Learning Scientific Trajectories to Predict Future Biomedical Discoveries
Authors: Chia-Hsuan Chang, Brian Ondov, Jiang Lv, Lan Jiang, Yunmai Wang, Haerin Song, Xueqing Peng, Qiaozhu Mei, Xiaolei Su, Hua Xu
Organizations: Department of Biomedical Informatics and Data Science, Yale School of Medicine, New Haven, CT 06510, USA · Department of Cell Biology, Yale School of Medicine, New Haven, CT 06520, USA · Interdisciplinary Program in Artificial Intelligence, Seoul National University, 1 Gwanak-ro, Gwanak-gu, Seoul 08826, Republic of Korea · School of Information, University of Michigan, Ann Arbor, MI 48109, USA
AI co-scientists have expanded the use of large language models for generating scientific hypotheses and guiding experimental investigation. Here we investigate whether temporal patterns in the semantic representations of complete biomedical conclusions can guide hypothesis generation. We introduce ForeSCITE, a framework that identifies thematic clusters of published conclusions, estimates a direction separating earlier from more recent findings within each cluster, and projects existing conclusion embeddings along these directions. An embedding language model trained to reconstruct conclusions decodes the projected coordinates into natural-language hypotheses. Across 407 clusters derived from approximately three million pre-cutoff conclusions, we identify temporal structure and illustrate its retention after decoding. In retrospective, time-sliced evaluations, ForeSCITE increases semantic coverage beyond the historical corpus, biomedical entity-pair novelty and post-cutoff pair re-observation relative to five prompt-based and retrieval-augmented baselines. Expert review of prioritized CAR-T hypotheses identified 37 candidates as both novel and biologically plausible, and expert justifications for five selected hypotheses illustrate their biological rationale and subsequent literature support where available. These findings support temporal guidance in semantic space as a scalable mechanism for proposing biomedical hypotheses for subsequent scientific evaluation.
Figures & tables
Figure 1: Temporal extrapolation of biomedical findings for hypothesis generation. Published conclusions in natural-language space are encoded as points in a shared semantic representation space. A temporal direction learned from historical findings guides extrapolation from a reference conclusion toward a candidate representation, which an embedding language model decodes into a natural-language hypothesis. Red and blue regions distinguish historical findings and candidate future findings, respectively, separated by the knowledge cutoff. The CAR-T example illustrates how prior findings can inform a new disease-specific hypothesis. The surface and temporal arrow are conceptual illustrations of the embedding space and learned extrapolation direction.
Figure 2: ForeSCITE workflow from historical literature to candidate biomedical hypotheses. Left, Conclusions extracted from PubMed abstracts are filtered for biological relations, partitioned by the knowledge cutoff and encoded as semantic vectors. Upper middle, Thematic clustering and early–late classification identify research trend directions. Lower middle, Alignment of paired conclusion texts and embeddings produces an ELM that serves as a representation-to-text decoder. Right, Reference conclusions selected by recency or citation impact are extrapolated along the corresponding temporal directions and decoded into candidate hypotheses. Bottom, ELM alignment reconstructs the original conclusion text from its embedding through an adapter and language model. Snowflakes and flames denote frozen and trainable components, respectively. FFN, feed-forward network. Implementation and training details are provided in Methods.
Figure 3: Publication-time structure, decoded temporal movement and early–late classification of biomedical conclusions. a, Publication-time structure and decoded temporal movement across three illustrative thematic clusters. Columns show Molecular Hematologic Malignancy Research (left), Diabetic Vascular and Renal Complications (middle), and Rare Monogenic Disease Genetics (right). Top row, Distributions of publication years predicted by linear regression on 1,024-dimensional conclusion embeddings, grouped by actual publication period. Blue and orange denote pre- and post-cutoff conclusions, respectively. The separate red “Extrapolated” group shows predictions for decoded, re-embedded hypotheses generated from the most recent 5% of pre-cutoff conclusions in each cluster; the parenthetical years identify the source period, not publication dates of the generated hypotheses. Bottom row, Two-dimensional bottleneck coordinates of a publication-year-predicting MLP, with hexagonal bins colored by publication year according to the shared color bar. Light-blue markers denote ten sampled recent pre-cutoff reference conclusions per cluster, red markers denote re-embedded ELM-decoded hypotheses after extrapolation along the cluster-specific research trend direction with α=0.25 , and gray lines connect matched pairs. Both endpoints are mapped by the same fitted MLP without refitting. The paired movements generally follow the early-to-late gradient.
