paper/sections/benchmark.tex

\section{NegBioBench: Evaluation Framework}
\label{sec:benchmark}

NegBioBench is a dual ML+LLM benchmark designed to evaluate both predictive models and language models on negative biological evidence. We describe the two evaluation tracks and our methodology for measuring genuine understanding versus memorization.

\subsection{ML Track}

The ML track evaluates whether predictive models can distinguish experimentally confirmed negatives from positives, and whether standard evaluation practices inflate reported performance.

\textbf{Tasks.} We define two task types: \emph{M1} (binary classification: negative vs.\ positive) across all three domains, and \emph{M2} (7-way failure category prediction) for CT only. Positive examples come from established sources: DAVIS actives for DTI, CTO successful trials~\citep{siah2021cto} for CT, and HuRI positive interactions~\citep{luck2020huri} for PPI.

\textbf{Splitting strategies.} We implement domain-appropriate cold splits to test generalization: cold\_drug and cold\_target (DTI), cold\_drug and cold\_condition (CT), cold\_protein and cold\_both via METIS graph partitioning~\citep{karypis1998metis} (PPI), plus random, temporal, scaffold~\citep{bemis1996murcko}, and degree-balanced (DDB)~\citep{zheng2020ddb} splits where applicable.

\textbf{Control negatives.} To measure the effect of negative source on model performance (\emph{Experiment~1}), we train identical models on NegBioDB negatives versus two control sets: uniform random pairs and degree-matched random pairs. This directly tests whether curated negatives carry different signal than assumed negatives.

\textbf{Models.} Three architectures per domain: DeepDTA~\citep{ozturk2018deepdta}, GraphDTA~\citep{nguyen2021graphdta}, and DrugBAN~\citep{bai2023drugban} for DTI; XGBoost~\citep{chen2016xgboost}, MLP, and GNN for CT; SiameseCNN, PIPR~\citep{chen2019pipr}, and MLPFeatures for PPI. Metrics include AUROC, LogAUC$_{[0.001,0.1]}$~\citep{mysinger2010logauc}, AUPRC, and MCC~\citep{matthews1975mcc}.

\subsection{LLM Track}

The LLM track evaluates language models across four levels of increasing cognitive demand:

\textbf{L1 (Multiple Choice).} Classification of negative evidence into domain-specific categories (4-way for DTI/PPI, 5-way for CT). Tests whether LLMs can recognize evidence types from textual descriptions.

\textbf{L2 (Extraction).} Structured JSON extraction of key fields from evidence text (compound/target identifiers, assay types, p-values). Tests whether LLMs can parse scientific evidence into machine-readable formats.

\textbf{L3 (Reasoning).} Open-ended scientific reasoning about why a negative result was observed and its implications. Evaluated by an LLM-as-judge on four dimensions: accuracy, completeness, reasoning quality, and specificity.

\textbf{L4 (Discrimination).} Binary classification of whether a given entity pair has been \emph{experimentally tested and found inactive} versus \emph{never tested}. This is the critical level: it tests whether LLMs possess genuine understanding of negative results or merely recall information from training data.

\textbf{Models and configurations.} Five models span the capability spectrum: Llama-3.3-70B~\citep{dubey2024llama3}, Qwen2.5-32B~\citep{yang2024qwen2}, GPT-4o-mini~\citep{openai2024gpt4o}, Gemini-2.5-Flash~\citep{google2025gemini}, and Claude Haiku-4.5~\citep{anthropic2025claude}. Each is evaluated in zero-shot and three few-shot configurations (different random example sets), yielding 4 configurations per model--level pair.

\subsection{Evaluation Methodology}

NegBioBench makes three methodological contributions relevant to the evaluation landscape:

\textbf{L4 as a contamination probe.} L4 discrimination performance reveals whether models have memorized negative results from training data. We incorporate temporal stratification (pre-2015 vs.\ post-2020 publication dates) to detect contamination: a performance gap exceeding 0.15 indicates likely memorization rather than reasoning~\citep{sainz2024contamination}.

\textbf{Cross-domain comparison.} By applying identical evaluation levels across three domains with different data accessibility profiles, we can isolate the effect of training data composition on LLM performance---independent of biological task difficulty.

\textbf{Anti-contamination design.} L1--L3 use paraphrased evidence text to reduce verbatim memorization advantage. L4 temporal holdouts ensure post-training-cutoff examples are included. All hallucinated evidence citations are tracked to measure confabulation rates.