ForeSCITE
A temporal-trajectory-guided framework for biomedical scientific hypothesis generation. It identifies the directions of early and recent findings within topic clusters, projects conclusion embeddings onto a temporal direction, and trains an embedding language model to decode them into natural-language hypotheses.
Timeline (last 90 days)
ForeSCITE is a framework that uses the temporal trajectories of semantic drift in biomedical conclusions over time to guide the generation of future hypotheses.
ForeSCITE runs on 407 topic clusters derived from approximately three million conclusions before the cutoff date.
ForeSCITE's method chain includes: identifying topic clusters, estimating a direction that distinguishes early from recent findings, projecting conclusion embeddings onto that direction, and decoding the projected coordinates into natural language hypotheses with an embedding language model.
Identifiable temporal structure exists in semantic space and is preserved after decoding.
In a retrospective evaluation sliced by time, ForeSCITE outperforms five prompt-based and retrieval-augmented baseline methods on three dimensions: semantic coverage, novelty of biomedical entity pairs, and the proportion of post-cutoff entity pairs that are re-observed.
Expert review judged 37 CAR-T hypotheses as both novel and biologically plausible.
ForeSCITE introduces time as a first-order signal into scientific hypothesis generation, rather than merely as a retrieval filter.
The sample size of the expert review is limited.
Temporal steering in semantic space is a scalable mechanism that does not rely on handcrafted rules from domain-specific knowledge bases, and is in principle transferable to disciplines beyond biomedicine.
All Facts (9)
ForeSCITE is a framework that uses the temporal trajectories of semantic drift in biomedical conclusions over time to guide the generation of future hypotheses.
50%UnverifiedForeSCITE runs on 407 topic clusters derived from approximately three million conclusions before the cutoff date.
50%UnverifiedForeSCITE's method chain includes: identifying topic clusters, estimating a direction that distinguishes early from recent findings, projecting conclusion embeddings onto that direction, and decoding the projected coordinates into natural language hypotheses with an embedding language model.
50%UnverifiedIdentifiable temporal structure exists in semantic space and is preserved after decoding.
50%UnverifiedIn a retrospective evaluation sliced by time, ForeSCITE outperforms five prompt-based and retrieval-augmented baseline methods on three dimensions: semantic coverage, novelty of biomedical entity pairs, and the proportion of post-cutoff entity pairs that are re-observed.
50%UnverifiedExpert review judged 37 CAR-T hypotheses as both novel and biologically plausible.
50%UnverifiedForeSCITE introduces time as a first-order signal into scientific hypothesis generation, rather than merely as a retrieval filter.
50%UnverifiedThe sample size of the expert review is limited.
50%UnverifiedTemporal steering in semantic space is a scalable mechanism that does not rely on handcrafted rules from domain-specific knowledge bases, and is in principle transferable to disciplines beyond biomedicine.
50%