Building an Explainable Toxicity Knowledge Graph

22 Sep 2026 — LynxKite Team & Hummingbird Bioscience

Developed in collaboration with LynxKite, supported by NVIDIA accelerated computing.

1. Framing the toxicity problem

Evaluating toxicity is a multi-faceted problem and highly context dependent. Rather than asking “is this molecule toxic?” we need to ask: what component causes the effect, through what mechanism, in which tissue, at what exposure, and in what biological context?

A practical map of major toxicity mechanisms

The categories below are deliberately broad and overlapping. They are useful to us as framing because they separate mechanistic questions that require different evidence and different validation strategies.

Mechanistic class	Core question	Examples of relevant evidence
On-target, off-tissue toxicity	The intended target is engaged, but in healthy cells or tissues where that pharmacology is harmful.	Target expression in normal tissue; exaggerated pharmacology; for oncology, often called on-target/off-tumor toxicity.
Off-target molecular pharmacology	The drug, payload, metabolite, antibody, or other component interacts with an unintended molecular target.	Receptor/enzyme cross-reactivity, reactive chemistry, or an unexpected protein interaction, poly-specificity.
Target-independent uptake or distribution	A drug reaches or enters a normal cell without the intended target-mediated delivery mechanism.	Nonspecific endocytosis, Fc- or lectin-receptor uptake, tissue trapping, transporter effects, or altered biodistribution.
Misplaced release or exposure	A toxic species is generated in the wrong compartment or at the wrong time.	Premature prodrug or linker cleavage, deconjugation, active metabolites, systemic free payload, or tissue accumulation due to altered drug clearance pathways.
Immune- and modality-mediated toxicity	The therapeutic format itself drives immune or platform effects beyond the intended pharmacology.	Cytokine release, infusion reactions, immunogenicity, Fc-mediated effects, aggregation, or complement activation.
Context-dependent susceptibility	The same molecular event produces different outcomes depending on the biological setting.	Species, tissue/cell state, genotype, immune status, disease state, organ function, dose, schedule, or concomitant therapy.

Why antibody–drug conjugates are a useful stress test

An antibody–drug conjugate (ADC) is a three-part therapeutic: a monoclonal antibody is chemically connected through a linker to a highly potent small-molecule payload. In the intended mechanism, the antibody binds a surface antigen on a cancer cell, the ADC is internalized, and the payload is released to kill the cell. The purpose is to widen the therapeutic index by concentrating a potent payload in tumor cells rather than normal tissue.

ADCs are particularly informative for a toxicity knowledge graph because they combine three coupled design systems — biologic targeting, linker chemistry, and payload chemistry — and can therefore fail through several routes at once. Toxicity can reflect normal-tissue target expression, target-independent uptake of intact ADC, premature or ectopic linker cleavage, free-payload exposure, payload-class pharmacology, or interactions between these mechanisms. Clinical experience also shows that ADCs using the same linker/payload class can share dose-limiting toxicities even when they target different antigens. [1]

This range of possible different toxicity mechanisms is helpful to stress test the knowledge graph design. Are we able to capture the relevant components? They are not meant to imply that the project is limited to ADCs; they provide tractable examples in which biological, chemical, clinical, and mechanistic evidence must be integrated.

The data problem is a context problem as much as a scale problem

Relevant evidence is distributed across regulatory documents and scientific literature and includes but is not limited to:

• FDA/EMEA and other regulators structured product labels and prescribing information
• Drugs@FDA pharmacology/toxicology and multidisciplinary reviews
• FAERS and other post-marketing adverse-event data
• ClinicalTrials.gov records, including terminated or halted studies
• primary literature, supplementary data, and patent filings
• company disclosures, annual reports, and other long-tail sources that may record safety-driven program changes

A toxicity observation is only interpretable together with its context. A claim such as “compound X causes hepatotoxicity” is incomplete unless we can recover, where available, the species, tissue or cell type, genotype or phenotype, immune context, dose or concentration, route, exposure duration, assay, disease state, and relevant metabolites. The same applies to biological relationships such as protein–protein interactions: a relation observed in liver cannot automatically be treated as the same relation in heart tissue or in a different cell state.

Initially, we want to evaluate: can we automate evidence collation and curation at useful scale while preserving the provenance and context that make expert-curated evidence relevant?

2. The evidence-to-knowledge-graph pipeline

The architecture has four linked functions: finding relevant evidence, extracting structured claims, representing context and provenance, and using the resulting graph to drive new retrieval or hypothesis generation.

Document acquisition. A run starts from seed articles, molecules, targets, mechanisms, or a researcher question. A query-generation layer expands these into searches across source connectors such as DailyMed, Drugs@FDA, ClinicalTrials.gov, and literature repositories. Retrieval is iterative: entities and mechanisms found in one document become candidates for the next search. The goal is not to crawl indiscriminately, but to follow evidence while keeping source selection auditable.

