Papers
Topics
Authors
Recent
Search
2000 character limit reached

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

Published 3 Sep 2026 in cs.LG and cs.AI | (2609.04494v1)

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.

Summary

  • The paper introduces Hakken, an advanced system that predicts future scientific relationships using temporal knowledge graphs and language models for biomedical discoveries.
  • Hakken demonstrated a high recall rate of 60.72% in predicting rare biomedical relations, though with lower precision, by leveraging both graph and text data.
  • The system successfully supported two out of three wet-lab experiments targeting predicted relationships, showing its potential for generating actionable hypotheses.

Hakken presents a system for predicting scientific relationships that have not yet been documented and for supplying graph-structured evidence that researchers can use to assess those predictions. The central claim is stronger than conventional literature-based discovery: rather than merely retrieving, summarizing, or recombining existing facts, Hakken attempts to identify future biomedical relationships before their appearance in the literature. The system combines a temporal knowledge graph, a LLM, calibrated multi-label prediction, and model-agnostic post-hoc explanation. Its empirical evaluation spans retrospective temporal forecasting, benchmark comparison, expert prioritization, and wet-lab testing. The paper reports that two of three selected hypotheses received experimental support, including previously undocumented relationships involving TP53–BAMBI and RAF1–TNF (2609.04494).

Scientific knowledge prediction as a temporal multi-label problem

Hakken formulates knowledge prediction over a temporal biomedical knowledge graph. Each fact is represented as a directed triple (s,r,o)(s,r,o), where ss and oo are ontology-grounded entities and rr is one of 23 relation types. For an entity pair, the target is not simply whether an edge exists, but which subset of relation labels applies. This is a multi-label setting because a pair may simultaneously support several relations, such as a general affective relation and a more specific expression or transcriptional relation.

The temporal formulation is essential. Facts are associated with their first observed publication year, and models are trained only on relations available before a cutoff. Relations first appearing after that cutoff constitute the forecasting target. The resulting evaluation is intended to approximate scientific prediction rather than ordinary knowledge-graph completion. In particular, a model must identify relations that are absent from the training graph but later become documented.

The dataset is substantial but highly processed. It begins with approximately 249.3 million raw triples and is reduced through entity normalization, relation consolidation, duplicate removal, domain assignment, and polarity-conflict resolution. The final graph contains 254,806 entities, 7,127,960 triples, 23 relation types, and 20 macro-domains. Relations are extracted from PMC Open Access, MEDLINE, and licensed biomedical data using a rule-based pipeline. The paper explicitly notes that negation and speculative statements are not represented, and that only the earliest occurrence of each fact is retained. These choices simplify temporal modeling but discard evidential polarity, replication frequency, and later corroboration.

The prediction target is also treated as positive-unlabeled. Observed relations are positives, whereas absent relations are not assumed to be genuine negatives. This assumption is appropriate for scientific literature, where an undocumented relationship may be unknown rather than false. It also creates a fundamental evaluation difficulty: false positives cannot reliably be distinguished from predictions that have not yet been tested or published. The paper therefore gives particular interpretive weight to recall and ranking metrics.

THiGERLLM: fusing temporal graph structure and language

The predictive component, THiGERLLM, extends the authors’ earlier THiGER model (2609.04494). THiGER modeled temporal and structural evidence but did not predict relation labels directly. THiGERLLM adds publication-derived text and performs multi-label relation prediction.

Its architecture has two jointly trained branches. The graph branch constructs ontology-informed node representations, applies GraphSAGE-style neighborhood aggregation independently across temporal graph snapshots, and produces a sequence of time-specific representations for the target entity pair. A hierarchical temporal Transformer then models persistence, accumulation, and change across snapshots. The sequence is progressively reduced by pairwise merging until a global graph representation is obtained, while intermediate temporal representations are retained for language-model conditioning.

A notable architectural detail is the pairmate-only attention mask. Immediately before each deterministic pairwise merge, attention between the two tokens that will be merged is suppressed. The motivation is to prevent the Transformer from learning a degenerate local mixing operation that ignores cross-group temporal interactions. The design forces each token to incorporate information from other temporal positions before merging.

Figure 1

Figure 1: Hierarchical temporal Transformer with pairwise merging and pairmate-only attention masking.

The graph representation is injected into Mistral-7B-Instruct-v0.3 through learned embedding-level tokens. A global pair vector replaces a special <htg> placeholder, while a short sequence of temporal graph vectors is prepended to the language-model input. The temporal tokens are projected into the LLM hidden space, augmented with Fourier time encodings, normalized, and modulated by a learned gate. This permits the language branch to process graph structure and textual context within a common Transformer computation.

Figure 2

Figure 2: Injection of the global node-pair embedding and temporal graph tokens into the LLM input sequence.

The language input consists of sentences mentioning each entity separately, rather than sentences that explicitly mention the entity pair. For each entity, the system samples 20 sentences from publications preceding the training cutoff. This design reduces direct task leakage and requires the model to integrate contextual evidence across separately described concepts. Nevertheless, the paper concedes that temporal leakage from the pretrained LLM cannot be completely excluded because the pretraining corpus and cutoff are unspecified.

The two branches produce relation-specific scores, which are combined using a label-wise ensemble. The resulting scores are calibrated through per-label Platt scaling, isotonic regression, or a blend of both. Thresholds are also selected separately by label. The reported confidence is therefore intended to approximate the empirical probability that a predicted triple is correct, although its interpretation remains constrained by incomplete labels and imperfect temporal coverage.

Training uses a positive-unlabeled objective with dynamic weighting, focal reweighting, downweighted unlabeled-negative terms, and a bounded confidence regularizer. An auxiliary ranking loss can further separate observed entity pairs from corrupted pairs. This objective reflects an important modeling choice: unobserved relations should influence the optimization less strongly than confirmed positives, rather than being treated as ordinary negatives.

The complete forward computation combines the graph and language modalities as follows:

Figure 3

Figure 3: THiGERLLM combines temporal graph evidence, publication-derived text, label history, calibrated relation scores, and confidence estimates.

Label history is incorporated in both branches. Relations already known for a pair at the query time are encoded as contextual information that can modify candidate relation scores. This is potentially useful for hierarchical or correlated relation types, but it also introduces dependence on the completeness and consistency of historical labels.

Benchmark performance and the long-tail trade-off

The principal benchmark uses a 2020 cutoff: relations first observed by 2020 are used for training, and relations first observed afterward are evaluated as future discoveries. THiGERLLM is compared with random, ComplEx, KNN, MLP, rule-based, tNodeEmbed, and THiGER baselines.

