Hakken: Predicting future discoveries to fill the gaps in today's knowledge

Tarek R. Besold; Uchenna Akujuobi; Pablo Sanchez; Alessandra Toniato; Kana Maruyama; Jihun Choi; Samy Badreddine; Frederick Gifford

cs.LGSep 3, 2026

Hakken: Predicting future discoveries to fill the gaps in today's knowledge

Authors: Tarek R. Besold, Uchenna Akujuobi, Pablo Sanchez, Alessandra Toniato, Kana Maruyama, Jihun Choi, Samy Badreddine, Frederick Gifford,

Organizations: SonyAI, Barcelona, Spain. · SonyAI, Zurich, Switzerland. · SonyAI, Tokyo, Japan. · SonyAI, London, UK. ·

Abstract

We present Hakken, a domain-agnostic prediction and explanation system performing knowledge prediction, i.e., growing scientific knowledge by establishing novel relationships, ones that are not limited to the deductive hull of previous knowledge. Hakken uses a transformer-based prediction model built on temporal sequences of knowledge graphs extracted from vast bodies of research publications, fused with an LLM's semantic knowledge, to predict the presence and define the type of as-yet undocumented relationships between scientific concepts. It then calls a model-agnostic explanation framework to provide accompanying information for each prediction that allows scientists to evaluate the suggested new relationship. While general purpose, we demonstrate Hakken's practical capabilities by applying it to the biomedical domain. There, Hakken's prediction model establishes a new benchmark for time-aware multi-label relation prediction, and we show that the model's output stays coherent and informative over extended time spans in historic data. In addition, we scored 1.5 million above-confidence-threshold hypotheses related to aging, qualitatively validated batches of these predictions with biologists and progressed three of them for empirical validation in wet-lab. Two predictions with potentially significant impact in the context of drug discovery and repurposing were confirmed, introducing previously undocumented interactions between TP53 and BAMBI, and between RAF1 and TNF, to biomedical science.

Figures & tables

Figure 1: THiGERLLM for multi-label relation prediction. Given an entity pair

(u,v)

, the model combines a hierarchical temporal graph encoder with a language model. The graph branch summarizes structural and temporal evidence into compact graph embeddings, which are projected as prefix tokens and injected into the LLM together with publication-derived text. Graph-based and text-based label scores are then combined to produce the final predicted relations, and the resulting probabilities are calibrated to yield confidence scores.

Figure 2: PHELInE providing explanations for THiGERLLM ’s predictions. A surrogate model is trained once (offline) to approximate prediction scores of THiGERLLM . At inference time (online), for each hypothesis—a candidate triple (subject, relation, object)—we enumerate relational paths in the knowledge graph that connect the entities. For each path, we query the surrogate on subgraphs with and without that path to compute sufficiency (how much the path supports the prediction) and necessity (how much removing it weakens the prediction). Paths are ranked by influence and reranked by diversity to produce final explanations.

Model	$\mathrm{Prec}_{\mathrm{macro}}$	$\mathrm{Rec}_{\mathrm{macro}}$	$F_{1,\mathrm{macro}}$	$\mathrm{Prec}_{\mathrm{w}}$	$\mathrm{Rec}_{\mathrm{w}}$	$F_{1,\mathrm{w}}$	$\mathrm{nDCG}_{\mathrm{mean}}$
Random	6.08	6.86	3.78	18.21	6.07	8.36	35.20
ComplEx	5.34	8.29	5.80	17.15	19.68	18.15	47.21
KNN	33.00	26.85	28.53	60.93	58.52	59.10	74.59
MLP	39.46	26.81	28.75	66.96	60.19	61.52	86.81
RuleBased	12.81	56.39	19.08	29.43	$\mathbf{91.08}$	42.22	78.98
TNodeEmbed	40.57	27.41	29.17	68.94	62.42	63.47	87.63

Table 1: Baseline comparison for time-aware multi-label relation prediction for a 2020 cut-off. All metrics are reported in percent to improve readability.