Parsing and claim extraction. NVIDIA NeMo™ Retriever is used for document parsing, including long PDF regulatory reviews. Named-entity recognition and NVIDIA Nemotron™ models generate schema-constrained claims rather than free-form summaries. A toxicity claim can include the therapeutic or component, structure identifier, intended and unintended targets, species, tissue/cell type, genotype or phenotype, dose/concentration, route, time point, observed phenotype, study type, source passage, and extraction confidence.

Context-first graph representation. The graph does not collapse a context-dependent observation into a single universal edge. Instead, the evidence assertion itself is represented with qualifiers. A protein interaction observed in a particular tissue, for example, remains linked to that tissue and experimental setting. Source metadata — authors, institutions, journal or regulatory source, study design — supports provenance and future evidence weighting, but it does not substitute for biological context.

Article and molecule graphs. Article-level nodes capture drugs, molecules, proteins, pathways, diseases, toxicity phenotypes, studies, and experimental contexts. Molecules are canonicalized using structure identifiers such as SMILES or InChI. Relationships are treated as claims supported by evidence, not as “proofs”: positive evidence, negative evidence, conflicting evidence, and uncertainty can coexist, and each assertion points back to the source passage from which it was derived.

Retrieval and iterative discovery. GraphRAG-style retrieval can answer questions from connected evidence rather than a single passage. Separately, graph embeddings, graph neural networks, rules, or other ranking models can be evaluated for downstream tasks such as prioritizing missing links, candidate mechanisms, or documents worth retrieving next. The model makes the prediction or ranking; the graph is the evidence substrate on which it operates.

3. From evidence graph to testable hypotheses

The graph should separate observed evidence from model-derived evidence. A clinical adverse event, a biochemical assay, an animal finding, and an in silico prediction are not equivalent, even if they point to the same mechanism. Each should be labeled by evidence type, context, method, and uncertainty.

The computational layer can then generate or prioritize hypotheses in several ways:

• target and tissue-expression analysis to assess on-target/off-tissue risk
• structural modeling, docking, and molecular dynamics to test the plausibility of specific molecular-interaction hypotheses
• protease-substrate and linker-cleavage models for release mechanisms
• PK, PBPK, or exposure models for tissue distribution, free-payload exposure, and species translation
• pathway or network models for downstream consequences of target perturbation
• graph-based ranking models that identify mechanisms or evidence gaps that are most informative to test next

The important boundary is that these methods prioritize evidence; they do not create ground truth. A docking pose does not establish binding, a docking score is not a reliable affinity measurement, and a predicted protease complex does not establish cleavage kinetics. Computational evidence is most useful when it is comparative, calibrated against suitable positive and negative controls, and designed to decide what should be tested experimentally next.

NVIDIA cuGraph accelerates the graph operations and graph-model workflows, while NVIDIA BioNeMo™ and related structural tools support selected molecular hypotheses. The exact predictive models should remain task-specific and should only be described as “toxicity prediction” once prospective performance has been measured on a clearly defined endpoint.

4. Toxicity Graph in action: two ADC case studies

The two examples below were selected because they expose different parts of the ADC toxicity problem. The first links a clinical liver signal to a reported payload-mediated protein interaction. The second links a clinical hematologic toxicity to a class-level linker-cleavage mechanism. In both cases, the value is not that the graph “discovers” a fact that no human or LLM could find; it is that the system can represent the relationship, its context, its provenance, and allows us to preserve the distinction between observation and inference in a reusable form.

T-DM1 and CKAP5: a payload-mediated hepatotoxicity hypothesis

What the graph connects. In the EMILIA phase III study of trastuzumab emtansine (T-DM1), the most common grade 3 or higher adverse events included thrombocytopenia (12.9%), increased AST (4.3%), and increased ALT (2.9%) — suggestive of liver damage/hepatotoxicity. [2] Separately, Endo and colleagues reported that the DM1-containing T-DM1 construct binds cell-surface CKAP5 on hepatocytes through the DM1 component and can induce membrane damage, calcium influx, microtubule disruption, and apoptosis in vitro. [3]

The correct graph-level conclusion is deliberately narrower than the current draft. These sources support a mechanistic hypothesis linking a clinically observed liver signal to an experimentally reported, HER2-independent DM1/CKAP5 interaction. They do not by themselves prove that CKAP5 is the dominant cause of the clinical transaminase elevations. The graph should preserve that distinction.