The results show a differentiated rather than uniformly superior performance profile.

Model Macro precision Macro recall Macro F1 Weighted F1 Mean nDCG
THiGER 60.20 46.13 50.98 77.97 92.19
THiGERLLM 53.77 60.72 51.16 73.73 90.30
Rule-based 12.81 56.39 19.08 42.22 78.98
MLP 39.46 26.81 28.75 61.52 86.81
tNodeEmbed 40.57 27.41 29.17 63.47 87.63

THiGERLLM achieves the highest macro recall, 60.72%, compared with 46.13% for THiGER, an increase of 14.59 percentage points. Its macro F1 is marginally higher, 51.16% versus 50.98%. However, THiGER remains stronger on macro precision, weighted F1, and mean nDCG. THiGERLLM therefore does not dominate its predecessor across all criteria. Its advantage is specifically broader recovery across relation types, while THiGER produces sharper rankings and better performance on frequent relations.

This trade-off is important under positive-unlabeled evaluation. A prediction counted as a false positive may represent a valid relation that has not yet entered the observed literature. Consequently, macro-recall improvement may be more informative for discovery than precision improvement, provided that predictions are subsequently filtered by experts and experiments. The cost is that the increased coverage of rare relations also produces more nominal false positives.

The per-relation analysis supports this interpretation. THiGERLLM’s recall gain is largest for low-support relation types, while its precision declines in the same region. The corresponding F1 slope is not significantly different from zero, indicating that the LLM component changes the precision–recall balance without producing a systematic F1 advantage as support varies. The paper’s contradictory claim is therefore that the textual branch both improves scientific usefulness and worsens conventional precision in the long tail.

Under a fixed top-mm prediction budget, THiGERLLM becomes increasingly competitive as more labels are retained. At m=5m=5, it obtains macro F1 of 38.26%, compared with 25.95% for THiGER, and weighted F1 of 51.21%, compared with 46.09%. At m=1m=1, however, the two models are nearly indistinguishable. This indicates that THiGER already identifies strong top-ranked candidates, whereas textual evidence provides additional useful coverage lower in the ranked list.

Temporal back-testing and persistence of predictive signal

The paper evaluates whether predictions remain useful over extended horizons. For cutoff years 1990, 2000, and 2010, future relations are divided into non-overlapping intervals over the subsequent decade. Recall is measured separately for each interval, allowing short-term predictions to be compared with longer-term predictions.

The reported macro-recall curves decline by approximately 8% from early to late horizons. This decline is interpreted as a moderate degradation rather than a collapse of predictive performance. The model retains a meaningful signal for relationships that appear substantially later in the literature. A null model that randomly reassigns scores while preserving their empirical distributions obtains recall near 0.24 in early intervals and declines only slightly thereafter; the gap between this baseline and THiGERLLM is presented as evidence that the model exploits input–output structure rather than merely calibrated score distributions.

Figure 4

Figure 4

Figure 4

Figure 4

Figure 4

Figure 4

Figure 4

Figure 4

Figure 4

Figure 4: Interval micro recall for a 1990 training cutoff across successive future discovery windows.

The temporal results exhibit a stable decay profile across training cutoffs. Although the underlying graphs differ substantially between 1990, 2000, and 2010, the relative pattern remains similar: near-term relations are easier to recover, long-term relations are harder, and performance declines at a comparable rate. Mean nDCG is also reported as relatively stable across cutoffs and horizons, suggesting that the model’s prioritization remains more stable than its absolute recall.

Figure 5

Figure 5: Mean nDCG across historical training cutoffs and future temporal horizons.

This result supports a specific interpretation: the biomedical literature contains persistent structural and semantic regularities that precede the formal publication of some relationships. It does not establish that the model predicts genuinely novel biology independently of research activity. Future publication probability is influenced by funding, experimental feasibility, community attention, terminology, and institutional priorities, all of which may be partially encoded in the literature. The observed signal may therefore represent predictability of future documentation as much as predictability of underlying biological truth.

The authors also report consistent performance across cutoffs at 1990, 2000, 2010, and 2020. THiGERLLM generally maintains an advantage in macro F1 and macro recall, except at the earliest cutoff for macro recall. The benefit of textual information becomes more pronounced when more temporal data are available, although the paper does not provide a complete ablation isolating every component of the graph, text, label-history, calibration, and loss-function design.

Figure 6

Figure 6

Figure 6

Figure 6

Figure 6

Figure 6: Benchmark performance under incrementally later training cutoffs and subsequent temporal evaluation windows.

PHELInE and the construction of explanations

Hakken pairs prediction with PHELInE, or Predicted Hypothesis Elucidation with Literature-Inferred Explanations. PHELInE is designed to provide relational explanations for a predicted triple using only query access to the original predictor. It does not require access to model weights, gradients, or training procedures.

The explanation pipeline first enumerates candidate paths connecting the subject and object in the observed knowledge graph. For the wet-lab study, candidate explanations were restricted primarily to shortest paths of two to four hops. A surrogate GraphSAGE model is trained offline to approximate the original predictor’s scores. At inference time, the surrogate is queried on graphs containing or excluding candidate paths.

PHELInE distinguishes two explanation objectives. Sufficiency measures whether an explanation alone can reproduce a high score for the hypothesis. Necessity measures how much the prediction decreases when the explanation is removed from the graph. Candidate paths are then ranked by influence and reranked for diversity, so that the final explanations represent distinct intermediate entities and relational mechanisms.

Figure 7

Figure 7: PHELInE ranks graph paths by sufficiency and necessity using a surrogate model and graph perturbations.

This procedure is computationally preferable to retraining THiGERLLM for every candidate path. However, its explanation faithfulness depends on the surrogate’s ability to approximate the original model under graph perturbations, not merely on its ability to fit the original model’s unperturbed scores. The paper acknowledges this limitation. A surrogate can reproduce predictions while misestimating counterfactual score changes, particularly when the original model uses text, temporal history, calibration, and multimodal interactions that are absent from the surrogate.

PHELInE explanations should therefore be interpreted as model-behavior explanations rather than causal biological mechanisms. A path that is necessary for the surrogate’s score does not demonstrate that the corresponding biological process causes the predicted relation. It identifies contextual graph evidence that influences the predictor.

