Learning Scientific Trajectories to Predict Future Biomedical Discoveries

1 Introduction

• The introduction asks whether chronological changes in scientific claims can guide detailed, testable hypotheses, as biomedical knowledge expands and promising research directions become consequential to identify.It frames literature change as a possible signal of emerging mechanisms, interventions, and experimental settings.Source: §1 · p. 2
• Unlike prior temporal work on concept connections, topic phrases, or concept combinations, this approach generates complete claims that can express interventions, context, effect direction, and proposed consequences.The cited precedents include temporal literature-based discovery, topic-embedding forecasts, and materials-science concept combinations [13].Source: §1 · p. 2
• ForeSCITE clusters historical conclusions, learns an earlier-to-later direction within each cluster, projects conclusion embeddings along it, and decodes the projections into natural-language hypotheses.Figure 1 depicts this extrapolation from a reference conclusion to a candidate representation across the knowledge cutoff.Source: §1 · p. 2, Figure 1 · p. 3
• ForeSCITE increases semantic coverage, biomedical entity-pair novelty, and post-cutoff pair re-observation relative to evaluated baselines; expert review found 37 CAR-T candidates novel and biologically plausible.Justifications for five selected hypotheses connect generated ideas to biological mechanisms and subsequent literature where available.Source: §1 · pp. 2–3

2.2 Conclusion embeddings encode publication time and preserve temporal movement after decoding

• Across 407 clusters, early-versus-recent linear classifiers achieved mean five-fold accuracy of 0.847, while decoded and re-embedded projections generally followed the early-to-late direction.The temporal signal was graded rather than an exact recovery of publication dates or complete separation of periods.Source: §2.2 · pp. 4, 6, Figure 3 · p. 7

2.3 Temporal extrapolation expands semantic and relational coverage of future biomedical hypotheses

• At α = 0.8, ForeSCITE reached approximately 90% of human semantic coverage with recency-based references and approximately 77% with impact-based references.In both analyses, prompt-based and retrieval-augmented baselines covered less future semantic space.Source: §2.3 · p. 6, Figure 4 · p. 9
• Across both reference-selection strategies, ForeSCITE generations showed higher entity novelty and future validation than prompt-based and retrieval-augmented baselines, with α controlling their tradeoff.The highest future-validation ratios occurred at intermediate extrapolation strengths.Source: §2.3 · pp. 8–9, Figure 4 · p. 9

2.4 ForeSCITE identifies novel and plausible hypotheses on CAR-T therapy

• Expert review identified 37 of 240 prioritized CAR-T hypotheses as both novel and biologically plausible, or 2.5% of all 1,465 generated hypotheses.Hypotheses scoring 0–3 were not included in this screening review, so 2.5% is the workflow-selected share, not an estimate of overall prevalence.Source: §2.4 · p. 10, Figure 5 · p. 11

Hypothesis 1 (‘NKG2D-based CAR-T cells’ + ‘synovial sarcoma’)

• A later preclinical study directly supported the generated NKG2D-based CAR-T hypothesis for synovial sarcoma.Earlier evidence covered NKG2D activity in other tumors and engineered T-cell approaches in synovial sarcoma.Source: Hypothesis 1 (‘NKG2D-based CAR-T… · p. 10, Figure 5 · p. 11

Hypothesis 2 (‘PEGylated CAR-T cells’)

• A September 2023 study corroborated the PEGylated CAR-T concept, reporting temporary surface spacing that buffered hyperactive cell interactions and prevented acute monocyte-driven cytokine release syndrome.As CAR-T cells expanded, the coating diluted, allowing antitumor cytotoxicity to recover after the acute-toxicity window.Source: Hypothesis 2 (‘PEGylated CAR-T cells’) · p. 12

Hypothesis 3 (‘PRDM1/TCF3 double knockdown’)

• A March 2024 study linked TCF3 to progenitor-exhausted T-cell states, but in vivo persistence and cytotoxic kinetics of PRDM1/TCF3 double-knockdown CAR-T cells remain unvalidated.The proposed dual intervention is presented as a multiplex-editing roadmap, not as an experimentally established CAR-T outcome.Source: Hypothesis 3 (‘PRDM1/TCF3 double… · p. 12

Hypothesis 4 (‘CD40L/FOXP3 co-expression’ + ‘advanced renal cell carcinoma’)

• The proposed CD40L/FOXP3 combination pairs immune recruitment and microenvironment disruption with metabolic adaptation, but its in vivo synergy in renal cell carcinoma remains unverified.The passage describes targeted preclinical studies as needed to test the configuration.Source: Hypothesis 4 (‘CD40L/FOXP3… · pp. 12–13