What we can model. Structural modeling could be used as an orthogonal plausibility check: identify credible CKAP5 structural regions or binding sites, examine whether DM1 produces reproducible poses across methods, and test whether candidate complexes remain physically plausible during molecular dynamics. Comparisons with matched negative-control proteins and alternative payloads would be essential. Because docking algorithms often generate apparently plausible complexes for non-binders, a positive computational result should rank the hypothesis for experimental follow-up rather than be presented as evidence of affinity or specificity. One way to probe this further in silico would be to carry out molecular dynamic (MD) simulations of the docked complexes to assess their stability in an aqueous environment.

Brentuximab vedotin and the valine–citrulline linker: a cleavage-mediated neutropenia hypothesis

What the graph connects. Brentuximab vedotin is a CD30-directed ADC carrying MMAE through a protease-cleavable valine–citrulline linker. Neutropenia is a clinically observed toxicity; in a pivotal phase II study in relapsed or refractory systemic anaplastic large-cell lymphoma, grade 3/4 neutropenia occurred in 21% of patients. [4] Separately, Zhao and colleagues showed in a human neutrophil-differentiation system that valine–citrulline–MMAE ADCs could undergo extracellular cleavage in the presence of differentiating neutrophils or their conditioned medium, with serine proteases implicated and a purified neutrophil elastase capable of releasing MMAE. [5]

This supports a class-level mechanism in which a linker intended to release payload after target-cell internalization can also be cleaved in an unintended extracellular compartment. It is important not to overstate the join: the Zhao study establishes a plausible mechanism for vcMMAE ADC-associated neutropenia, not definitive proof that neutrophil elastase is the sole or dominant cause of neutropenia in patients receiving brentuximab vedotin.

What we can model. Protease cleavage is a catalytic event, not simply a binding event. A useful in silico workflow would combine known substrate preferences, protease–linker structural modeling, and molecular dynamics focused on whether catalytically competent geometries are accessible. The result should be a ranked set of proteases or linker variants for biochemical cleavage assays. Similar docking scores between cathepsin B and neutrophil elastase would not, on their own, justify a claim of equivalent cleavage or lack of selectivity. Comparable binding of the VC-PABC motif to cathepsin B and neutrophil elastase in cleavage-amenable conformations (vicinity to catalytic residues) and stability of their docked complexes in MD, along with sufficient negative controls (e.g. Cathepsin D and Alpha chymotrypsin), would however support the GNN-inferred and in-vitro-suggested node interactions.

What the case studies demonstrate

Our PoC demonstrates that we are able to:

• capture a clinically or experimentally observed toxicity phenotype
• connect it to a molecule, component, target, pathway, or release mechanism across independent sources
• retain the biological and experimental context for every assertion
• state explicitly which links are observations and which are hypotheses/predictions
• use computational methods to rank the most informative hypotheses or molecule variants
• validate the prioritized hypotheses with biochemical, cellular, or in vivo experiments

In terms of impact, we achieved a 5x expansion in structured toxicity evidence coverage inferred from the available literature after extraction. Over time, as the graph accumulates high-quality, context-rich observations and negative evidence, it can become training data for better-defined predictive tasks. The expansion is driven by newly identified articles containing compound-level information. However, substantial work remains to convert this literature into a reliable knowledge graph in which entities are represented as nodes and evidence-supported claims as edges.

In our initial benchmark, Nemotron 3.5 Lightning achieved 13% precision in extracting correct relationship claims. NVIDIA Nemotron 3.5 Lightning is a customizable open model for always-on agents to complete specialized tasks faster with leading domain-specific accuracy across local, on-premises, and cloud deployments. Applying multiple rounds of agentic validation with Nemotron 3.5 Lightning increased precision to 35%. While this represents a meaningful improvement, the remaining errors illustrate why rigorous validation, and experimentation with various model variants, are still required. Examples include:

• Incorrect relationship direction: myeloma cells -inhibits-> HSP90. HSP90 may be inhibited in myeloma cells, but the cells themselves do not inhibit HSP90.
• Incorrect relationship between correctly identified entities: ADC -cleaved_by-> BTG. BTG is used to conjugate the ADC; it does not cleave it.
• Biologically correct but unsupported claim: lintuzumab -binds_to-> CD33. Although lintuzumab binds to CD33, the cited passage discusses a pharmacokinetic model and does not provide evidence for this binding relationship.

5. What the first phase has taught us

The first phase was run on LynxKite deployed on Nebius with eight parallel NVIDIA H100 80 GB GPUs. The NVIDIA-Nemotron™-3.5-Lightning-30B-A3B NIM™ microservice packages NVIDIA-Nemotron™-3.5-Lightning-30B-A3B, a large language model trained by NVIDIA, as an NVIDIA inference microservice. NVIDIA NIM™, part of NVIDIA AI Enterprise, is a set of easy-to-use microservices designed for secure, reliable deployment of high-performance AI model inferencing across clouds, data centers, and workstations.