Expert filtering and wet-lab validation

The practical evaluation restricts prediction to a curated aging-related gene set. The system considers 1,386 entities, corresponding to 959,805 unordered entity pairs, and produces 1,543,297 above-threshold hypotheses. After removing relations already present in the reference graph, applying a top-three-per-entity filter, and characterizing recency and shortest-path length, 2,804 hypotheses remain. Biomedical experts select three predictions with confidence above 0.8 for experimental testing.

Hypothesis Relation Model confidence Experimental result
TP53–BAMBI Affects expression 0.85391 Supported
RAF1–TNF Decreases expression 0.83114 Supported
SOAT1–STAT3 Affects transcriptional activity 0.84183 Not supported

The validation rate is two of three selected hypotheses, or 66.7%, but this figure must not be interpreted as an unbiased estimate of system-wide precision. The three candidates were selected by experts from a highly filtered pool, and the experiments were conducted in specific cell lines and assay conditions. The result demonstrates that Hakken can generate experimentally actionable candidates; it does not establish general biomedical validity across relation types, tissues, species, or experimental protocols.

TP53–BAMBI

TP53 activation was induced with Nutlin-3 in HepG2 cells. BAMBI expression increased modestly and reproducibly at high Nutlin-3 concentration. In the initial assay, the largest changes ranged from 17% to 33% across seeding densities. In the confirmatory time course, BAMBI expression increased by 16% at 24 hours and 38% at 48 hours following treatment with 10 μ\muM Nutlin-3. P53 protein induction reached approximately 11-fold in the confirmatory experiment.

The result supports a functional relationship between TP53 activity and BAMBI expression, but the direction and mechanism remain unresolved. Lower Nutlin-3 concentrations produced decreases of up to 18%, and the magnitude of the positive response was modest relative to the TGF-β\beta control. The experiments establish regulatory association under the selected conditions, not a complete mechanistic pathway or direct transcriptional binding relationship.

RAF1–TNF

RAF1 modulation was tested in THP-1 cells using RAF inhibitors and TNF transcript measurements. GW5074 at 1,000 nM increased TNF expression by a log2 fold change of 1.1 at four hours and 1.8 at 24 hours in the initial assay. In the confirmatory assay, a 24-hour increase of approximately 0.9 log2 fold change was again observed. LPS controls produced much larger increases, with log2 fold changes of 4.2–5.0.

The result was complicated by inhibitor-specific behavior. AZ 628 reduced TNF expression at four hours, whereas ZM336372 induced ERK phosphorylation across concentrations. GW5074 also increased ERK phosphorylation, consistent with paradoxical RAF inhibitor-mediated pathway activation. Secreted TNF protein remained below the quantification limit, even when TNF mRNA increased. The authors consequently establish a reproducible transcriptional relationship under one inhibitor and concentration, but not a consistent protein-level effect. The claim that RAF1 modulation affects TNF is supported, whereas the more specific interpretation that RAF1 inhibition decreases TNF expression is not directly confirmed by the reported assays.

SOAT1–STAT3

SOAT1 was activated through LDL loading in HepG2 cells, with cholesterol ester measurements confirming pathway engagement. LDL increased cholesterol ester formation approximately twofold, and nevanimibe inhibited this effect by up to 50% under some conditions. However, LDL did not induce STAT3 phosphorylation or significantly alter STAT3 mRNA. IL-6 positive controls produced the expected STAT3 responses.

This negative result is consequential because it demonstrates that high model confidence does not guarantee experimental confirmation. The hypothesis may be false, context-dependent, or unsupported in HepG2 cells under the tested exposure, timing, and readouts. The paper appropriately limits its conclusion to the experimental design rather than treating the failed validation as definitive evidence that no SOAT1–STAT3 relationship exists.

Data, evaluation, and interpretive limitations

The most important limitation is incomplete and potentially biased supervision. The graph is extracted from literature using rule-based methods, and the paper acknowledges that negated and speculative statements are excluded. Extraction errors, ontology ambiguity, relation normalization, and polarity conflicts can therefore propagate into both training and evaluation. Removing all but the first occurrence of a fact also prevents the model from distinguishing isolated claims from repeatedly replicated findings.

Temporal leakage is reduced but not eliminated. Publication-derived text is restricted by date, yet the underlying Mistral model was pretrained on corpora with an unspecified cutoff. The model may therefore encode later biomedical knowledge. This is particularly relevant to claims of forecasting future discoveries.

The evaluation labels all relations first observed after a cutoff as future positives and treats relations not observed during the evaluation period as negatives for conventional precision calculations. This is unavoidable for some benchmark metrics but conflicts with the paper’s own positive-unlabeled formulation. The authors correctly emphasize that temporal precision is ill-defined; consequently, weighted precision, F1, and threshold-dependent comparisons should not be interpreted as direct estimates of scientific truth.

The wet-lab study is also narrow. Only three hypotheses were tested, all selected by domain experts after substantial filtering, and the assays used HepG2 or THP-1 cells. The two supported hypotheses involve modest or context-dependent molecular effects, and one does not replicate at the protein level. No independent, blinded, prospective evaluation of a larger randomly sampled set is reported.

PHELInE has a separate faithfulness limitation. Its explanations are based on surrogate perturbations and graph paths. The surrogate is not the original multimodal predictor, and sufficiency or necessity is not equivalent to causal influence. The paper leaves open how explanation fidelity should be measured quantitatively and how uncertainty in the explanations should be propagated into experimental prioritization.

Finally, the representation is restricted to categorical concepts and relations. Quantitative values, experimental conditions, dosage, tissue specificity, species, temporal intervals, and causal qualifiers are not represented in the predicted triples. The TP53–BAMBI and RAF1–TNF experiments illustrate why this matters: biological relationships can depend strongly on concentration, timing, cell type, and assay modality.

Conclusion

Hakken defines scientific knowledge prediction as temporally separated multi-label relation forecasting and implements it through THiGERLLM, which integrates temporal graph representations with publication-derived language representations. Its strongest benchmark result is a substantial increase in macro recall, particularly for rare relation types, although this comes with lower precision and weaker weighted ranking performance than THiGER. PHELInE supplements predictions with path-based sufficiency and necessity explanations, but its faithfulness depends on surrogate-model fidelity.

