Antibody as Agent: Graded Proto-Roles for Fc-Effector Text

When an immunology paper says "the antibody killed the tumor cells" versus "tumor regression was observed," the underlying experiment may be identical, but the causal story a reader — or a machine — extracts is very different. The first sentence names an instigator and a mechanism; the second names nothing. Linguists call this perspectivization: syntax and word choice systematically foreground or bury who did what, without changing the facts. A small literature now measures this computationally, showing that language models systematically misread agency — for example treating the subject of "the window broke" as a volitional doer — and that iterative machine processing of text actively deletes agents and converts actions into abstract nouns. Meanwhile, immunology has its own agency problem: therapeutic antibodies are engineered precisely to do things — recruit natural killer cells through Fc gamma receptors, drive phagocytosis, deplete regulatory T cells — and whether a text preserves or erases that Fc-mediated mechanism determines whether a reader can tell passive regression from actively engineered cytotoxicity.

No prior work connects these two fields. But the connection is not arbitrary: the foundational paper on property-based semantic roles explicitly raised "a human vs. an antibody" as the canonical case where a categorical Agent label breaks down, and then never tested it Semantic Proto-Roles. This proposal takes that untested example as its subject. The core deliverable is a graded model of molecular agency: an annotation scheme, a benchmark corpus of Fc-effector text, and a set of measurements establishing (i) whether current models can represent a non-sentient instigator, and (ii) whether automated summarization and translation pipelines — increasingly the machinery of scientific synthesis — systematically strip Fc-mechanism causality out of the literature.

Core Claim

The claim has two parts. First, categorical semantic roles (Agent, Patient) cannot represent Fc-mediated effector claims, but Dowty-style graded properties — instigation, causation, change of state, and crucially absence of volition and sentience — can. An antibody in "the antibody depleted the Tregs" is a maximal instigator and minimal volition entity; no categorical label captures that, which is exactly why the 1991-inspired annotation work flagged it Semantic Proto-Roles and why large language models, which already mislabel unaccusative subjects as Agents on general English benchmarks On the Role of Semantic Proto-roles in Semantic Analysis: What do LLMs know about agency?, should fail even harder on "the Tregs depleted," where the missing instigator is a designed Fc function. Second, when LLM pipelines rewrite or summarize immunology text, the Fc-mechanism attribution is the part that dies first — collapsing "FcγRIIIa engagement drove NK-mediated lysis" into "responses were observed" — with measurable consequences for downstream automated evidence synthesis.

Research Program

Three components, in increasing order of consequence.

The first is a benchmark corpus. We will build an Fc-Effector Proto-Role corpus: roughly 3,000–5,000 sentences drawn from abstracts and results sections of the Fc-engineering and receptor-targeting literature — the vaccine-vector and checkpoint-inhibitor papers here are natural seed sources, spanning Fc-competent versus Fc-silent designs (e.g. the anti-TIGIT antibody where the wild-type Fc, not the Fc-silent variant, carried the efficacy An Fc-Competent Anti-Human TIGIT Blocking Antibody Ociperlimab (BGB-A1217) Elicits Strong Immune Responses and Potent Anti-Tumor Efficacy in Pre-Clinical Models), receptor-targeted delivery (Engineered virus-hunter vaccine overcomes HBV immune tolerance), and NK engagers that bypass Fc entirely (Derivation and Preclinical Characterization of CYT-303, a Novel NKp46-NK Cell Engager Targeting GPC3). Each sentence gets two annotation layers. Layer one follows the established protocol of Likert-scale (1–5) ratings over proto-role properties per argument, so our data is directly comparable to the SPR1/SPR2 tradition Semantic Proto-Roles. Layer two is new: a mechanism-attribution label per stated outcome — which effector process (ADCC, ADCP, trogocytosis, complement, cross-presentation, none), whether it is Fc-dependent or Fc-bypassing, and crucially whether the sentence attributes it or leaves it unattributed ("depletion was observed"). Agentless passives and nominalized outcomes get explicit "unattributed" gold labels rather than being treated as annotation noise, because they are the phenomenon.

The second component is model evaluation on this corpus. We test three families: masked-language-model probes of the kind that mapped BERT representations onto Dowty properties with partial least squares regression Does BERT Recognize an Agent? Modeling Dowty’s Proto-Roles with Contextual Embeddings; dependency-parse-plus-lexicon pipelines in the RIVETER mold Riveter: Measuring Power and Social Dynamics Between Entities; and prompted frontier LLMs, where the prior evidence is cautionary — adding explicit proto-role instructions to GPT-4o decreased its role-labeling accuracy, a documented "complexity tax," and all tested models fell into the unaccusative trap On the Role of Semantic Proto-roles in Semantic Analysis: What do LLMs know about agency?. Our prediction is that the unaccusative trap becomes a mechanism trap in this register: models will label "the Tregs depleted" with an implicit self-agent rather than recognizing a deleted Fc-mediated instigator, and will rate antibodies as more volitional than their gold (floor) ratings.

