Genetic testing routinely tells clinicians which variants a patient carries. Knowing why a disease is behaving the way it is (why a tumor stopped responding to a drug after eight months, why a lab value fell off a cliff) requires something else: literature that shows how those variants act biologically, through experiments such as cell assays, gene editing, or expression studies. Most variant–disease links are only statistical associations, and the rarer papers that test the mechanism are hard to find. Finding them is a retrieval problem with two parts that pull in opposite directions. The molecular part (a specific pair of co-occurring mutations, a variant identifier, a drug name) needs exact string matching. The clinical part ("partial response, then progression on second-line therapy with rising liver enzymes") needs matching on meaning. Current systems also treat the patient as a single snapshot, so they cannot ask which change in the patient's history a paper would explain.
We propose delta-keyed hybrid retrieval. The patient's record becomes two things. The first is an exact-match molecular key, scored so that documents must mention the whole combination of mutations. The second is a set of dense vectors, one for each clinically meaningful change in the timeline. A document is scored by the molecular key plus its best match to any single change event, with a learned preference for documents that report functional validation over those that report only association. The central claim can be tested directly. When the molecular key is held fixed, the relevant mechanism literature depends on the trajectory: resistance mechanisms, initiation mechanisms, and toxicity mechanisms for the same co-mutation are different papers. A retriever that encodes the timeline change by change should find them, and one that compresses the whole timeline into a single query should miss them.
Why this combination, and what already exists
Three strands of prior work motivate the design. Comprehensive Analysis of Co-Mutations Identifies Cooperating Mechanisms of Tumorigenesis analyzes 29,559 TCGA and ICGC patients. It shows that co-mutations such as TP53:KRAS in pancreatic cancer carry much stronger prognostic signal (HR 2.87) than either gene alone, often through disrupted RNA-binding-protein and microRNA sites, and that 106 of 144 prognostic co-mutation genes are absent from major clinical panels. The unit of evidence we need is therefore a combination, often involving non-coding regions, and curated single-variant resources will not cover it. That paper uses survival statistics only and does no retrieval. ClinSeekAgent shows that actively seeking sparse evidence in longitudinal records beats reading curated summaries. Claude Opus 4.6 moves from 47.5 to 62.6 F1 on its multimodal tasks. The finding we reuse is that decisive events get lost when a history is flattened. Its tools are SQL lookups, imaging tools, and generic web search, and it has no route to molecular mechanism literature. BioViL-T separates the current state from temporal change in chest X-ray pairs. We borrow its idea of representing change explicitly, but we apply it to retrieval queries rather than image encoding, and to histories longer than two time points.
Two papers supply the retrieval method. Hybrid First-stage Retrieval Models for Biomedical Literature concatenates BM25 sparse vectors with dense BERT embeddings into one searchable vector and trains the dense part on synthetic PubMed queries. It shows that dense-only retrieval confuses near-identical entity names ("JASPAR 2016" vs "2010"). Sparse, Dense, and Attentional Representations for Text Retrieval shows that a single fixed-length vector loses its ability to separate relevant from irrelevant documents as text length grows, and that multi-vector encodings help. We apply that argument to the query side: a multi-year timeline is a long query, and one vector will blur it.
The closest bridging work is An Ontology-driven Treatment Article Retrieval System for Precision Oncology. It combines lexical Elasticsearch search, ontology expansion (including OncoKB variant aliases), and a perceptron re-ranker on TREC 2017 Precision Medicine topics. It works on static single-variant vignettes, uses no dense representations, and targets treatment articles rather than mechanism evidence. Our proposal differs on all three points.
Core Mechanism
For a patient we build a molecular key $K$, the set of variant identifiers, co-mutated genes, and current drugs expanded through ontologies. We also build change events $e_1,\dots,e_T$, short natural-language descriptions of detected changes such as "progression after initial response to erlotinib" or "new grade 3 hepatotoxicity." Events come from rules over structured fields (therapy switches, RECIST-style progression labels where available, lab threshold crossings), with a language model phrasing each event. A document $d$ is scored as
$$s(K,E,d) = \alpha\, \mathrm{BM25}_{\wedge}(K,d) + \max_{t} \langle \mathbf{u}(e_t), \mathbf{v}(d)\rangle + \gamma\, m(d).$$
Here $\mathrm{BM25}_{\wedge}$ is BM25 restricted to documents containing every co-mutation element of $K$, $\mathbf{u}$ and $\mathbf{v}$ are the dense query and document encoders, $m(d)$ is a classifier's probability that $d$ reports functional validation (assays, editing, expression, binding-site disruption) rather than association alone, and $\alpha,\gamma$ are tuned weights. The maximum over events means each retrieved document is attributed to the specific change it explains. That attribution is the explanation shown to a user. The sparse and dense parts share a single concatenated index, as in Ma et al. The dense encoder is trained on synthetic event-style queries generated from PubMed abstracts that describe resistance, progression, or toxicity mechanisms, which extends their question-generation recipe.
Research Program
The work has three parts. The first is a benchmark, which the field lacks, since no dataset in our material pairs genomics, longitudinal course, and literature relevance. We will build it from TCGA cases that have mutation calls and clinical follow-up. Each topic is a molecular key plus a timeline, and PubMed documents are judged for whether they give mechanistic evidence relevant to a specific event. Ma et al. warn that biomedical relevance labels created with term-search tools favor BM25. To avoid that bias, the judging pool will draw equally from lexical, dense, hybrid, and agent systems.