The system’s practical contribution is supported by a targeted discovery cycle in which two of three expert-selected hypotheses received experimental support. These results establish that Hakken can produce experimentally actionable biomedical candidates, while the failed SOAT1–STAT3 validation and the context dependence of the positive results demonstrate that confidence scores and graph explanations cannot replace biological experimentation. The principal question left open is whether the same forecasting signal and validation rate persist under prospective, larger-scale, independently sampled experiments with richer representations of evidence, polarity, quantitative conditions, and replication.

Paper to Video (Beta)

No one has generated a video about this paper yet.

Sign Up to Generate All Videos Subscribe on YouTube

Whiteboard

Explain it Like I'm 14

1. What is the paper about?

This paper introduces Hakken, an artificial intelligence system designed to help scientists discover new scientific ideas.

Scientists publish millions of papers, making it impossible for one person to read everything. Hakken reads information from scientific research and looks for possible connections that have not yet been officially discovered or written about.

The researchers tested Hakken mainly in biomedicine, especially in research about genes, diseases, aging, and cancer.

A useful way to imagine Hakken is as a very fast research assistant that:

  1. Reads and organizes scientific knowledge.
  2. Notices patterns that people may have missed.
  3. Suggests new relationships between scientific ideas.
  4. Explains why each suggestion might be reasonable.
  5. Helps scientists decide which ideas are worth testing in the laboratory.

2. What questions did the researchers ask?

The paper focuses on several main questions:

  • Can AI predict scientific relationships that may be discovered in the future?
  • Can it identify not just that two things might be connected, but also how they are connected?
  • Can it make useful predictions even when the relationship has never appeared in the scientific literature?
  • Can it explain the evidence behind its predictions?
  • Will scientists find the predictions useful enough to test in real laboratory experiments?
  • Can the system help with important biomedical problems, such as cancer, aging, and drug development?

For example, Hakken might ask whether two genes are connected and then predict a specific relationship, such as:

“Gene A may increase or decrease the activity of Gene B.”

This is more detailed than simply saying that the two genes are somehow related.

3. How did the researchers build and test Hakken?

Organizing scientific knowledge as a network

Hakken represents scientific information as a knowledge graph. A knowledge graph is like a giant map made of:

  • Points, representing things such as genes, drugs, diseases, or proteins.
  • Lines, representing relationships between those things.

For example:

1
2
3

Drug A ── treats ──> Disease B
Gene X ── affects ──> Gene Y
Protein P ── interacts with ──> Protein Q

The researchers also included time. This means the system could see when a relationship first appeared in published research. In this way, it could learn from the past and try to predict what might appear in the future.

Combining graphs with written text

Hakken uses two major sources of information:

  1. The knowledge graph: This shows how scientific concepts are connected and how those connections change over time.
  2. Scientific text: This provides the meanings and background information found in research papers.

The prediction model, called THiGERLLM, combines these sources. It uses a type of AI called a transformer, which is also related to the technology behind many modern LLMs.

The graph part is similar to looking at a social network. If two people are not directly connected, we might still guess they know each other because they share many friends. Hakken uses a similar idea, but with scientific concepts and relationships.

The text part helps the system understand what each concept means. This is especially useful when the graph contains only a few connections.

Testing on past and future information

The researchers trained Hakken using older scientific information and then tested whether it could predict relationships that appeared in later publications.

This is called temporal testing. It is similar to giving a student only information available before 2020 and asking whether the student can predict discoveries published after 2020.

The researchers compared Hakken with several other methods, including:

  • Random guessing.
  • Methods that only study the structure of the knowledge graph.
  • Methods that compare similar concepts.
  • Earlier AI models created by the same research group.
  • Rule-based systems.

Explaining predictions with PHELInE

Hakken also includes an explanation system called PHELInE.

PHELInE looks for chains of known facts that connect the two concepts in a prediction. For example:

1
2
3

Gene A affects Protein B
Protein B is involved in Process C
Process C affects Gene D

This chain might help explain why Hakken thinks Gene A and Gene D could be connected.

PHELInE checks two ideas:

  • Sufficiency: Is this group of facts enough to support the prediction?
  • Necessity: Does the prediction become weaker if these facts are removed?

This is similar to solving a mystery by asking both:

“Is this clue enough to support the explanation?”

and:

“Would the explanation fall apart without this clue?”

Laboratory testing

The researchers created more than 1.5 million possible hypotheses related to aging and selected three for laboratory testing with biomedical scientists.

These predictions were:

  1. TP53 may affect BAMBI.
  2. RAF1 may affect TNF.
  3. SOAT1 may affect STAT3.

The researchers then worked with an independent research organization to test them in laboratory experiments.

4. What did the researchers find?

Hakken performed well at predicting future relationships

Hakken was generally one of the best-performing systems in the computer tests.

Its older version, THiGER, was slightly better at some measures, especially for common types of relationships. However, THiGERLLM was better at finding a wider variety of relationships, including less common ones.

This is important because unusual or rare relationships may be especially valuable in science. They might represent:

  • New areas of research.
  • Previously overlooked biological processes.
  • Possible targets for new medicines.

Its predictions remained useful over time

The researchers tested whether Hakken could predict discoveries several years into the future.

Its performance became somewhat weaker for more distant predictions, which is expected: predicting something ten years ahead is harder than predicting something two years ahead. However, the system still performed much better than random guessing.

The researchers also found that Hakken usually continued to place the most promising predictions near the top of its list. This means scientists could use the ranking to decide which ideas to investigate first.

Two predictions were confirmed in the laboratory

Two of the three selected predictions received supporting evidence:

  • TP53 and BAMBI: The experiments supported a functional relationship between the gene TP53 and the expression of BAMBI.
  • RAF1 and TNF: The experiments supported a relationship between RAF1 and the regulation of TNF.

The third prediction, involving SOAT1 and STAT3, was not supported under the experimental conditions used.

These results are important because they show that Hakken did more than find patterns in computer data. It helped suggest relationships that scientists had not previously recorded and that could be tested successfully in the laboratory.

Hakken could provide useful explanations

Scientists who tried the system liked the explanations that came with the predictions. Rather than receiving only a statement such as “these two genes may be connected,” researchers could also see known facts and pathways that helped support the suggestion.

This makes the predictions easier to judge and may help scientists decide whether an idea is worth testing.

5. Why does this research matter?

Hakken could help scientists deal with the enormous amount of published research. Instead of reading every paper themselves, researchers could use Hakken to search for promising connections that might otherwise remain hidden.

Possible uses include:

  • Finding new drug targets.
  • Suggesting ways to reuse existing medicines.
  • Understanding how genes contribute to cancer or aging.
  • Identifying connections between diseases and biological processes.
  • Exploring new materials or agricultural methods.
  • Helping researchers choose which experiments to perform.