Panel of Figure 3 — Figure 3: Historic validation results for THiGERLLM . First row (a–c): Discovery coverage across disjoint horizons. Each curve reports micro recall for new relations that first appear in each future window. Second row (d–e): Discovery coverage across disjoint horizons. Each curve reports macro recall for new relations that first appear in each future window. Third row (f–g): Stability of prioritization across disjoint horizons. Mean nDCG quantifies whether relations that first appear in a given future window concentrate near the top of the ranked list.

Appendix figures & tables23 assets

Supplementary material from the paper’s appendix.

Appendix

Figure 4: Temporal modeling via hierarchical Transformer with pairwise merging and pairmate-only attention masking.

Figure 5: Token-based injection by replacing <htg> with global node-pair embedding and prepending

M

temporal graph tokens

Figure 6: Temporal evaluation across historical cut-offs (ranking over all candidates). Mean nDCG computed over the full candidate set (including both transductive and inductive cases). Scores are reported for temporal cut-offs at 1990, 2000, 2010, and 2020.

Figure 7: Benchmark performance of models trained incrementally up to different cutoff years. Rows correspond to training cut-offs at 1990, 2000, 2010, and 2020 (top to bottom). Each model is evaluated on the following decade, except the model trained up to 2020, which is evaluated on data through 2023. Columns show different metrics.

Figure 8: Per-relation-type performance gain of THiGERLLM over THiGER as a function of relation-type test support. Left: For each relation type, the change in Precision, Recall, and F1 is plotted against the number of test facts (

\log_{10}

-scaled). Each point represents one relation type. Dashed lines show ordinary least-squares fits on log-scaled support. The inset reports the slope, p-value, and

R^{2}

for each metric. Right: Distribution of test-set support across relation types, shown on a

\log_{10}

scale. This figure illustrates how the relative performance gain of THiGERLLM varies with relation frequency.

$m=1$
Model	$\mathrm{Prec}_{\mathrm{macro}}^{m}$	$\mathrm{Rec}_{\mathrm{macro}}^{m}$	$F_{1,\mathrm{macro}}^{m}$	$\mathrm{Prec}_{\mathrm{w}}^{m}$	$\mathrm{Rec}_{\mathrm{w}}^{m}$	$F_{1,\mathrm{w}}^{m}$
Random	6.04	4.99	3.19	19.12	5.01	7.34
ComplEx	5.34	4.14	3.95	17.15	9.76	12.29
KNN	33.13	23.54	24.34	61.64	52.16	53.19
MLP	37.51	25.54	26.43	65.50	58.46	57.74
RuleBased	38.92	18.53	19.78	58.37	48.84	47.02

Appendix

Table 2: Baseline comparison for time-aware multi-label relation prediction under a 2020 cut-off using top-

m

evaluation (

m=1,3,5

). For each entity pair, only the top-

m

predicted relation labels are treated as positive predictions. Macro- and weighted precision, recall, and F1 are reported in percent. Unlike threshold-based evaluation, this setting compares models under a fixed prediction budget.

Scenario	Issue	Resolution
Only AFFECTS present (e.g., A affects B @ t1)	No inconsistency; base relation already present.	Do nothing. AFFECTS is kept.
Only INCREASES present (e.g., A increases B @ t1)	Missing the more general parent relation.	Add AFFECTS with same timestamp (t1); keep INCREASES.
Only DECREASES present (e.g., A decreases B @ t2)	Missing the more general parent relation.	Add AFFECTS with same timestamp (t2); keep DECREASES.
Both INCREASES and AFFECTS present (e.g., A affects B @ t1; A increases B @ t1)	No inconsistency; parent + child relation already aligned.	Do nothing. Both are kept.
Both DECREASES and AFFECTS present (e.g., A affects B @ t3; A decreases B @ t3)	No inconsistency; parent + child relation already aligned.	Do nothing. Both are kept.
Both INCREASES and DECREASES present (e.g., A increases B @ t1; A decreases B @ t4)	Mutually contradictory child relations. A general relation (AFFECTS) is missing.	Remove INCREASES and DECREASES , add AFFECTS with the earlier timestamp (t1).

Appendix

Table 3: Examples and rules used to resolve inconsistencies among AFFECTS , INCREASES , and DECREASES .