The second part adds trajectory-swap topics to the benchmark. The same molecular key is paired with different constructed trajectories: response, primary resistance, acquired resistance, and toxicity. Separate relevance judgments per trajectory let us measure whether rankings shift the way they should.
The third part is the retriever itself and its ablations.
Baselines
The naive reference is BM25 over the molecular key alone, which ignores the timeline. It shows how much the trajectory adds at all.
The strongest lexical competitor is the ontology-driven system of Chen et al. (An Ontology-driven Treatment Article Retrieval System for Precision Oncology), reimplemented with the same ontologies. On TREC 2017 Precision Medicine it reports P5 of 0.7800, against 0.7000 for the top participating system. Variant nomenclature is exactly where lexical expansion is strong, so this is the baseline most likely to be hard to beat. We will run it on our benchmark with trajectory keywords appended to the query, so that it has access to the same information. We will also run our method on the original TREC 2017 topics. There we must reach at least roughly 0.78 P5, because a method that loses on static topics has traded away exact-match fidelity.
The strongest hybrid competitor is Ma et al.'s unified sparse–dense retriever (Hybrid First-stage Retrieval Models for Biomedical Literature), using the whole timeline flattened into one query. It reports average MAP 0.4056 across BioASQ 8 batches with reranking. We will first reproduce it near that number, so that our backbone is shown to be sound. On our benchmark it gets the same training data, encoder size, and tuning budget as our method.
A language-model query-rewriting baseline reads the full timeline, writes a search query, and runs it through the same hybrid index. It gets the same number of tuning trials.
An agent baseline in the style of ClinSeekAgent is given a PubMed search tool and a fixed call budget. ClinSeekAgent's reported numbers (62.6 F1 on ClinSeek-Bench) come from a different task, and no reported result was found on Lacuna for any agent on mechanism-evidence retrieval.
Three ablations isolate the key ingredients: one dense vector for the whole timeline instead of per-event vectors, plain BM25 instead of the all-elements-required version, and $\gamma=0$ to remove the functional-validation preference.
Beating the field means higher nDCG@10 and mechanism-precision@10 than every baseline on trajectory-conditioned topics, with a gap larger than the variation across topic bootstraps, and a trajectory-swap sensitivity that the single-vector ablation does not show.
Why not just…?
Why not just use BM25 with ontology expansion, which already does well on precision oncology? It is a real threat for molecular matching, and we adopt its ontology expansion inside the molecular key. What it cannot do is rank two papers that both mention TP53 and KRAS by whether they explain acquired resistance or tumor initiation. Clinical course is written in varied language that keyword expansion covers poorly, the vocabulary-mismatch failure described in the hybrid retrieval literature. The trajectory-swap test settles this. If lexical ranking with appended trajectory keywords shifts its rankings as much as ours does, then the dense event component is not needed and we will report that.
Why not just have a long-context language model read the timeline and write the query, or search as an agent? This is the most serious alternative. ClinSeekAgent's own gains came from targeted seeking rather than reading everything, which suggests a single rewritten query may still drop sparse events. The same work also reports agents wandering among irrelevant records. The rewriting and agent baselines test this directly. If rewriting matches per-event retrieval, the contribution narrows to the benchmark and the functional-validation preference. We would still check whether our method matches rewriting at a small fraction of the inference cost.
Why not just use curated databases like OncoKB or ClinVar? The co-mutation analysis shows that most prognostic co-mutation genes are missing from standard panels, and curated resources mostly index single coding variants. We will measure coverage directly: the fraction of benchmark topics where the curated resource returns any mechanism entry for the full combination.
Why not just fuse separate BM25 and dense rankings instead of a joint score? Rank fusion is cheap and often competitive, so we will include it as a variant of the Ma et al. baseline. What fusion does not provide is the per-event maximum or the requirement that all co-mutation elements appear. The single-vector and plain-BM25 ablations show whether those two ingredients carry the gain.
Decisive Checks
The first check is trajectory swap. With the molecular key fixed, the overlap between top-10 lists for different trajectories should track the overlap in their relevance judgments. The single-vector ablation should show rankings that barely move. If our method's rankings also barely move, the core claim fails.
The second check is timeline length. Following Luan et al.'s capacity argument, the gap between per-event and single-vector querying should widen as the number of events grows. We will plot it by event count.
The third check is exact-match preservation, reported as P5 on TREC 2017 Precision Medicine and MAP on BioASQ 8 against the reported 0.7800 and 0.4056.
The fourth check is mechanism precision. Independent annotators will judge whether top-ranked documents contain functional evidence. This tests whether the functional-validation preference improves mechanism evidence rather than rewarding the use of assay vocabulary.
References
- Comprehensive Analysis of Co-Mutations Identifies Cooperating Mechanisms of Tumorigenesis
- ClinSeekAgent: Automating Multi-modal Evidence Seeking for Agentic Clinical Reasoning
- Learning to Exploit Temporal Structure for Biomedical Vision-Language Processing
- Hybrid First-stage Retrieval Models for Biomedical Literature
- Sparse, Dense, and Attentional Representations for Text Retrieval
- An Ontology-driven Treatment Article Retrieval System for Precision Oncology