Literature-Derived Priors

• Literature-derived priors are probabilistic models that use published research findings to set Bayesian hyperparameters, offering data-efficient regularization.
• They aggregate evidence through meta-analysis, language model parsing, and standardized fact extraction to construct context-aware priors.
• Empirical applications demonstrate their ability to improve predictive performance and robustness in small-sample or data-sparse environments.

Literature-derived priors are probabilistic models or prior distributions for Bayesian inference whose hyperparameters are calibrated using results, statistics, or facts extracted from published research rather than directly from the data at hand. These priors encode domain knowledge, empirical findings, and experimental constraints distilled from the academic literature and, when carefully constructed, offer principled, data-efficient regularization, improved model interpretability, and enhanced predictive performance, especially in small data regimes or domains with sparse or costly historical data. They find formal instantiations in survey methodology, generalized linear models, symbolic regression, therapeutic design, and elsewhere, leveraging systematic meta-analyses, LLM parsing, and information-theoretic criteria to aggregate evidence and construct priors with explicit operational guarantees.

1. Formal Construction Paradigms

The construction of literature-derived priors follows several prototypical workflows, shaped by the inferential context and structural properties of the underlying model:

• Meta-analytic pooling for parametric priors: In Bayesian regression for daily response propensity, each coefficient $\beta_c$ receives an independent normal prior $$N(\mu_c^{(\text{lit})},\, \sigma_c^2{}^{(\text{lit})})$$ whose mean and variance are computed as the average and average variance of log-odds (or rescaled probit) coefficients reported in published studies. Covariances between coefficients are set to zero, with literature coverage completeness determining the prior informativeness. When no literature exists for a coefficient, a weakly informative baseline prior is substituted (West et al., 2019).
• Knowledge extraction from unstructured text: In therapeutic design, pipelines such as Medex employ large-scale retrieval and tagging of scientific paragraphs, normalization of chemical and protein entities (e.g., SMILES or RefSeq IDs), and fact extraction using fine-tuned LLMs. This produces entity–fact pairs that can be leveraged to pretrain property prediction networks or LLM adapters, with literature-derived "fact priors" acting as explicit regularizers or constraints in generation and downstream supervised or zero-shot tasks (Jones et al., 14 Aug 2025).
• Language-model-based structure priors: In symbolic regression, priors over expression trees are learned from corpora of canonical equations using n-gram LLMs. The prior $P(\mathcal{T})$ for a candidate tree $\mathcal{T}$ reflects the statistical regularities of operator and operand co-occurrences in the literature, providing non-uniform, context-aware preference for plausible forms (Bartlett et al., 2023).
• Normalization and discount for historical borrowing: In the normalized power prior (NPP) framework, the degree of trust in historical (and often literature-derived) data is controlled via a discount parameter $\delta$, to which a Beta (or more general) prior is assigned. Its hyperparameters are optimized to balance borrowing when literature and current data agree and withholding when there is conflict—using Kullback–Leibler or mean squared error criteria (Shen et al., 2023).
• Objective and asymptotically unbiased priors: Literature-motivated methods provide analytic forms or PDE-based characterizations (for linear, random-effects, or non-i.i.d. models) that ensure desirable frequentist properties, such as second-order unbiasedness, for Bayesian posteriors (Sakai et al., 2024).
• Compound prior mixtures from prior comparisons: Mixtures of $g$-priors for GLMs unify literature-driven proposals under a parametric family—often the truncated Compound Confluent Hypergeometric (tCCH) distribution—enabling the analytical comparison of model-selection and estimation properties across previously published priors (Li et al., 2015).

2. Meta-analysis, Information Aggregation, and Elicitation

Effective literature-derived priors depend on principled aggregation of heterogeneous published results:

• Meta-analytic averaging: For parametric models, aggregate all eligible studies reporting compatible parameter estimates, properly rescale coefficients (e.g., mapping probit links to log-odds via established multiplicative factors), and average both the coefficients and their estimated variances to form the prior parameters (West et al., 2019).
• Fact curation and normalization: In large-scale knowledge extraction, tagged objects are normalized to standard biomedical identifiers. Extracted facts must be consistently formulated (e.g., binary toxicities, continuous bioactivity, physiochemical properties) to enable effective pretraining and parameter sharing across downstream models (Jones et al., 14 Aug 2025).
• Hyperparameter optimization: For discount parameters in NPPs, numerical optimization of prior hyperparameters is conducted via convex combinations of KL divergences or expected MSE over plausible agreement/conflict scenarios, providing explicit control over the informativeness and robustness of the prior (Shen et al., 2023).
• Language-model-based structure likelihoods: By parsing equation corpora into symbolic trees and constructing n-gram frequency tables, a statistical model of plausible structure is distilled, mimicking the implicit inductive biases of domain experts and bench scientists (Bartlett et al., 2023).