#	Original relation	Action	Transformed relation
1	INDUCES	keep	–
2	IS_A_BIOMARKER_FOR	keep	–
3	OF	merge into	IS_A_BIOMARKER_FOR
4	INHIBITS	keep	–
5	TREATS	keep	–
6	IS_TREATED_BY	swap (s $\leftrightarrow$ o)	TREATS

Appendix

Table 4: Relation-type renaming, merging, or removal based on expert review.

Step	Input hypotheses	Output hypotheses	Notes
prepare	1,543,297	1,524,046	Removes existing graph relations.
filter_topk_entities	1,524,046	2,804	Top-3 filter applied.

Appendix

Table 5: Filtering statistics for hyothesis selection pipeline.

Path length	Count
3	104
4	2,068
5	631
6	1

Appendix

Table 6: Shortest-path-length distribution among final Batch 2 hypotheses.

E1	E2	Relation	Recency E1	Recency E2	MPL	Confidence
SOAT1	STAT3	affects_transcriptional_activity_of	2021	2021	–	0.84183
RAF1	TNF	decreases_expression_of	2021	2011	–	0.83114
TP53	BAMBI	affects_expression_of	2021	2018	–	0.85391

Appendix

Table 7: List of the selected hypotheses for wet-lab validation. E1=entity 1 in the relation, E2=entity 2 in the relation, MPL=minimum path length exisiting between the two entitites.

E1	E2	Relation	Validation Summary
SOAT1	STAT3	affects_transcriptional_activity_of	Denied
RAF1	TNF	decreases_expression_of	Supported
TP53	BAMBI	affects_expression_of	Supported

Appendix

Table 8: List of the selected hypotheses for wet-lab validation. E1=entity 1 in the relation, E2=entity 2 in the relation.

Figure 9: Jess (Simple Western) of STAT3/pSTAT3 under induction of SOAT1. pSTAT3 induction in IL-6 or LDL treated HepG2 lysates at 4 and 24 hours under low serum conditions. HepG2 cells were pre-treated with nevanimibe (0, 0.3, 1 $\mu$ M) and then stimulated with IL-6 (100 ng/mL) for 4 or 24 hours. Total STAT3 was detected at the expected molecular weight in all treatment conditions. The induction of pSTAT3 was only observed in the presence of IL-6 and not induced by any of the LDL treatment conditions.

Figure 10: RT-qPCR of STAT3 under SOAT1 induction. Time-dependent STAT3 expression changes in response to LDL and nevanimibe treatment in HepG2 cells (4 and 24 hours, 1% serum). STAT3 expression levels were compared as a percent change relative to the vehicle control baseline (medium only). IL-6 (100 ng/mL) was included as a positive control for STAT3 induction and showed increased expression at both timepoints. Treatment with LDL failed to significantly alter STAT3 expression levels at either timepoint. Nev: + (300 nM), ++ (1000 nM)

Figure 11: TNF expression in THP-1 cells following treatment with RAF1 inhibitors GW5074 and AZ 628. THP-1 cells (8

\times

10 4 per well) were treated for 4 hours or 24 hours with 0.2% DMSO (vehicle control), 100 ng/mL LPS (positive control), or the indicated concentrations of GW5074 (+:10 nM, ++: 100 nM, +++: 1000 nM) or AZ 628 (+:30 nM, ++: 300 nM, +++: 3000 nM) RAF1 inhibitors. TNF expression was normalized to ACTB and presented as Log 2 fold change. Each condition represents n=4 biological replicates; bars represent mean

\pm

SEM. Statistically significant differences compared to the vehicle control are indicated (*p

\leq

0.05, **p

\leq

0.01, ***p

\leq

0.001).

Figure 12: TNF expression in THP-1 cells following GW5074 and ZM336372 treatment (confirmatory assay). THP-1 cells (8

\times

10 4 per well) were treated for 4 hours or 24 hours with 0.2% DMSO (vehicle control), 100 ng/mL LPS (positive control), or the indicated concentrations of GW5074 (+:10 nM, ++: 100 nM, +++: 1000 nM) or ZM336372 (+:100 nM, ++: 300 nM, +++: 1000 nM) RAF1 inhibitors. Consistent with the initial Stage 3 assay, a modest increase in normalized TNF expression was observed at 24 hours after treatment with 1000 nM GW5074. Each condition represents n=4 biological replicates; bars represent mean

\pm

SEM. Statistically significant differences compared to the vehicle control are indicated (*p

\leq

0.05, **p

\leq

0.01, ***p

\leq

0.001).

