AI that reads the unpublished: a three-layer protocol flags four blind spots in CPEB biology and yields a testable sleep-forgetting hypothesis
Synopsis
AI systems for scientific discovery usually learn only from published positive results, leaving them structurally blind to failures and suppressed directions. A three-layer operational protocol turns that gap into executable steps: it scores theoretically possible but undocumented relationships in a knowledge graph, uses an adversarial classifier to surface features systematically suppressed in unpublished, rejected or failed data, and generates hypotheses that diverge from consensus yet pass viability filters. In a CPEB3/CPEB4 neurobiology case, it flagged four structural blind spots and produced a testable hypothesis about astrocytic translation during sleep and active forgetting.
Interpretation
Discovery-oriented AI is fed published literature, so it learns only successes and consensus and has no representation of failures, suppressed directions or untried ones. An earlier counter-inductive architecture argued why such a system is needed but specified no inputs, algorithms, parameters or validation criteria, so it could not be built directly. The protocol writes out each layer's operational steps, default hyperparameters and validation criteria, so the three layers can be chained as one pipeline.
The pipeline builds a knowledge graph, scores relationships that are theoretically possible but absent from the literature, and flags high scorers as structural blind spots. Conventional practice counts only documented relationship patterns; absence itself is not a search target. In the CPEB3/CPEB4 case the graph held 247 entities and 1,832 edges; at a threshold of 0.7 it flagged CPEB3–astrocytes, CPEB4–microglia, CPEB3–active forgetting and CPEB3–sleep, with absence surprise scores of 0.89, 0.84, 0.91 and 0.87.
The second layer pools unpublished preprints, rejected manuscripts and trial registrations that missed endpoints with the published literature, trains a classifier to separate them, and reads off the features that most distinguish the two. Such material normally stays out of literature reviews and model training, so its information is discarded. The adversarial classifier reached an AUC-ROC of 0.83 on a held-out test set, above the 0.80 threshold; suppressed features included glial cell types (suppression score 3.42), sleep states (2.87), forgetting processes (2.65), non-neuronal mRNA translation (2.31) and circadian regulation of translation (2.18).
The third layer has a generative model connect blind spots with suppressed features into testable statements, then ranks them by divergence from consensus minus viability penalties. Conventional generation rewards agreement with the literature, treating high-consensus hypotheses as the plausible ones. The model produced 47 hypotheses and 12 survived viability filtering; the top one had a consensus probability of 0.012 and a counter-inductive loss of 0.73, with all four viability penalties zero, and proposed knocking down CPEB3/4 in astrocytes during sleep and measuring memory retention in a fear-conditioning paradigm.
Perspective
The protocol applies to domains where a knowledge graph can be built and where negative results, failed experiments or unpublished data exist or can be digitized; the case centers on the CPEB3/CPEB4 protein system. Intended readers are computational biology, neuroinformatics and AI engineering researchers with knowledge-graph, natural language processing and generative modeling experience. The authors propose next steps: automating dark-data acquisition and normalization, developing viability functions that generalize across domains, and applying the protocol to other fields such as mathematics.
The authors name three limits: dark-data access may be constrained by legal, ethical or institutional rules; the viability filters are currently binary and expert-defined, awaiting more nuanced continuous functions; and the protocol has been explored only in the CPEB domain, so generalization to other fields remains to be tested.