Counterfactual NMR: Benchmarking Minimal Spectral Interventions for Interpretable Structure Elucidation

Susanna Di Vita

Keywords: Counterfactual reasoning, Causal inference, NMR spectroscopy, Automated structure elucidation, Spectral interventions, Explainable AI in chemistry, Causal machine learning, Interpretable molecular models, Structure–spectra relationships

Abstract: NMR-based structure elucidation models are often accurate yet opaque: they rarely indicate \emph{which specific measurements} drove a call or \emph{what to acquire next}. We introduce \emph{Counterfactual NMR}, a causal audit that asks: \emph{What minimal, chemically plausible edit to the spectrum would flip this prediction?} Edits are single-peak interventions constrained to expert ppm windows for $^1$H/$^{13}$C (e.g., carbonyl $^{13}$C 190–220\,ppm); a fast search selects the smallest change that maximally increases a target probability. Effects are quantified by per-sample treatment $\hat{\tau}$, cohort Average Treatment Effect (ATE) with 95\% CIs, and \emph{lift} versus a window-matched randomized baseline to separate specificity from generic sensitivity; mechanistic ablations ($^1$H-only/$^{13}$C-only/both) test alignment with textbook chemistry. On near-boundary cohorts ($0.35{<}p{<}0.65$), minimal $^{13}$C interventions produce large, precise shifts—e.g., \textit{ketone} ATE $0.336$ (95\% CI $[0.315,0.357]$, flip $0.684$) and \textit{alcohol} ATE $0.272$ (95\% CI $[0.261,0.283]$, flip $0.800$)—with targeted effects $2$–$4\times$ stronger than random edits under the same constraints. Ablations confirm chemistry (carbonyl/ketone are $^{13}$C-driven; alcohol shows balanced $^1$H/$^{13}$C$+$synergy); edits remain sparse and realistic. Counterfactual NMR turns interpretability into \emph{actionable recourse}, enabling trustworthy auditing, targeted data curation, and principled next-experiment selection in functional group prediction workflows.

Submission Number: 56