3 Discussion

• Temporal extrapolation expanded semantic coverage and entity-pair novelty beyond prompting and retrieval, while expert review identified 37 CAR-T hypotheses as novel and biologically plausible.The five selected examples illustrate biological rationale and differing levels of subsequent literature support.Source: §3 · p. 13
• Unlike temporal literature-based discovery’s future links between predefined concepts, ForeSCITE models complete conclusions, but its comparisons do not establish superiority over full multi-agent systems.The authors describe ForeSCITE as supplying a data-derived direction before deliberation begins.Source: §3 · p. 14
• Intermediate step sizes aligned best with future literature; larger displacements increased novelty but eventually compressed generated hypotheses into a lower-rank region.The authors interpret α as a search policy and recommend selecting it using displacement and population diversity, not novelty alone.Source: §3 · p. 14
• A higher entity novelty ratio does not establish scientific utility, and about 2% of generated pairs at the best settings were novel pairs observed after cutoff.The authors interpret this as a conservative re-observation rate under incomplete ascertainment; the comparisons do not establish practical predictive value.Source: §3 · p. 14
• The five selected CAR-T examples connect findings across settings and mechanisms, but their retrospective selection does not establish a prospective success rate.Some proposed combinations still require direct experimental evaluation.Source: §3 · p. 15
• One cluster-level direction may miss branching, convergence, reversals, abrupt changes, and smaller-subfield trajectories, especially in sparse or rapidly reorganizing domains.The authors propose local, time-varying, nonlinear, or multiple-direction models and uncertainty-aware generation as possible improvements.Source: §3 · p. 15
• Published conclusions and each pipeline stage can introduce bias or information loss; temporal holdouts also cannot eliminate all pretraining or selection leakage.The authors call for prospective evaluation with corpora, models, hypotheses, and ranking rules frozen before new evidence becomes available.Source: §3 · p. 15
• A proposed hybrid would use ForeSCITE to enumerate trajectory-consistent candidates, then apply deliberation, critique, causal analysis, and feasibility constraints to refine them.The authors identify prospective testing as the next important evaluation.Source: §3 · p. 15

4.1 Preparing literature data & embedding space

• 3,305,718 conclusions with 7,749,555 PubTator relations were retained; 3,005,395 pre-cutoff conclusions trained the model and 236,057 post-cutoff conclusions formed the holdout.Records were drawn from PubMed and split at March 31, 2023; reference conclusions were selected by recency or citation impact and encoded in 1,024 dimensions.Source: §4.1 · p. 16

4.2 Training embedding language model

• The ELM reconstructs conclusion text from embeddings, using an adapter to map embeddings into the base LLM token space and LoRA to tune query and key weights.The base token-embedding layer remains frozen; training aligns the embedding space with biomedical conclusion language.Source: §4.2 · p. 17

4.3 Estimating research trend directions

• 407 thematic clusters were identified after reducing pre-cutoff embeddings to 10 UMAP components; a linear classifier estimates each cluster’s early-to-late direction.The classifier separates conclusions from the cluster’s early and late publication-year quartiles; its normalized gradient gives a stable first-order trend direction without assuming true evolution is linear.Source: §4.3 · pp. 17–18

4.4 Generating hypotheses by temporal extrapolation

• Each pre-cutoff reference embedding is shifted along its cluster’s trend direction, with α controlling semantic displacement, and the ELM decodes the projection into a hypothesis.α is an embedding-space step size, not elapsed calendar time.Source: §4.4 · pp. 18–19

4.5.1 Linear prediction of publication year

• Publication-year regression maps each cluster’s 1,024-dimensional conclusion embeddings to predicted years, which are grouped by actual publication periods to inspect temporal ordering across the cutoff.The visualizations distinguish pre- and post-cutoff conclusions by color.Source: §4.5 · p. 19, §4.5.1 · p. 19

4.5.2 Supervised two-dimensional mapping and decoded movement

• A supervised two-dimensional MLP mapping places reference and decoded, re-embedded conclusions in a shared cluster view to test whether decoding retains temporal movement.The display connects matched references and generated points while holding the mapping fixed; it also compares predicted-year distributions for extrapolated conclusions with historical groups.Source: §4.5.2 · pp. 19–20

4.6 Exploring and selecting the temporal step size

• The operating regime uses α < 1.0, selected by jointly considering paired cosine similarity, effective-rank loss, and shared UMAP geometry to limit semantic drift and concentration.Values below 1.0 kept generated conclusions closer to their references than the typical within-cluster similarity range while limiting rank loss and visible concentration.Source: §4.6 · pp. 20–21

4.7.1 Baselines

• The five LLM baselines included three prompt-only formulations and two retrieval-augmented conditions, with RAG + Iter. adding a novelty-directed revision turn.The prompt-only variants framed the task as continuation, temporal projection, or novelty; Novel RAG retrieved five nearest-neighbor conclusions, and RAG + Iter. revised its first draft against the reference and retrieved neighbors.Source: §4.7.1 · pp. 21–22

4.7.2 Quantifying semantic space expansion

• Semantic Space Growth (∆Cov) measures the marginal increase in embedding-space volume contributed by generated hypotheses.The #Circles metric estimates coverage as the volume of a union of density-threshold hyperspheres centered on embeddings.Source: §4.7.2 · p. 22