For example, the TP53–BAMBI and RAF1–TNF findings might eventually contribute to research on cancer, inflammation, or new treatments. However, these discoveries still require more experiments before they can be used as medical treatments.

Conclusion

The paper shows that AI can help scientists move beyond simply summarizing what is already known. Hakken attempts to predict what might become known next.

Its main strengths are that it:

  • Combines scientific text with a network of known facts.
  • Considers how knowledge changes over time.
  • Suggests specific new relationships.
  • Explains the evidence behind its predictions.
  • Produces ideas that can be tested in real experiments.

The system is not a replacement for scientists. Its predictions can be wrong, and they must be checked through experiments. Instead, Hakken acts like a powerful brainstorming and research tool. By helping scientists notice promising possibilities earlier, it could speed up discovery and lead to new medicines, technologies, and scientific breakthroughs.

View Paper Prompt  View All Prompts 

Knowledge Gaps

Knowledge gaps, limitations, and open questions

  • Limited domain validation: Hakken is evaluated primarily in biomedicine; its claimed domain generality remains untested through systematic, comparative evaluations in fields such as materials science, agriculture, physics, or social science.
  • Small wet-lab validation sample: Only three hypotheses were experimentally tested, with two receiving supporting evidence. This is insufficient to estimate the system’s general precision, reproducibility, or practical discovery rate.
  • Selection bias in experimental validation: The three hypotheses were chosen by experts from a large pool based on subjective judgments of relevance and interestingness, so the reported validation rate may not represent the performance of randomly selected or automatically prioritized predictions.
  • Restricted biological scope: Wet-lab testing focused on a curated set of aging-related genes and selected experimental systems, leaving unclear whether Hakken performs similarly for other diseases, molecular entities, cell types, tissues, organisms, and biological mechanisms.
  • Incomplete characterization of confirmed relationships: The experiments establish functional associations between TP53–BAMBI and RAF1–TNF but do not fully determine causal direction, molecular mechanism, dose dependence, pathway mediation, or context dependence.
  • Unresolved false-positive rate: Because unobserved relations are treated as potentially undiscovered rather than definitively false, the paper cannot establish how many high-confidence predictions are incorrect.
  • No systematic prospective evaluation: The study does not report a large-scale, preregistered prospective trial in which predictions are generated before independent researchers attempt validation under predefined criteria.
  • Temporal leakage remains possible: Although task-specific text is restricted by publication date, the pretrained Mistral model may have absorbed information published after the nominal cutoff. The extent and effect of this leakage are not quantified.
  • Publication and reporting bias: The knowledge graph is derived from published literature, which overrepresents positive, popular, and well-funded findings while underrepresenting null results, failed experiments, unpublished work, and research from less visible communities.
  • Extraction-error propagation: The paper does not quantify errors from entity recognition, relation extraction, entity normalization, ontology mapping, duplicate removal, or timestamp assignment, although these errors may directly affect both predictions and explanations.
  • Ambiguity in relation semantics: The 23 relation types are not shown to have consistent meaning, granularity, directionality, or inter-annotator agreement across biomedical subdomains. The effect of relation-label noise on performance is unresolved.
  • Insufficient treatment of contradictory evidence: The representation records the first occurrence of a fact but does not adequately model disagreement, refutation, uncertainty, or changes in experimental context across publications.
  • Overreliance on first occurrence dates: Treating the first reported occurrence as the decisive temporal signal ignores delayed recognition, repeated rediscovery, retractions, preprints, publication delays, and differences in research intensity across topics.
  • No analysis of entity and relation drift: The historical experiments mention terminology and entity drift but do not separately measure how changes in naming, ontology structure, biological understanding, or relation definitions affect long-horizon performance.
  • Potential confounding by popularity: Predictions may favor entities and relations that are already highly studied or densely connected. The paper does not establish whether Hakken can discover genuinely low-resource or structurally isolated relationships.
  • Limited baseline and ablation evidence: The paper does not provide sufficiently detailed ablations isolating the contributions of the LLM, temporal encoder, graph branch, label history, ontology features, calibration, or text sampling strategy.
  • Unclear benefit of language-model fusion: THiGERLLM improves some macro-level metrics but underperforms THiGER on weighted F1 and nDCG. The specific conditions under which textual evidence helps or harms prediction remain unexplained.
  • Restricted text input: The language branch uses 20 randomly sampled sentences mentioning each entity separately, without requiring sentences to mention the entity pair. It remains unclear whether this sampling strategy captures relational evidence or introduces substantial irrelevant context.
  • Random sentence sampling is not assessed: The robustness of predictions to different sampled sentences, numbers of sentences, publication sources, and retrieval methods is not reported.
  • Calibration validity is uncertain under distribution shift: Confidence scores are calibrated on held-out data, but the paper does not evaluate calibration across future years, rare relations, new entities, new domains, or substantially different publication distributions.
  • Confidence scores lack decision-theoretic interpretation: The reported confidence represents estimated truth probability under the observed data and model assumptions, but the relationship between confidence, experimental success probability, and research utility is not established.