The third component is the degradation and consequence study. Following the method that ran 600 BBC articles through 12 iterations of LLM back-translation across 17 model/language configurations and found factual retention falling to 57% (FActScore), nominalization density rising 21%, and with-agent passives disproportionately collapsing into agentless ones The Broken Telephone Changes Tone: Examining Nuanced Linguistic Cues in LLM Chains-of-Translation, we run the same loop design on our corpus. Whether the hedged, hedging-dense register of immunology is more or less vulnerable than journalism is a genuinely open question — we test it rather than assume it. We then measure the downstream cost: original versus iterated texts are fed to a standard claim extractor in the span-oriented information-extraction tradition Span-Oriented Information Extraction - A Unifying Perspective on Information Extraction, and we score whether the extracted claims still distinguish Fc-effector-driven protection from neutralization-driven protection. This is not a hypothetical concern: in the HVTN 505 vaccine trial, the protective signal lived specifically in Fc-mediated properties — antibody-dependent cellular phagocytosis with odds ratio 0.47 and anti-Env IgG3 breadth with odds ratio 0.326 — while Env-specific IgA actively spoiled that protection Antibody Fc effector functions and IgG3 associate with decreased HIV-1 risk. A synthesis pipeline that strips mechanism conflates exactly the distinctions that trial turned on.

We define one quantity precisely: the mechanism survival rate after $k$ processing iterations,

$$S_k = \frac{\text{number of gold mechanism attributions still recoverable after } k \text{ iterations}}{\text{number of gold mechanism attributions in the source text}},$$

where "recoverable" means a reader-model given only the processed text can reproduce the mechanism-attribution label. $S_k$ decoupling from overall factuality (which the prior work tracked at 57% after 12 iterations) would show that mechanism is lost faster than facts — the most damaging possible outcome for automated meta-analysis.

A smaller fourth component tests restoration: rewriting agentless outcome statements back into mechanism-explicit form, scored by a Harmonic Mean of mechanism-explicitness and factual similarity to the source, directly following the femicide-news perspective-transfer setup where GPT-3 iterative few-shot achieved a Harmonic Mean of 4.98 with similarity 7.97, and where aggressive shifting traded off against fidelity Responsibility Perspective Transfer for Italian Femicide News. In our setting fidelity is not just similarity but biochemical correctness — restoring "ADCC" where the source never established it is hallucination, and the metric must punish it.

Baselines

The naive reference point is a curated keyword-and-regex lexicon of effector-function terms (ADCC, ADCP, FcγR, trogocytosis, D265A, phagocytosis, depletion) run over the same sentences. It will detect explicit mechanism mentions but is definitionally blind to unattributed outcomes and implicit framing; its recall on the unattributed half of our layer-two labels is expected near zero, and beating it means substantially higher mechanism-attribution F1 on sentences lacking any lexicon term.

The strongest pipeline competitor is a RIVETER-style system: spaCy dependency parsing plus coreference plus connotation-frame verb lexicons Riveter: Measuring Power and Social Dynamics Between Entities. It is the strongest existing open tool for scoring entity power and agency — on social dyads it cleanly separated high-power from low-power roles (mean power 0.265 versus 0.0364, p < 0.05) — and it generalizes to arbitrary entity types. To be fair, we will run it on identical input sentences and allow its verb lexicons to be extended from our training split, giving it the same supervision every other system gets; its known weakness, static lexicons that fail when verbs fall outside social domains, is exactly what our corpus stresses, since its lexicons contain nothing for opsonize, cross-present, or trogocytose.

The strongest overall competitor is a prompted frontier LLM (GPT-4o class) given the annotation instructions and few-shot examples, evaluated zero-shot and few-shot. This is the system the prior evaluation work actually tested on general English, where it showed both the complexity tax and the unaccusative trap; no reported result found on Lacuna for its raw SPR1 or SPR2 accuracy, so we will establish that number on our corpus ourselves. We will give it the same sentences, the same few-shot budget, and the same decoding settings as our trained models. Beating it means higher token-level Spearman correlation with gold property ratings and higher mechanism-attribution F1 under matched conditions — and, separately, demonstrating whether its complexity tax replicates in the biomedical register.