At pilot scale, the graph contained on the order of hundreds of thousands of nodes and 6 edges per node on average, with scale tests extending into the millions of nodes. Three practical lessons have emerged:

1. Reading at scale is not the bottleneck. Our extractor doesn’t read a passage once. It proposes claims — subject, relation, object, and the sentence that supports each one — then checks every claim automatically: is that quote really in the text, are those real biological entities, is that relation one we allow? Anything that fails goes back to the model to repair. The loop runs until nothing is left to fix, or five rounds, whichever comes first. That is five times the work of a single pass, and it still isn’t what limits us. One H100, NVIDIA Nemotron 3.5 Lightning 30B-A3B, sustains about two passages a second at that setting — roughly 144 GPU-hours per million passages, or around 18 hours of wall clock on our eight-GPU pilot allocation.
2. Finding relevant evidence is harder than parsing it. Generic ingestion tools can process large document sets, but toxicity evidence is sparse, heterogeneous, and often described indirectly. Much of the engineering effort therefore went into iterative discovery: using compounds, targets, organs, adverse events, mechanisms, and cited studies found in one source to guide the next retrieval step, while controlling duplication and topic drift.
3. Schema-first extraction is materially easier to reconcile. Constraining evidence generation to the curation schema at extraction time produces more consistent entities and relationships than extracting free-form text and attempting to normalize everything afterward. The remaining challenge is not only entity normalization but representing context without flattening biologically distinct observations.
4. From months of model setup to weeks of actionable hypotheses. By orchestrating NVIDIA NIM™ for AlphaFold2, ESMFold2, NVIDIA NIM™ for Boltz-2 and docking workflows across the same compounds and protein regions, we can systematically compare alternative structural and binding hypotheses rather than relying on a single model or docking result. For CKAP5, this allowed us to evaluate the same compound set across multiple regions, compare predicted binding locations and relative ligand rankings, and identify where independent methods reinforce, or challenge, one another. The result is a much faster way to narrow a large hypothesis space to the most plausible compound–region interactions for deeper investigation. Candidates emerging from this multi-model screening can then be prioritized for molecular dynamics.

What previously could require months of assembling, configuring and connecting separate computational workflows can increasingly be executed in weeks as a repeatable pipeline using LynxKite and the NVIDIA accelerated AI stack. For toxicity research, this means generating and testing substantially more hypotheses before committing to expensive experimental validation.

6. Working together

Hummingbird Bioscience brings the domain knowledge, AI and structural modelling, internal datasets, validation data, and biological interpretation. LynxKite contributes graph software, data curation, graph analytics, model development, and user-facing tooling.

The NVIDIA Inception program for startups provides guidance through the NVIDIA platform and ecosystem, which includes technical guidance and an accelerated computing and open-model stack that includes NVIDIA Nemotron 3 and NVIDIA Nemotron 3.5 Lightning, NeMo™ Retriever, NeMo™ Curator, BioNeMo™, and cuGraph.

The longer-term ambition is to determine whether a sufficiently large, context-rich, provenance-aware toxicity evidence base can support specialized predictive models and reusable toxicity representations. That is an empirical question. The immediate milestone is more concrete: build an auditable system that makes fragmented toxicity evidence easier to find, connect, challenge, and test.

Selected references

[1] Nguyen TD, Bordeau BM, Balthasar JP. Mechanisms of ADC Toxicity and Strategies to Increase ADC Tolerability. Cancers (Basel). 2023;15(3):713. doi:10.3390/cancers15030713.

[2] Verma S, et al.; EMILIA Study Group. Trastuzumab Emtansine for HER2-Positive Advanced Breast Cancer. N Engl J Med. 2012;367:1783–1791. doi:10.1056/NEJMoa1209124.

[3] Endo Y, et al. Payload of T-DM1 binds to cell surface cytoskeleton-associated protein 5 to mediate cytotoxicity of hepatocytes. Oncotarget. 2018;9:37200–37215. doi:10.18632/oncotarget.26461.

[4] Pro B, et al. Brentuximab vedotin (SGN-35) in patients with relapsed or refractory systemic anaplastic large-cell lymphoma: results of a phase II study. J Clin Oncol. 2012;30:2190–2196. PMID:22614995.

[5] Zhao H, et al. A Potential Mechanism for ADC-Induced Neutropenia: Role of Neutrophils in Their Own Demise. Mol Cancer Ther. 2017;16:1866–1876. doi:10.1158/1535-7163.MCT-17-0133.

---

This article was originally published by Hummingbird Bioscience and is reproduced here with their approval.