Protein Safety Knowledge Graph

• Protein Safety Knowledge Graph is a semantically typed network that integrates heterogeneous data from biochemical, clinical, and curated evidence sources to quantify protein biosafety.
• It employs a multipartite, directed, and weighted graph schema to enable predictive analytics, hypothesis generation, and transparent safety–efficacy trade-off analyses.
• The framework supports advanced machine learning methods, including reinforcement learning and preference optimization, to reduce toxicity in protein generation while preserving functional integrity.

A Protein Safety Knowledge Graph (PSKG) is a structured, semantically typed network integrating heterogeneous evidence concerning protein function, safety, and risk, with the explicit aim of quantifying or improving biosafety profiles in domains such as drug development, pharmacovigilance, and protein LLM optimization. PSKGs unify disparate data sources—ranging from curated biochemical knowledgebases to clinical and post-market safety signals—within a formal graph architecture. This paradigm enables predictive analytics, hypothesis generation, and stringent safety assurance for both approved biotherapeutics and de novo protein designs (Jackson et al., 17 Feb 2026, Wang et al., 15 Jul 2025).

1. Data Sources, Curation, and Integration

Protein safety knowledge graphs depend on rigorous, multi-source integration. In the context of drug safety (e.g., protein kinase inhibitors), primary data inputs include ChEMBL (drug-target binding data), PubMed/PMC (RCTs, systematic reviews), ClinicalTrials.gov (trial metadata and endpoints), FAERS (post-marketing adverse event reports), and the Small Molecule Suite (experimental affinities). Nodes are standardized, including synonym resolution for drugs, and evidence from literature is mapped to biomedical ontologies such as UMLS conditions and GO terms (Jackson et al., 17 Feb 2026).

For biosafety in protein generation, PSKG construction involves:

• Nodes: “Harmful” proteins (e.g., UniProt “toxin” and “antigen” keywords), “benign” proteins (Swiss-Prot entries not labeled as harmful), and Gene Ontology (GO) terms.
• Edges: Protein–GO annotation links; GO–GO hierarchical relationships.
• Evidence curation: Deduplication, association via UniProt cross-referencing, and direct acquisition of GO hierarchy from the Gene Ontology Consortium (Wang et al., 15 Jul 2025).

2. Formal Graph Schema and Edge Weighting

A multipartite, directed, weighted graph $G = (V, E)$ is instantiated, where $V$ comprises drugs, proteins, targets, adverse events, trials, papers, and for biosafety: P_H (harmful proteins), P_B (benign proteins), and GO terms. Edge types encode mechanistic or evidentiary relationships: drug→target mappings, drug→adverse event disproportionality (e.g. FAERS PRR), protein–GO annotation, and hierarchical GO–GO relations (Jackson et al., 17 Feb 2026, Wang et al., 15 Jul 2025).

Edge weighting is context-dependent. For drug–condition edges, a “paper weight” is computed as:

$$w_p = 1 + \text{normalize}(year_p) + \text{normalize}(1 - bias_p) + \text{normalize}(\log(1 + citations_p)) + \text{normalize}(dataVolume_p)$$

and aggregated over all supporting papers as $w_{d \to c} = \sum_{p} w_p$. Drug–target edges use normalized binding affinities, and Drug→AE edges leverage PRR scores (Jackson et al., 17 Feb 2026). In the biosafety context, no edge weighting beyond presence/absence is performed before pruning (Wang et al., 15 Jul 2025).

3. Knowledge Graph Pruning and Statistical Refinement

Biosafety-focused PSKGs undergo a two-stage weighted-metric pruning to improve computational tractability and focus on the most informative “bridges” between harmful and benign proteins:

• GO-node scoring: $C(g_z) = \gamma R(g_z) + \delta O(g_z)$ ranks GO terms by their potential to link harmful and benign sequences.
• Benign protein scoring: $S(p_B) = \alpha C_{GO}(p_B) + \beta C_{Deg}(p_B)$ prioritizes benign proteins closely connected to the selected GO terms and of high local degree.
• Complexity: Practical implementation reduces $O(|P_H| \cdot |P_B|)$ per GO node scoring to $O(|E|)$ via adjacency lists. Sorting costs $O(|G| \log |G| + |P_B| \log |P_B|)$. The procedure retains $Q$ top GO nodes and $V$0 top benign proteins to define a pruned subgraph (Wang et al., 15 Jul 2025).