  • Negative-label assumptions remain problematic: Benchmark precision and F1 treat relations not observed after the cutoff as negatives, despite the paper acknowledging that many may be undiscovered truths. This makes conventional classification metrics difficult to interpret.
  • Limited evaluation of ranking utility: Although nDCG and recall are reported, the study does not quantify how efficiently researchers can identify experimentally valuable hypotheses under realistic budgets, such as the top 10, 100, or 1,000 candidates.
  • No cost-sensitive evaluation: Predictions are not evaluated according to experimental cost, feasibility, expected effect size, safety, or translational value, even though the paper proposes applications in drug discovery.
  • Long-horizon validation is observational: Historic back-testing measures whether relations later appeared in the literature, not whether they were biologically true before publication. This conflates scientific predictability with publication and discovery processes.
  • Stable ranking may reflect dataset structure rather than durable scientific insight: The reported nDCG stability is not tested against popularity-matched, degree-matched, or field-specific null models that could distinguish genuine predictive structure from persistent citation or co-occurrence patterns.
  • PHELInE explanations are surrogate-based: The explanations approximate THiGERLLM through a GraphSAGE surrogate, but the paper does not quantify surrogate fidelity, explanation faithfulness, or how often the surrogate’s influential paths disagree with the original model.
  • Explanation quality is not independently evaluated: PHELInE explanations are described as useful and received positive qualitative feedback, but there is no blinded expert assessment against human-generated explanations, mechanistic validity, sufficiency, or completeness.
  • Path-based explanations may be misleading: A graph path that influences a prediction need not represent a causal or biological mechanism. The paper does not distinguish mechanistic evidence from associative or extraction-induced paths.
  • Explanation search is constrained: Wet-lab explanations are limited to shortest paths of typically 2–4 hops, potentially excluding longer, convergent, conditional, or non-path-based mechanisms.
  • Necessity and sufficiency estimates may not reflect retraining effects: PHELInE approximates score changes under graph perturbations using a surrogate rather than actually retraining THiGERLLM, so the validity of its counterfactual interpretation is unresolved.
  • Explanations omit textual and temporal evidence: PHELInE primarily exposes graph paths, while the prediction model also uses publication text, temporal signals, label history, and ontology information. The explanations may therefore fail to account for important sources of model behavior.
  • Human workflow impact is undermeasured: Feedback from researchers at seven organizations is described as overwhelmingly positive, but the study does not report sample size, participant selection, task-based usability measures, inter-rater agreement, time savings, or effects on hypothesis quality.
  • Risk of automation bias is unexplored: The paper does not test whether confidence scores and explanations cause researchers to overtrust incorrect predictions or overlook plausible hypotheses that receive lower scores.
  • Reproducibility is limited: Key details such as data-cleaning rules, entity-pair construction, candidate-pair sampling, training hyperparameters, thresholds, calibration procedures, and full prediction outputs are not sufficiently specified in the provided text.
  • Commercial data limits independent replication: Part of the dataset is commercially licensed, potentially preventing external researchers from reproducing the benchmark and validating the reported results.
  • No robustness analysis across corpora: Performance is not systematically compared across PMC, MEDLINE, PubTator3, and other biomedical databases to determine whether results depend on a particular corpus or extraction pipeline.
  • Potential institutional and geographic bias is unexamined: The literature sources and language-model training data may underrepresent research from particular countries, institutions, languages, or publication venues.
  • No evaluation of retractions or corrected findings: The system may preserve relationships extracted from retracted, superseded, or methodologically discredited publications, but the handling of such evidence is not described.
  • Quantitative and conditional knowledge is absent: The current representation cannot express effect sizes, concentration or quantity ratios, environmental conditions, cell-state dependence, temporal order, or uncertainty, limiting the scientific specificity of generated hypotheses.
  • No prediction of when discoveries will occur: Hakken ranks candidate relationships but does not forecast their likely discovery or publication time, despite temporal information being central to the model.
  • No prediction of downstream scientific impact: The system does not estimate whether a predicted relationship will generate follow-up discoveries, therapeutic applications, citations, or changes in a research field.
  • Experimental-design generation remains unimplemented: The paper identifies experimental design as a major next step but does not determine how Hakken could translate a predicted relation into a valid, discriminating, and resource-aware experiment.
  • Causal discovery is not established: The framework predicts relations and uses associative graph evidence; it does not demonstrate that it can identify causal effects or distinguish direct interactions from confounding, mediation, or correlation.
  • Generalization to genuinely novel entities is unclear: Evaluation largely concerns entities already present in the graph. Performance on newly discovered entities, sparse entities, emerging concepts, or entities absent from training data remains unresolved.
  • Ethical and safety implications are not assessed: The paper does not examine risks from generating biomedical hypotheses that could enable unsafe experimentation, misdirect clinical research, amplify biased evidence, or be misused outside expert oversight.
View Paper Prompt  View All Prompts 