3. Theoretical Properties and Desiderata

Several desiderata are demanded of literature-derived priors across contexts:

• Measurement and reparametrization invariance: Priors should not depend on arbitrary scale or location of the predictors; families such as the tCCH and $g$-priors maintain invariance under such transformations (Li et al., 2015).
• Model-selection and intrinsic consistency: For variable selection, priors must ensure that the Bayes factor for the true model grows decisively (or decays to zero for false models) as sample size increases, and priors themselves remain proper and non-degenerate in the large-sample limit (Li et al., 2015). Normalized power priors are constructed to explicitly avoid pathological borrowing of information in cases of data–prior discrepancy (Shen et al., 2023).
• Unbiasedness and small-sample performance: Analytically constructed priors—such as second-order unbiased or Jeffreys-type priors—reduce estimator bias in finite samples, with simulation-backed reductions in mean squared error, especially for variance components in hierarchical and mixed models (Sakai et al., 2024).

4. Empirical Applications and Benchmarks

The utility of literature-derived priors is demonstrated empirically across diverse domains:

• Survey methodology: Literature-based normal priors for response-propensity coefficients lead to substantial improvements in early and mid-field predictive accuracy for daily response propensity, outperforming standard non-informative priors and approaching the performance of fully data-driven priors where historical data exist. Even with incomplete prior coverage (≈46% of coefficients), these priors stabilize extreme estimates and offer practical utility when direct historical data are unavailable (West et al., 2019).
• Therapeutic design and molecular optimization: Multimodal models pretrained on literature-extracted entity-fact pairs (Medex priors) outperform or match resource-intensive large models in toxicity prediction, pharmacokinetics, and molecular screening tasks, and can impose experimentally justified constraints in Bayesian optimization, yielding high-quality and safe molecular proposals (Jones et al., 14 Aug 2025).
• Symbolic regression and scientific discovery: Inclusion of literature-driven, n-gram-inferred structure priors in model comparison elevates physically plausible or canonical forms (e.g., in cosmological relations) to top posterior probability, suppressing spurious, low-penalty solutions favored by unregularized likelihood or minimum description length (Bartlett et al., 2023).
• Generalized linear models: Marginal likelihoods and Bayes factors under mixture $g$-priors—with parameters mapping directly to previously published proposals—enable model selection and estimation that satisfies theoretical criteria and allows transparent investigation of prior sensitivity (Li et al., 2015).
• Small-sample random effects estimation: Asymptotically unbiased priors constructed using literature-derived analytic expressions achieve near-nominal coverage and minimized estimator bias in simulation, especially for variance components, compared with conventional inverse-Gamma or Jeffreys priors (Sakai et al., 2024).

5. Limitations, Challenges, and Guidelines

Despite their rigor, literature-derived priors are constrained by several practical and statistical factors:

• Coverage limitations: Only predictors or design elements well represented in the literature can be meaningfully informed by meta-analytic priors; unreported covariates revert to minimally informative defaults or require additional elicitation (West et al., 2019).
• Variable definition mismatch: Subtle differences in how predictors are constructed or measured across studies can induce systematic bias; careful alignment and cross-walking of covariates are mandatory before pooling (West et al., 2019).
• Fact provenance and weighting: Large-scale fact aggregation (e.g., Medex) currently treats all literature-derived facts equally, without source-quality weighting or corroboration across documents, which may propagate literature imbalances or errors (Jones et al., 14 Aug 2025).
• Finite-sample adaptivity: Some mixture priors or structure regularizers require careful hyperparameter tuning to balance between under- and over-shrinkage in finite or data-sparse regimes (Shen et al., 2023, Li et al., 2015).
• Consistency with “real-world” processes: For priors grounded in physical or biological constraints, success depends critically on the stability of underlying mechanisms over time and between the literature corpus and the target application (West et al., 2019).

Practitioners constructing literature-derived priors are advised to:

1. Undertake systematic, cross-domain searches for relevant published parameter estimates, property facts, or canonical forms.
2. Align statistical definitions, params, and measurement scales rigorously across sources.
3. Aggregate using formal meta-analysis, or, for symbolic or structured prior construction, leverage statistical language modeling.
4. Quantitatively assess the informativeness, uncertainty, and coverage of the resultant prior.
5. Where hyperparameters govern borrowing or regularization, optimize against explicit information-theoretic or prediction criteria.
6. Critically evaluate posterior robustness via simulation and, where possible, update priors as bodies of literature evolve.

6. Comparative Analysis and Future Directions

Literature-derived priors unify heuristic and theoretically grounded approaches to prior construction, enabling explicit, interpretable regularization in scenarios where domain expertise or published evidence is available but in-sample data are limited. By mapping both classic and modern methodological innovations—normalized power priors, NPP-based hyperparameter optimization, empirical Bayes shrinkage, symbolic structure models, and g-prior mixtures—onto a corpus-driven foundation, contemporary literature-derived prior frameworks provide a systematic bridge between historical knowledge and practical statistical inference.

Future work emphasizes expanding literature coverage, richer provenance modeling (e.g., weighting facts or estimates by their empirical support), joint modeling of fact interdependencies (e.g., through cross-document graphs), and adaptive updating regimes to reflect new scientific discoveries. These advances promise to further enhance the robustness, interpretability, and domain fidelity of Bayesian inference in scientific and data-scarce domains (Jones et al., 14 Aug 2025, West et al., 2019, Bartlett et al., 2023, Sakai et al., 2024, Shen et al., 2023, Li et al., 2015).