Figure 4: Publication-time structure, decoded temporal movement and early–late classification of biomedical conclusions (continued). b, Distribution of five-fold cross-validation accuracy for linear classifiers distinguishing early from recent conclusions across the 407 thematic clusters (mean =0.847 , s.d. =0.048 ; median =0.857 ). c, Accuracy versus cluster size, defined here as the combined number of early- and late-period conclusions used to train the classifier. Each point denotes one cluster. The red line shows a regression on log-transformed cluster size and the shaded region its 95% confidence interval (slope =0.0421 , intercept =0.7043 , r=0.364 , P=3.69×10−14 ).
Figure 5: ForeSCITE expands semantic coverage and generates novel entity pairs that align with future literature under two reference-selection strategies. a,b, Semantic space growth, ΔCov , measured by the #Circles metric at radius r=0.7 for recency-based and impact-based reference conclusions, respectively. Coverage is reported as a fraction of the coverage achieved by the 236,057 human-authored post-cutoff conclusions; error bars denote s.d. Labels above the points report the corresponding absolute coverage values. c,d, Bootstrap distributions of the Entity Novelty Ratio and Future Validation Ratio calculated from co-occurring biomedical entity pairs for the recency-based and impact-based reference conclusions, respectively. For each method and selection strategy, hypotheses were resampled with replacement 1,000 times and both ratios were recomputed for every sample (9,000 bootstrap samples per panel). Each point represents one bootstrap resample. The vertical dashed line marks the human-written post-cutoff Entity Novelty Ratio of 25.72%, calculated from 161,625 novel pairs among 628,434 pairs extracted from 236,057 conclusions. The horizontal dotted line marks the degree-weighted sampling baseline for the Future Validation Ratio (0.689%); the vertical axis is broken to display this baseline alongside the generation results. Both the recency-based and impact-based sets contain 236,057 pre-cutoff conclusions (Appendix Table A1 ). Across both analyses, ForeSCITE generations show a tradeoff between increasing entity novelty and maintaining alignment with future literature.
Figure 6: Expert assessment and temporal evidence for ForeSCITE-generated CAR-T hypotheses. a, ForeSCITE generated 1,465 CAR-T hypotheses with α=0.25 . GPT-5 assessed novelty on an ordinal scale from 0 (very low) to 5 (very high); counts and percentages are shown for each score, with scores 4 and 5 highlighted. Two immunology cell biologists reviewed all 240 hypotheses with scores of 4 or 5 (16.4% of all generated hypotheses) for novelty and biological plausibility. Expert assessment identified 37 hypotheses as both novel and plausible: 15.4% of the reviewed subset and 2.5% of all generated hypotheses. b, Temporal evidence for the NKG2D-based CAR-T hypothesis in synovial sarcoma. Pre-cutoff studies supported NKG2D-directed activity in other tumors and engineered T-cell approaches in synovial sarcoma. White, green and peach boxes denote pre-cutoff literature, the ForeSCITE-generated hypothesis and post-cutoff literature, respectively. The later preclinical study provides direct support for the disease-specific hypothesis. The timeline marks the March 31, 2023 knowledge cutoff and a buffer period ending January 1, 2024.
Appendix figures & tables3 assets
Supplementary material from the paper’s appendix.
Appendix
Year
Recency-based selection
Impact-based selection
2023
37,591
10,235
2022
141,502
27,605
2021
56,964
24,269
2020
—
21,715
2019
—
19,223
2018
—
17,431
Appendix
Table A1: Year-wise distribution of pre-cutoff reference conclusions under recency-based and impact-based selection. Impact-based references satisfy RCR >2 ; the listed counts total 236,057.
Figure B1: Sensitivity analysis for the HDBSCAN minimum cluster size. Density-based clustering validation (DBCV) score and the number of points assigned to noise across candidate minimum cluster sizes. A minimum cluster size of 500 retains a strong DBCV score while assigning fewer points to noise than a minimum cluster size of 250.
Figure C2: Extrapolation step size controls the tradeoff between semantic displacement and embedding diversity. a, Average cosine similarity between 236,057 paired reference and ForeSCITE-generated conclusion embeddings (left axis) and effective-rank loss of the generated embedding matrix relative to the reference matrix (right axis) across temporal step sizes α . The horizontal line and band show the mean within-cluster cosine similarity and ±1 s.d.; the vertical line marks the estimated crossover at α∗=1.17 . Increasing α produces more distant embeddings but greater rank loss. b, Joint UMAP projection fitted to all training embeddings and ForeSCITE-generated embeddings at α∈{0.05,0.2,0.4,0.8,1.2} . Generated embeddings broadly follow the training manifold at small α but become progressively concentrated as α increases, illustrating space collapse.