Practical Applications

Immediate Applications

The paper’s validated contribution is an AI-assisted workflow for generating, ranking, and explaining previously undocumented relationships in scientific knowledge graphs. The following applications are deployable now using the demonstrated system, although most should support—not replace—expert review and empirical validation.

  • Biomedical hypothesis discovery and literature-based discovery — healthcare and life sciences. Researchers can use Hakken to identify candidate gene–gene, gene–disease, drug–target, enzyme–gene, and drug–disease relationships that are not explicitly documented in the literature. A practical workflow is:

    1. select a disease, pathway, gene set, or therapeutic area;
    2. generate candidate relations with calibrated confidence scores;
    3. inspect PHELInE’s supporting relational paths;
    4. prioritize hypotheses for review or laboratory testing. This is directly supported by the validated predictions linking TP53–BAMBI and RAF1–TNF. Feasibility depends on access to sufficiently comprehensive, high-quality biomedical literature, reliable entity normalization, and expert assessment of whether a predicted relation is biologically meaningful.

  • Drug discovery and drug repurposing prioritization — pharmaceutical and biotechnology industries.

    • prioritize compounds for mechanism-of-action studies;
    • identify existing drugs with plausible new indications;
    • rank target combinations for preclinical investigation;
    • reduce the number of hypotheses sent to expensive assays.
    • The score should be interpreted as evidence from the literature rather than as a probability of clinical efficacy. Dependencies include regulatory-grade validation, proprietary data integration, experimental confirmation, and controls for publication bias.
  • Researcher-facing hypothesis exploration through a graphical interface — academia and industry.
    • predicted hypotheses;
    • calibrated confidence scores;
    • explanatory graph paths;
    • related or potentially relevant papers;
    • filters for recent, older, or less-connected literature.
    • This could become a laboratory or research-group “hypothesis triage” workflow. Its usefulness depends on intuitive visualization, up-to-date indexing, transparent provenance, and integration with existing reference-management and laboratory-information systems.
  • Explainable AI-assisted scientific review — academia, research management, and peer review.
    • hypotheses grounded in several independent mechanisms;
    • predictions driven by a single potentially spurious path;
    • relationships supported by indirect or weakly connected evidence.
    • This is immediately useful for preparing grant proposals, planning literature reviews, and evaluating AI-generated hypotheses. The main assumption is that graph paths correspond sufficiently well to meaningful scientific mechanisms; explanations remain model-based evidence, not causal proof.
  • Systematic literature review and knowledge-gap mapping — academia and policy research.
    • research-program design;
    • review-paper scoping;
    • identification of neglected diseases or mechanisms;
    • mapping of fragmentation between subfields.
    • This application depends on coverage and representativeness of the source literature. Missing publications, terminology changes, language restrictions, and uneven publication rates can distort the apparent knowledge gaps.
  • Experimental prioritization in biology — laboratories and contract research organizations. Research teams can combine Hakken predictions with domain-specific feasibility criteria—available cell lines, assay cost, expected effect size, biosafety, and reproducibility—to select a small number of experiments. The paper demonstrates this workflow by generating over 1.5 million aging-related hypotheses, filtering them, and selecting three for wet-lab testing, two of which received confirmatory evidence. A critical dependency is that laboratory validation remains necessary; one of the three selected hypotheses was not supported under the chosen experimental design.
  • Cross-domain discovery in agriculture and materials science — industrial and academic R&D.
    • identifying combinations of agricultural practices associated with crop outcomes;
    • finding links between soil, climate, and cultivation concepts;
    • suggesting undocumented chemical or materials relationships;
    • mining historical technical literature for overlooked ideas.
    • Deployment requires domain-specific ontologies, relation schemas, entity resolution, and sufficient timestamped literature. The examples in the paper are prototypes rather than fully validated production systems.
  • Research-policy intelligence and funding allocation — public agencies and foundations. Funding bodies could use temporally indexed predictions to detect emerging scientific opportunities, identify underexplored connections, and compare areas where future discovery potential appears high. The system could support—not automate—portfolio reviews and calls for proposals. This depends on avoiding the use of model scores as direct measures of scientific importance, societal value, or funding priority, since the model primarily reflects patterns in available literature.
  • Personal and team-level research assistance — daily professional workflows. Scientists can use Hakken as a structured “second reader” that proposes connections they may not have considered while drafting a research plan or investigating a new topic. A practical workflow is to submit a known entity pair or candidate relation, inspect the score and explanations, then verify the underlying papers manually. This is feasible now but requires user training to prevent overconfidence in plausible-sounding predictions.

Long-Term Applications

These applications require improvements in representation, validation, scale, or integration beyond what the paper currently demonstrates.

  • Quantitative and conditional scientific prediction — biomedical, chemical, agricultural, and engineering sectors. Future versions could represent not only that a relation exists, but also its magnitude, direction, dose, time dependence, environmental conditions, and experimental context. For example, the system could predict that a compound affects a target only within a specified concentration range or under a particular cellular condition. This requires richer knowledge graphs, standardized quantitative extraction, uncertainty modeling, and methods that distinguish correlation from causal and conditional effects.
  • Time-to-discovery forecasting — research planning and technology foresight.
    • short-, medium-, and long-term validation horizons;
    • when a hypothesis is likely to become experimentally testable;
    • how rapidly an emerging research area may develop.
    • This depends on modeling changes in scientific attention, funding, experimental capability, terminology, and publication behavior. Long-horizon predictions are particularly uncertain, despite the paper’s finding that useful ranking signal persists over extended periods.
  • Forecasting the downstream impact of discoveries — science policy and strategic R&D. A future system could estimate whether a predicted relationship is likely to unlock many subsequent findings, treatments, materials, or technologies. Such an “impact forecast” could help prioritize research with potentially high network-level consequences. It would require historical models linking discoveries to later citations, experiments, patents, clinical developments, and industrial adoption, while correcting for citation and publication biases.
  • Automated experimental-design generation — laboratory automation and AI co-scientist systems.
    • model systems and controls;
    • perturbations and dosages;
    • assays and readouts;
    • replication strategies;
    • alternative explanations and falsification tests.
    • This requires integration with laboratory protocols, safety constraints, robotics, statistical power calculations, and laboratory execution platforms. Human scientists and institutional oversight would remain essential, particularly in biomedical and high-risk settings.
  • Closed-loop AI-guided experimentation — robotics and industrial R&D. Hakken predictions could feed robotic laboratories or materials-discovery platforms, with experimental results returned to the knowledge graph for continual updating. Potential applications include automated testing of gene perturbations, catalysts, formulations, or material combinations. This requires reliable machine-readable protocols, standardized experimental metadata, reproducible instrumentation, active-learning methods, and safeguards against reinforcing erroneous or biased predictions.
  • Clinical decision support and therapeutic discovery — healthcare. Confirmed gene–pathway and drug–disease predictions could eventually inform biomarker discovery, target validation, and individualized treatment research. However, clinical deployment would require substantially more evidence: replicated studies, causal validation, clinical trials, patient-level data, bias audits, explainable risk estimates, and regulatory approval. The current system should not be used to diagnose patients, select treatments, or make clinical decisions.
  • Materials and chemical invention platforms — energy, manufacturing, and sustainability. A scaled version could combine predicted chemical relationships with molecular simulation, synthesis planning, and property-prediction models to discover catalysts, batteries, photovoltaics, polymers, or low-carbon materials. This would require representing stoichiometry, reaction conditions, physical quantities, synthesis feasibility, and negative results—features not currently central to Hakken’s concept-level relation representation.
  • Agricultural decision support and climate-resilient production — agriculture and environmental policy. Expanded models could connect crops, practices, soil properties, pests, weather conditions, and yields to propose locally appropriate interventions. Practical deployment would depend on field trials, regional data, seasonal variability, farmer feedback, and careful treatment of causal and environmental confounding. Predictions extracted from historical literature alone may not transfer reliably to current climates or local farming systems.
  • A continuously updated scientific knowledge infrastructure — academia, industry, and government. Hakken could evolve into a shared service that continuously ingests new publications, patents, datasets, protocols, and experimental results; updates temporal knowledge graphs; recalibrates scores; and tracks which predictions have been validated or falsified. Such infrastructure could support federated scientific discovery across institutions. Dependencies include licensing and data-governance agreements, persistent entity identifiers, reproducible model versioning, provenance tracking, privacy protections, and mechanisms for correcting erroneous extracted facts.
  • More reliable scientific AI through evidence-aware calibration — all sectors. Future systems could improve confidence estimates by incorporating the number, quality, independence, recency, and consistency of supporting sources rather than treating the first occurrence of a fact as the primary temporal representation. This would make scores more useful for decision-making, but requires methods for detecting duplicated evidence, retracted studies, contradictory findings, and publication bias. Calibration must also be evaluated separately across domains, relation types, and levels of scientific maturity.
  • Collaborative human–AI discovery ecosystems — education and professional training. In advanced forms, Hakken-like systems could teach students how to formulate hypotheses, trace evidence paths, compare competing mechanisms, and design validation studies. This could support graduate training and interdisciplinary collaboration. The principal dependency is pedagogical validation: interfaces must encourage critical reasoning and source checking rather than turn model outputs into unexamined answers.
View Paper Prompt  View All Prompts 