4.7.3 Quantifying novelty and future validation ratio

• The Entity Novelty Ratio measures generated entity pairs absent from training, while the Future Validation Ratio measures novel generated pairs observed in post-cutoff conclusions.The post-cutoff target set consists of co-occurring pairs absent from the pre-cutoff corpus; generated text was annotated with PubTator Tagger for comparison.Source: §4.7.3 · p. 23
• The 25.72% human-written novelty reference is the fraction of 628,434 post-cutoff entity pairs absent from the pre-cutoff corpus.It was calculated across 236,057 human-written post-cutoff conclusions and used as a fixed reference.Source: §4.7.3 · p. 23
• Uniform and degree-weighted sampling supplied chance references for future overlap among unseen entity pairs, not natural-language hypothesis-generation baselines.Both sampled unseen pairs from the pre-cutoff vocabulary; degree-weighted sampling favored historically well-connected entities without exactly preserving the training degree sequence.Source: §4.7.3 · pp. 23–24

4.8 Model-based prioritization and expert assessment of CAR-T hypotheses

• 37 of the 240 CAR-T hypotheses with GPT-5 novelty scores of 4 or 5 were judged both novel and biologically plausible by two immunology cell biologists.The 240 reviewed candidates were a subset of the 1,465 generated hypotheses; five selected hypotheses received qualitative justifications.Source: §4.8 · p. 24

A Pre-cutoff Reference Selection

• The recency-based set spans 2021–2023, while the impact-based set spans 2011–2023 and includes only papers with RCR > 2.Both reference sets contain 236,057 conclusions.Source: A Pre-cutoff Reference Selection · p. 28, Table A1 · p. 29

B HDBSCAN Minimum Cluster Size Selection

• A minimum cluster size of 500 retained a strong DBCV score while assigning fewer points to noise than a size of 250.Figure B1 compares DBCV and noise-point counts across candidate minimum cluster sizes.Source: B HDBSCAN Minimum Cluster Size Selection · p. 29

C Temporal Step-Size Selection

• The diagnostics motivated using α < 1.0 for downstream hypothesis generation, with larger step sizes retained for sensitivity analyses of semantic drift and space collapse.Figure C2 shows paired reference–generation cosine similarity and effective-rank loss, alongside training and generated embeddings in a shared UMAP projection.Source: C Temporal Step-Size Selection · p. 29

D Baseline Prompt Templates

• {reference} denotes the evaluation conclusion, while {top_k_results} denotes five retrieved nearest-neighbor conclusions.Source: D Baseline Prompt Templates · p. 29
• Figure B1 compares DBCV score and noise assignments across HDBSCAN minimum cluster sizes; 500 retains a strong score with fewer noise points than 250.Source: Figure B1 · p. 30

D.1 Direct Prompt

• The Direct Prompt asks for a possible future conclusion building on the reference and requests only the conclusion text.Source: §D.1 · p. 30

D.2 Temporal Prompt

• The Temporal Prompt asks for a hypothetical future conclusion arising from novel hypotheses and requests only the new conclusion.Source: §D.2 · p. 30

D.3 Novel Prompt

• The Novel Prompt asks for a future conclusion based on a novel hypothesis and explicitly prohibits repeating or paraphrasing the reference.Source: §D.3 · pp. 30, 32
• Figure C2 shows that increasing extrapolation step size α produces more distant embeddings but greater rank loss, with generated embeddings concentrating as α rises.Source: Figure C2 · p. 31

D.4 Novel RAG and RAG + Iter., Turn 1

• Novel RAG and RAG + Iter. use a shared first-turn template presenting the primary reference and similar existing conclusions before requesting a novel future conclusion.Source: §D.4 · p. 32

D.5 RAG + Iter., Turn 2

• RAG + Iter. adds a second-turn prompt asking for a substantially different conclusion from the reference and existing conclusions.Source: §D.5 · p. 32

• GPT-5 evaluated the novelty of each ForeSCITE-generated CAR-T hypothesis.Source: E CAR-T Hypothesis Novelty Assessment · p. 32
• The evaluation used developer-role instructions and a user-role template, replacing {testing_conclusion} with the hypothesis under assessment.Source: E CAR-T Hypothesis Novelty Assessment · p. 32
• Request parameters were listed separately from the prompt text.Source: E CAR-T Hypothesis Novelty Assessment · p. 32

E.1 Developer-role prompt

• The prompt frames the evaluator as an expert in CAR-T therapy, immuno-oncology, and translational cancer biology assessing hypothesis-like statements.Source: §E.1 · p. 32
• Novelty is defined by introducing targets, mechanisms, tissue or immune interactions, or non-standard therapeutic combinations not broadly known in CAR-T literature.Source: §E.1 · p. 32
• The rubric scores novelty from 0 to 5, ranging from well-known claims to unusual mechanistic insights, targets, or strategies not previously reported in CAR-T research.Source: §E.1 · p. 33
• The prompt excludes correctness, feasibility, and writing quality, and requests a JSON novelty score with concise reasoning.Source: §E.1 · p. 33

E.2 User-role prompt template

• The template asks the evaluator to assess the novelty of {testing_conclusion}.Source: §E.2 · p. 33