Figure 13: BAMBI expression changes in response to Nutlin-3 treatment at 24 hours. HepG2 cells were treated with the indicated concentrations of Nutlin-3 for 24 hours. Expression was quantified using the

\Delta\Delta

Ct method and normalized to ACTB. Data are presented as percentage change relative to the vehicle control baseline (0.2% DMSO). Three initial seeding densities (1, 2, or 4

\times

10 4 cells per well) are indicated by bar colors. Each condition represents n=3–4 biological replicates; bars show mean

\pm

standard error. TGF

\beta

1 (10 ng/mL) was included as a positive control for BAMBI induction. Statistically significant changes compared to the vehicle control are indicated (*p

\leq

0.05, **p

\leq

0.01, ***p

\leq

0.001).

Figure 14: Time-dependent BAMBI expression changes with Nutlin-3 treatment. HepG2 cells were treated with the indicated concentrations of Nutlin-3 for 8, 16, 24, or 48 hours. BAMBI expression was quantified using the

\Delta\Delta

Ct method and normalized to ACTB. Data are shown as percent change relative to the vehicle control baseline (0.2% DMSO). TGF

\beta

1 (10 ng/mL) was included as a positive control for BAMBI induction. Each condition represents n=4 biological replicates; bars show mean

\pm

standard error. The 48-hour TGF

\beta

1 treatment exceeded the y-axis range, with a mean percent change of 214% (

\pm

14%). Statistically significant changes compared to the vehicle control are indicated (*p

\leq

0.05, **p

\leq

0.01, ***p

\leq

0.001).

Equipment/Software Name	Cat No./ID No./Serial No./Version
Vac-Man 96 Vacuum Manifold	A2291
MiniAmp Thermal Cycler	A37834
Quantstudio 6 Flex Real-Time PCR System	4485689
Agilent 50 TS Plate Washer	S/N 2302141D
BioTek Cytation5	S/N 21041316
Gen5 software	3.13.15

Appendix

Table 9: List of equipment used for wet-lab validation.

Hypothesis	Cell Line	Vendor	Catalog No.	Morphology	Tissue	Disease
SOAT1-STAT3	HepG2	ATCC	HB-8065	Epithelial-like	Liver	Hepatocellular carcinoma
TP53-BAMBI	HepG2	ATCC	HB-8065	Epithelial-like	Liver	Hepatocellular carcinoma
RAF1-TNF	THP-1	ATCC	TIB-202	Monocyte	Peripheral blood	Acute monocytic leukemia

Appendix

Table 10: List of cell lines used for wet-lab validation.

Category	Reagent Description	Vendor	Catalog No.
Cell culture	Eagle’s minimum essential medium (EMEM)	ATCC	30-2003
Cell culture	FBS	ATCC	30-2020
Cell culture	Ca 2+ /Mg 2+ free HBSS	Merck (Sigma Aldrich)	H6648
Cell culture	Ca 2+ /Mg 2+ free PBS	Merck (Sigma Aldrich)	D8537
Cell culture	Penicillin-Streptomycin (100x)	Merck (Sigma Aldrich)	P4333
Cell culture	Recombinant Human IL-6	Fisher Scientific	200-06-50UG

Appendix

Table 11: List of reagents used for SOAT1-STAT3 hypothesis validation.

Category	Reagent Description	Vendor	Catalog No.
Cell culture	Eagle’s minimum essential medium (EMEM)	ATCC	30-2003
Cell culture	FBS	ATCC	30-2020
Cell culture	Ca 2+ /Mg 2+ free HBSS	Merck (Sigma Aldrich)	H6648
Cell culture	Ca 2+ /Mg 2+ free PBS	Merck (Sigma Aldrich)	D8537
Cell culture	Penicillin-Streptomycin (100x)	Merck (Sigma Aldrich)	P4333
Cell culture	Transforming growth factor- $\beta$ 1	R&D systems	240-B-002

Appendix

Table 12: List of reagents used for TP53-BAMBI hypothesis validation.