Glossary

  • ADMET score: A measure of a substance’s absorption, distribution, metabolism, excretion, and toxicity properties. “alongside other mechanisms like ADMET scores”
  • Auxiliary ranking loss: An additional training objective that encourages correct items to receive higher scores than incorrect ones. “the graph branch is additionally trained with an auxiliary ranking loss”
  • Back-testing: Evaluating a model by simulating predictions using only information that would have been available at an earlier historical time. “Back-testing the Model on Historic Data”
  • Calibration: The process of adjusting predicted scores so that they correspond more accurately to empirical probabilities. “the resulting probabilities as confidence scores”
  • Candidate explanation: A proposed set of facts or paths used to explain a model’s prediction. “candidate explanations were restricted to shortest subject--object paths”
  • Class-balanced multi-label objective: A training loss that compensates for imbalanced label frequencies when multiple labels may apply simultaneously. “We therefore use a class-balanced multi-label objective”
  • Counterfactual probability: The estimated probability of an event under a hypothetical alternative condition. “representing the counterfactual probability that it could already exist in the literature”
  • Deductive closure: The complete set of conclusions that can be logically derived from an existing set of facts and rules. “the deductive hull of previous knowledge”
  • Domain-agnostic: Designed to operate across subject areas without being restricted to one domain. “a domain-agnostic prediction and explanation system”
  • Domain-general: Applicable to multiple scientific fields rather than specialized for one field. “have intentionally been built to be domain-general”
  • Entity drift: Changes over time in the meaning, representation, or contextual associations of an entity. “shifts in terminology, or entity drift”
  • Entity marker: A token or annotation identifying a particular entity in input text. “When entity markers are available”
  • Explainability framework: A system that produces information about the factors supporting a model’s output. “a model-agnostic explanation framework”
  • F1 score: The harmonic mean of precision and recall, used to summarize classification performance. “Weighted F1 emphasizes frequent relation types”
  • Graph embedding: A vector representation that encodes structural information about a graph node, subgraph, or graph-related object. “The encoder produces a global pair embedding”
  • Graph perturbation: A deliberate modification to a graph, such as adding or removing nodes or edges, used to study model behavior. “estimate prediction changes under graph perturbations”
  • GraphSAGE: A graph neural-network method that generates node representations by aggregating information from neighboring nodes. “a GraphSAGE-style encoder”
  • Held-out validation set: Data excluded from model training and used to calibrate or evaluate the model. “calibrated on a held-out validation set”
  • Hypothesis generation: The automated production of proposed, testable relationships or explanations. “the full cycle from hypothesis generation to wet-lab validation”
  • Incomplete supervision: A learning setting in which missing labels are not necessarily treated as negative examples. “Training is designed for incomplete supervision”
  • Influence score: A numerical estimate of how strongly a fact or explanation affects a model prediction. “After computing these influence scores”
  • Knowledge graph: A graph whose nodes represent entities and whose edges represent typed relationships between them. “A knowledge graph is represented as”
  • Knowledge prediction: The task of predicting previously undocumented relationships or facts. “We present Hakken, a domain-agnostic prediction and explanation system performing knowledge prediction”
  • Label history: Previously known relation labels for an entity pair, used as additional predictive context. “We also incorporate label history”
  • Literature-Based Discovery (LBD): The identification of potentially useful connections by analyzing relationships reported across scientific publications. “Technologies like Literature Based Discovery (LBD)”
  • Macro recall: Recall calculated separately for each relation type and then averaged, giving equal weight to each type. “Here the recall is first computed per relation type and then averaged”
  • Model-agnostic: Usable with different predictive models without requiring access to their internal parameters or training procedures. “a model-agnostic explanation framework”
  • Multi-hop path: A graph path connecting two entities through multiple edges or intermediate entities. “typically forming multi-hop paths between subject and object”
  • Multi-label classification: A classification setting in which one instance may receive several labels simultaneously. “We formulate this as a multi-label classification problem”
  • nDCG: Normalized discounted cumulative gain, a ranking metric that gives greater importance to correctly placing relevant items near the top of a list. “the quality of the predicted ranking”
  • Null baseline: A comparison model that preserves score distributions while removing the meaningful relationship between inputs and outputs. “we introduce a null baseline”
  • Ontology-based representation: An entity representation derived from its position or descriptors in an ontology. “the ontology-based node representations of ss and oo”
  • Ontology: A structured specification of concepts, categories, and relationships within a domain. “ontology files were provided that map each entity to one or more hierarchical domain descriptors”
  • Prefix token: A token inserted at the beginning of a language-model input to convey additional information. “projected as prefix tokens and injected into the LLM”
  • Pseudo-retraining: Approximate retraining performed with a surrogate model rather than repeatedly retraining the original model. “subsequent surrogate-based pseudo-retraining”
  • Relational pattern: A recurring configuration of connected facts or relations in a knowledge graph. “through relational patterns, i.e., contextual subgraphs of facts”
  • Surrogate model: A simpler or substitute model trained to approximate the outputs of another model. “a surrogate model trained to approximate prediction scores of THiGERLLM”
  • Temporal leakage: The unintended use of information from after the permitted time period during prediction or evaluation. “residual temporal leakage cannot be fully excluded”
  • Temporal knowledge graph: A knowledge graph represented as a sequence of time-indexed graph snapshots. “We consider a temporal knowledge graph represented as a sequence of TT graph snapshots”
  • Temporal encoder: A model component that represents patterns and dependencies across time. “which is processed by a temporal encoder”
  • Temporal horizon: The future time interval over which predictions are evaluated. “long-horizon prediction”
  • Temporal cut-off: A time boundary separating information available for training from information reserved for evaluation. “using a 2020 temporal cut-off”
  • Transformer: A neural-network architecture based primarily on attention mechanisms for processing sequential data. “a transformer-based prediction model”
  • Wet-lab validation: Experimental testing conducted in a laboratory using biological materials or processes. “progressed three of them for empirical validation in wet-lab”
  • Zero-shot: The ability to make predictions for tasks or relationships without having received direct labeled examples of them during training. “the first approach that predicts future relation labels in a multi-label setting”
View Paper Prompt  View All Prompts 

Open Problems

We haven't generated a list of open problems mentioned in this paper yet.

Generate Now

Tweets

Sign up for free to view the 2 tweets with 827 likes about this paper.

Sign Up for Free