The key ablation isolates the domain ingredient: the same property-prediction model trained only on general-domain SPR1/SPR2-style data versus trained with our Fc-effector corpus added. The BERT-probe work provides the reference technique and a general-domain reference point of Spearman's ρ = 0.74 for type-level property prediction from BERT layers 13–16 Does BERT Recognize an Agent? Modeling Dowty’s Proto-Roles with Contextual Embeddings; if general-domain training alone suffices, the gap is in the annotation, and if not, the gap is in the register — the ablation distinguishes these. A second ablation replaces graded properties with binary Agent/Patient labels to confirm that the graded representation, not merely extra training data, carries the improvement, echoing the finding that roughly 800 distinct property configurations exist in the original data Semantic Proto-Roles.

Why not just…?

Why not just use keyword matching? The lexicon-free-metric literature exists precisely because lexicons miss implicit framing: AnthroScore was built as a masked-language-model metric because keyword approaches cannot capture context-dependent anthropomorphism AnthroScore: A Computational Linguistic Measure of Anthropomorphism, and the RIVETER authors themselves flag static lexicon coverage as the system's bottleneck. On our corpus the failure is structural: an unattributed outcome ("clearance was achieved") contains no mechanism keyword by construction, so recall there is zero by definition. We still run the keyword baseline, because if it ties on overall F1 — possible if most corpus sentences do name their mechanisms explicitly — the graded-property machinery is unjustified, and we would report that honestly.

Why not just use off-the-shelf semantic role labeling? PropBank-style ARG0/ARG1 labeling is the standard extraction answer, but the categorical scheme is what created the problem: the original proto-role work showed monolithic Agent labels erase distinctions the antibody case depends on, with nearly 800 distinct property configurations collapsing into a handful of roles Semantic Proto-Roles, and the LLM evaluation showed models defaulting to ARG0 for unaccusative subjects On the Role of Semantic Proto-roles in Semantic Analysis: What do LLMs know about agency?. "Tregs depleted" will get Tregs as ARG0 whether or not an antibody did the depleting. The decisive experiment is direct: we run a standard SRL configuration over our corpus and measure how often its ARG0 coincides with the graded instigator versus the graded patient — the mismatch rate on unaccusative and agentless-passive sentences is the number that settles it.

Why not just prompt GPT-4o to do the whole analysis in one shot? This is the most serious alternative, and we treat it as the primary baseline rather than arguing around it. The evidence against it is the complexity tax — GPT-4o's role-labeling accuracy dropped when proto-role properties were demanded alongside role labels in structured output On the Role of Semantic Proto-roles in Semantic Analysis: What do LLMs know about agency? — plus documented hallucination risk when generative models rewrite factual text under perspective pressure Responsibility Perspective Transfer for Italian Femicide News. But the honest position is that the comparison has never been run on biomedical text, and a frontier model might well win. Our contribution then stands on the benchmark itself: the corpus, gold labels, and mechanism-survival measurements are valuable regardless of which system wins, and the head-to-head under matched budgets is the experiment that produces the answer.

Decisive Checks

Four results would make the case. First, on the corpus: an ablation showing graded properties beat categorical labels on token-level property correlation and mechanism-attribution F1 by a margin that survives the same-data, same-budget comparison against the prompted frontier LLM — with the specific signature that the win concentrates on unaccusative and agentless-passive sentences. Second, a documented failure rate for existing systems: RIVETER-derived and SRL baselines misassigning instigation on the effector sentences, quantified per sentence type. Third, the degradation result: a measured mechanism survival rate $S_k$ after 12 iterations, with the key comparison being whether $S_k$ falls below the general-factuality retention benchmark of 57% established on news text The Broken Telephone Changes Tone: Examining Nuanced Linguistic Cues in LLM Chains-of-Translation — if mechanism dies faster than facts, automated synthesis is selectively blind to exactly the causal content that distinguishes, say, an Fc-silent from an Fc-competent design, a distinction that in the checkpoint-inhibitor setting was the whole efficacy story An Fc-Competent Anti-Human TIGIT Blocking Antibody Ociperlimab (BGB-A1217) Elicits Strong Immune Responses and Potent Anti-Tumor Efficacy in Pre-Clinical Models. Fourth, the consequence check: a claim extractor fed iterated versus original HVTN-505-style text losing the ADCP/IgG3 protective attribution Antibody Fc effector functions and IgG3 associate with decreased HIV-1 risk at a rate that scales with measured mechanism erosion. If all four hold, the field gains both a reusable benchmark for molecular agency and a concrete, quantified warning about what LLM-mediated synthesis does to mechanistic causality in the biomedical literature.