Category	Reagent Description	Vendor	Catalog No.
Cell culture	RPMI-1640 medium	ATCC	30-2001
Cell culture	FBS	ATCC	30-2020
Cell culture	2-Mercaptoethanol	Fisher	11508916
Cell culture	Ca 2+ /Mg 2+ free HBSS	Merck (Sigma Aldrich)	H6648
Cell culture	Penicillin-Streptomycin (100x)	Merck (Sigma Aldrich)	P4333
Cell culture	Lipopolysaccharides from Escherichia coli O127:B8 (LPS)	Merck	L4516

Appendix

Table 13: List of reagents used for RAF1-TNF hypothesis validation.

Scientific Hypothesis Generation

Scientific hypothesis generation develops methods for proposing testable explanations and research directions from literature, data, and experiments.

See all 90 papers

Explainable Forecasting of Scientific Breakthroughs from Concept Network Dynamics
Jun 2, 2026Thomas Maillart, Thibaut Chataing, Ntorina Antoni +4Explainable Artificial Intelligence Temporal Link Prediction
Geneva School of Economics and Management, University of Geneva, Geneva, Switzerland · Faculty of Medicine, University of Geneva, Geneva, Switzerland · TU Eindhoven, The Netherlands +2
SciPredict: Can LLMs Predict the Outcomes of Scientific Experiments in Natural Sciences?
Apr 12, 2026Udari Madhushani Sehwag, Elaine Lau, Haniyeh Ehsani Oskouie +14AI-Assisted Scientific Research Scientific Reasoning in Language Models
Scale AI · University of California, Los Angeles · Human Frontier Collective, Scale AI +2
GIANTS: Generative Insight Anticipation from Scientific Literature
Apr 10, 2026Joy He-Yueya, Anikait Singh, Ge Gao +5Large Language Model-Guided Scientific Discovery Research Idea Generation
Stanford University · New York University

Jun 2, 2026J/K move · Enter open · S save

Explainable Forecasting of Scientific Breakthroughs from Concept Network Dynamics

Thomas Maillart, Thibaut Chataing, Ntorina Antoni, David Dosu+3

Geneva School of Economics and Management, University of Geneva, Geneva, Switzerland · Faculty of Medicine, University of Geneva, Geneva, Switzerland · TU Eindhoven, The Netherlands+2

Explainable Artificial Intelligence Temporal Link Prediction

Open paper AskPDF arXiv

We introduce an explainable machine-learning approach that forecasts the structural precursors of scientific breakthroughs -- the emergence and intensification of links between research concepts -- by modelling how OpenAlex concept networks evolve over time. Using 59 semantic and topological features, a two-stage LightGBM model jointly predicts the formation and the future weight of concept pairs, adding a regression stage that quantifies expected intensity to prior link-existence forecasts. Relative to the state of the art, the approach improves accuracy and explainability at once: comparative validation across four technology and biomedical domains yields ROC-AUC in [0.954, 0.967] at all horizons without re-tuning, exceeding the roughly 0.90 of prior models, while every forecast rests on structural, auditable features rather than opaque embeddings. Classification performance is high (AUC about 0.95) and regression remains stable (RMSLE 0.45 to 0.6 over one to five years). Feature attribution shows that structural factors -- particularly Adamic-Adar similarity and degree-based Hadamard measures -- consistently drive accuracy, suggesting that breakthrough-relevant recombinations emerge in tightly connected sub-networks. Two expert-anchored cases, quantum annealing and AI-enabled quantum architectures, show the model surfacing technological convergence consistent with expert expectations. We then outline a three-layer decision architecture -- detection, expert translation, institutional integration -- that turns these forecasts into evidence-based research strategy and policy, anchored in open data and explainable features.

Abstract

Scientific Hypothesis Generation

Explore similar work

Explainable Forecasting of Scientific Breakthroughs from Concept Network Dynamics

SciPredict: Can LLMs Predict the Outcomes of Scientific Experiments in Natural Sciences?

GIANTS: Generative Insight Anticipation from Scientific Literature