For drug safety KGs, target–adverse event (AE) relationships are refined by constructing binary drug–target and drug–AE matrices, calculating Pearson correlation $V$1, $V$2 tests, and partial correlations. Target–AE links’ strengths are thus grounded in data-driven associations across the integrated evidence sources (Jackson et al., 17 Feb 2026).

4. Analytical and Machine Learning Methods

PSKGs facilitate both classical and machine learning analyses:

• Pairwise preferences: In controllable protein generation, knowledge-guided preference optimization (KPO) uses PSKGs to train LLMs to prefer benign over harmful sequences. The DPO-formulated loss is:

$V$3

where $V$4 is the sigmoid function and $V$5 a scaling factor.

• Reinforcement learning (RL) integration: Rewards are based on embedding or graph distances to harmful proteins.
• Community detection and similarity: Modularity-maximization (e.g., Louvain), Adamic-Adar index, and embedding-based clustering are used for subnetwork identification and hypothesis prioritization in both safety and efficacy dimensions (Jackson et al., 17 Feb 2026, Wang et al., 15 Jul 2025).

5. Applications: Drug Safety, Pharmacovigilance, and Protein Generation

Drug Safety and Pharmacovigilance

Application to protein kinase inhibitors demonstrates the PSKG’s capacity to:

• Contextually score and compare drugs for efficacy endpoints (hazard ratio, PFS, OS) and adverse event likelihoods.
• Infer mechanistic target–AE correlations (e.g., B-RAF ⟶ hand-foot syndrome, r ≈ 0.42; ABL ⟶ cytopenias, r ≈ 0.37).
• Predict candidate molecules for specific indications (e.g., non-small cell lung cancer) and discover pharmacologically meaningful target communities (e.g., ErbB, ALK, VEGF).
• Enable transparent joint safety–efficacy trade-off analyses, e.g., plotting medianHR against AE count (Jackson et al., 17 Feb 2026).

Safe Protein Generation

In the context of generative protein PLMs, PSKG-guided KPO robustly reduces the risk of toxic sequence generation:

• BLAST/MMseq2 similarity to harmful test proteins is halved versus baselines.
• Predicted toxicity by ToxinPred3 is reduced by >60%, while mutation-fitness scores (on GB1, PhoQ, UBC9, GFP) are preserved or improved.
• Embedding-space clustering shows clear separation from known harmful sequences post-fine-tuning (Wang et al., 15 Jul 2025).

Representative results over three PLMs (ProtGPT2, ProGen2, InstructProtein) confirm consistent improvements:

Model	BLAST↓	MMseq2↓	ToxinPred3↓
ProtGPT2	0.269	0.325	0.070
ProtGPT2 + KPO	0.138	0.149	0.024
InstructProtein	0.410	0.285	0.031
InstructProtein + KPO	0.086	0.079	0.003

6. Limitations, Extensibility, and Future Directions

PSKG-based methods currently inherit all biases and incompleteness from underlying sources (UniProt, GO, FAERS). Structural safety constraints (binding specificity, off-target immune response) are not directly encoded in sequence-based graphs. Extending schemas to include “Mutation,” “Pathway,” “Cell-line,” or integrating structural, social, or high-dimensional omics data is straightforward within the modular KG framework (Jackson et al., 17 Feb 2026, Wang et al., 15 Jul 2025).

Key limitations include:

• Sequence-level annotation without explicit 3D or immunological risk modeling.
• Static graphs; real-time curation (e.g., new toxins) is under development.
• For protein generation, there is potential for misuse; controlled-access to harmful protein sets is strongly advised.

Extensibility encompasses:

• Additional node and edge types.
• Alternate edge weighting: e.g., antibody–target interactions via surface plasmon resonance K_D, post-marketing safety from sentiment data.
• Swap-in of advanced classifiers for target–AE prediction (logistic regression, random forests), or alternate community detection methods (Infomap, Leiden).
• Incorporation of journal impact, clinical phase, or additional domain-specific features (Jackson et al., 17 Feb 2026, Wang et al., 15 Jul 2025).

7. Significance and Impact

Protein safety knowledge graphs constitute an orthogonal, extensible approach to risk analysis and mitigation. They are not designed as replacements but as advances over single-source, shallow evidence integration. The transparent, modular schema enables reproducibility, extensibility to novel drug classes or engineered proteins, and supports both hypothesis generation and prospective safety assurance. The convergence of PSKGs and preference-optimized learning architectures is establishing new benchmarks for safety-aware AI systems in both clinical and synthetic biology (Jackson et al., 17 Feb 2026, Wang et al., 15 Jul 2025).