Identifying promising research directions in fast-moving subareas is one of the most cognitively expensive tasks in modern AI research. Existing LLM-driven scientific discovery systems are typically limited to one-shot prompting on static literature snapshots and are validated only against contemporary judges such as human reviewers, agent peer review, wet-lab assays, or self-evaluation, leaving open whether they can anticipate future trends. We present Continuous Knowledge Metabolism (CKM), an AI workflow for hypothesis generation with three key capabilities: (i) continuous literature metabolism via sliding windows that maintain an evolving knowledge state; (ii) predictive evaluation, which grades hypotheses against papers published after the generation window; and (iii) practitioner-grade failure detection that diagnoses workflow failure modes from its outputs. On a 50-topic machine learning benchmark, CKM-Lite produces at least one validated hypothesis on 72% of topics (36 out of 50), more than doubling a one-shot baseline (30%) at approximately 3 dollars per topic and achieving 91% lower token cost. Validated hypotheses precede their matched papers by an average of 404 days (55 hits across 36 topics; median 399 days, range 66-757 days). Broadly, predictive validation against future literature provides a falsifiable, low-cost alternative to contemporary-judge evaluation protocols and can be applied wherever a corpus has dated publication records.
Jinkai Tao, Yubo Wang, Xiaoyu Liu +1
TsingyuAI, Beijing, China · School of Information, Central University of Finance and Economics, Beijing, China · Thrust of Artificial Intelligence, The Hong Kong University of Science and Technology (Guangzhou), Guangzhou, China +1
Scientific breakthroughs often emerge from synthesizing prior ideas into novel contributions. While language models (LMs) show promise in scientific discovery, their ability to perform this targeted, literature-grounded synthesis remains underexplored. We introduce insight anticipation, a generation task in which a model predicts a downstream paper's core insight from its foundational parent papers. To evaluate this capability, we develop GiantsBench, a benchmark of 17k examples across eight scientific domains, where each example consists of a set of parent papers paired with the core insight of a downstream paper. We evaluate models using an LM judge that scores similarity between generated and ground-truth insights, and show that these similarity scores correlate with expert human ratings. Finally, we present GIANTS-4B, an LM trained via reinforcement learning (RL) to optimize insight anticipation using these similarity scores as a proxy reward. Despite its smaller open-source architecture, GIANTS-4B outperforms proprietary baselines and generalizes to unseen domains, achieving a 34% relative improvement in similarity score over gemini-3-pro. Human evaluations further show that GIANTS-4B produces insights that are more conceptually clear than those of the base model. In addition, SciJudge-30B, a third-party model trained to compare research abstracts by likely citation impact, predicts that insights generated by GIANTS-4B are more likely to lead to higher citations, preferring them over the base model in 68% of pairwise comparisons. We release our code, benchmark, and model to support future research in automated scientific discovery.
Modern artificial intelligence excels at prediction but cannot explain. From large language models to AI-for-science systems, today's machines answer what by recombining patterns already present in the human literature, yet they cannot reason out why a phenomenon must arise from underlying principles even though explanation, not prediction, lies at the heart of scientific discovery. Here we ask whether the structure of scientific explanation can be operationalized to guide how a machine generates hypotheses. We introduce DN-Hypo-Pipeline, a hypothesis-generation framework that adopts a layered, explanation-theoretic scaffold: Hempel's deductive-nomological (DN) model supplies the output form and deductive validity of a hypothesis, Salmon's causal-process account supplies an organizing constraint on where to search for the governing laws, and Armstrong's view of laws as relations between universals supplies the bridge from a phenomenon's constituent processes to the laws that may be associated with it. Rather than searching the space of what has been written, the framework searches the space of what principles govern a phenomenon: given an explanandum, it abstracts the universals instantiated in the phenomenon's formation process, retrieves the laws relating those universals, and deductively reconstructs a new, testable explanation. Evaluated in data-science modeling and judged by both LLMs and human experts, hypotheses generated through this principled reasoning significantly outperform those from direct prompting. Crucially, we translated the two highest-scoring hypotheses into novel algorithms one that reduces the Transformer's theoretical complexity with only minimal performance loss, and another that achieves competitive accuracy with substantially fewer parameters.
Lei Lin, Ronghao Wang, Chunbao Zhou +2
Computer Network Information Center, Chinese Academy of Sciences, China