limited reversible kinetics with a fixed enzyme budget and internal concentration bounds of $Z \in \mathbb{R}^{P \times d}$2 nM to $Z \in \mathbb{R}^{P \times d}$3 M. The central result is conditional rather than absolute: the natural glycolytic trunk ranks second on average over sampled parameter space but is the maximal-flux solution in the typical physiological box, whereas some alternatives outperform it only under atypical energy, redox, or pyrophosphate regimes, or when concentration bounds are ignored. A crucial consequence is that high nominal flux is often biologically irrelevant if intermediate concentrations become infeasible.

The 2026 computational framework for multistep metabolic pathway design introduces a deep-learning-guided retrobiosynthesis pipeline that combines template expansion with learned plausibility scoring (Zhang et al., 15 Apr 2026). The assembled training data comprise $Z \in \mathbb{R}^{P \times d}$4 KEGG reactions and $Z \in \mathbb{R}^{P \times d}$5 pathways initially, reduced after cleaning to $Z \in \mathbb{R}^{P \times d}$6 reactions and $Z \in \mathbb{R}^{P \times d}$7 compounds, then expanded to $Z \in \mathbb{R}^{P \times d}$8 mono-product reactions. Backward template libraries include $Z \in \mathbb{R}^{P \times d}$9 BNICE rules and $D=1536$00 RetroRules templates. Two neural classifiers are trained: NN1PR for 1-step plausibility with a 1024-bit concatenated fingerprint and one 256-unit hidden layer, and NN2PR for 2-step plausibility with a 1536-bit input and hidden layers of 512 and 128 units. NN1PR places about $D=1536$01 of positives in the top 10, versus about $D=1536$02 for the Tanimoto baseline, and NN2PR improves top-10 coverage for 2-step ranking by about $D=1536$03 over NN1PR. In the non-natural BDO case, the exact literature route ranks $D=1536$04 with NN1PR only but improves to $D=1536$05 with NN1PR+NN2PR.

Automated chemical ProtoPathway construction is addressed by the Chemoton protocol for context-driven exploration of complex reaction networks (Simm et al., 2017). Starting from initial reagents, the method generates conformers, identifies reactive atoms by conceptual electronic-structure heuristics, assembles intermolecular or intramolecular reactive complexes, performs constrained scans, refines transition states by freezing-string and eigenvector-following procedures, and verifies minimum-energy paths by IRC. In the formose application, exploration was bounded to tetroses and used Q-Chem 4.3 with PBE and a double-$D=1536$06 basis. The reported scale is $D=1536$07 geometry optimizations, $D=1536$08 constrained scans, $D=1536$09 freezing-string calculations, $D=1536$10 transition-state searches, and $D=1536$11 IRCs, producing $D=1536$12 unique molecular configurations and $D=1536$13 minimum-energy paths. Tree traversal then identified $D=1536$14 distinct paths to d-erythrose and ten autocatalytic cycles. The significance of this work is not only automation but verification: routes are retained only after transition-state and IRC confirmation.

6. Annotation rules, threshold optimality, and broader significance

A protein-annotation interpretation of ProtoPathway appears in association-rule mining for prokaryotic metabolic pathway involvement (Boudellioua et al., 2015). Each UniProtKB protein is represented as a transaction containing InterPro signatures, taxonomic lineage nodes, and pathway labels. Apriori mining with minimum confidence of $D=1536$15 produces $D=1536$16 raw rules, which are reduced by SkyRule to $D=1536$17 representative rules and aggregated into $D=1536$18 pathway prediction models covering $D=1536$19 validated pathways. On two independent runs of five-fold cross-validation, the system attains precision $D=1536$20, recall $D=1536$21, F1-score $D=1536$22, and AUC $D=1536$23. Applied to $D=1536$24 TrEMBL prokaryotic reference proteome entries, it annotates $D=1536$25 proteins, including $D=1536$26 that previously lacked pathway annotations. Here, ProtoPathway is best understood as an interpretable, taxon-aware rule layer in which pathway involvement is predicted by explicit antecedents rather than opaque embeddings.

A more abstract ProtoPathway formulation is provided by the threshold-sensing model for path formation in Physarum polycephalum (Proverbio et al., 16 Jul 2025). The environment is represented as a graph $D=1536$27 with node potentials $D=1536$28 and link flows $D=1536$29. Link dynamics combine capacitive storage and thresholded memristive conductance, and the steady-state problem is shown to be equivalent, under ideal thresholds, to minimizing

$D=1536$30

subject to flow conservation $D=1536$31. For oriented links with $D=1536$32, this becomes a linear program minimizing $D=1536$33, and for $D=1536$34 the solution is a shortest path minimizing cumulative threshold cost. The paper’s point is conceptual as much as algorithmic: the mould does not compute a global path explicitly, yet global optimality emerges from local threshold sensing and conservation laws.

These lines of work delimit both the promise and the boundaries of the ProtoPathway idea. The collected evidence suggests that pathway-native representations can improve interpretability, biological plausibility, or computational efficiency when compared with formulations that treat pathways only as downstream annotations. At the same time, every major instance is conditioned by explicit assumptions: stable pathway vocabularies in multimodal oncology, incomplete gold standards in signaling reconstruction, equilibrium and spherical-diffusion assumptions in protocell models, fixed cofactor environments in metabolic trunk analysis, heuristic reactive-site rules in chemical exploration, curated feature availability in rule mining, and threshold idealization in path-formation models. ProtoPathway therefore denotes not a single algorithm but a family of structured pathway formalisms in which the central scientific question is how much mechanistic meaning can be attached to a path once the path itself becomes the computational primitive.