Showing the Ambiguity: Contrastive View Recommendation for Underspecified Genome Browser Queries

A researcher types "rs1447295 prostate cancer" into a genome browser. A genome browser is a web tool that draws stacked data tracks (gene models, sequencing signal, chromatin contacts) along a chromosome coordinate axis. Today the browser centers on the variant and shows whatever tracks were configured in advance. The query is ambiguous in ways that matter biologically. The researcher may want the gene the variant sits inside (CASC8) or the gene it physically contacts in 3D (MYC, megabases away). The researcher may want lung fibroblast chromatin data because that is what is loaded, or prostate data because that is the disease. And it is unclear whether the right window is a few kilobases or a few megabases. Each reading produces a different picture, and picking the wrong one silently produces the "nearest gene" error that HiView documents.

We propose that in coordinate genomics the clarifying question should itself be a pair of views, not a sentence. The system keeps an explicit distribution over structured interpretations of the query: cell type, target-gene reading, data modality, and window size. It renders the two candidate views whose difference is most informative about which interpretation the user holds. It attaches to each view a rationale that is a path in a biological knowledge graph, with every edge of the path drawn as a visible feature on screen. The user's click or zoom then works as the answer to the clarifying question. The rationale cannot claim anything the view does not show, so it is faithful by construction rather than by post-hoc checking.

Core Claim

Ambiguity in a genomic query can be measured as disagreement among the views its interpretations would render, and it can be resolved more cheaply and more accurately by contrastive view recommendation than by text clarification or by guessing. Let $Z$ be the hidden interpretation, $q$ the query, $p(z \mid q)$ the system's current belief over interpretations, and $C$ the user's choice between candidate views $v_1$ and $v_2$. The system shows the pair

$$ (v_1, v_2)^* = \arg\max_{v_1, v_2} \; I(C; Z \mid q, v_1, v_2), $$

where $I$ is the mutual information between the user's choice and the interpretation. In plain terms, the system shows the two pictures whose contrast best splits the plausible readings. When $p(z \mid q)$ is concentrated, it shows a single view and asks nothing.

The rationale for view $v$ is a path $\pi$ through a knowledge graph, for example variant → overlaps → CASC8 and variant → Hi-C contact in cell type X → MYC promoter. The path is admissible only if every edge corresponds to a rendered element: an arc, a peak, or a gene model. When the only available contact data come from a mismatched tissue, as with the IMR90 lung data HiView used for a prostate variant, that mismatch becomes an explicit flagged edge in the rationale and is not left hidden.

How This Differs From Prior Work

Three lines of work each supply one piece and leave the conjunction open.

Visualization recommenders work on generic tabular charts. LLM4Vis prompts an LLM to pick among four chart types and generate natural-language rationales, but those rationales are free text that need not reflect what drove the choice. KG4Vis learns knowledge-graph embeddings that yield readable design rules, but it is static and has no biological entities. Hier-SUCB adapts interactively with a semi-bandit (a learner that updates from partial feedback on each recommended combination), but it explains nothing and asks users for fatiguing per-component feedback. We take from Hier-SUCB the idea of factoring a view into separable components. We differ by asking for one binary choice between whole views, and by making explanations structural rather than generated.

Clarification research supplies the uncertainty machinery. CLAIM measures ambiguity as disagreement among sampled answers and picks clarifying questions by information gain. We reuse that criterion but apply it to views instead of text questions. Tree of Clarifications expands ambiguous questions into a tree of readings, at up to 20 LLM calls per query, which is incompatible with the fast viewport updates that ABrowse treats as essential. CLAMBER constrains our design. It shows instruction-tuned LLMs tend to guess rather than clarify, and that few-shot and chain-of-thought prompting often fail to help. We therefore do not trust an LLM's self-report of ambiguity. Instead, we derive ambiguity from the structured interpretation space and from which data actually exist.

Biomedical grounding comes from Knowledge-based citation reasoning, which explains citations as triplets drawn from curated databases. We borrow the triplet form of explanation but tie each triplet to a rendered track instead of to text.

Genome browsers such as HiView, dbEmbryo and OmnibusX provide the coordinate substrate and track stacking. They have no recommender, no ambiguity model and no explanations.

Research Program

The build has three parts.

The first part is an interpretation parser. It maps a free-text query to a distribution over structured slots: locus or variant, target-gene reading (overlapping, nearest, or contacted), cell type, modality (linear signal or 3D contact arcs), and window scale. An LLM proposes candidate slot fillings by sampling, which is how CLAIM measures disagreement. A knowledge-graph filter then removes fillings for which no data exist and adds readings the LLM missed, such as long-range contacts present in the loaded Hi-C maps. This hybrid is the guard against the overconfidence CLAMBER documents.

The second part is a view compiler. It turns each interpretation into a concrete browser state: tracks, coordinates, and Hi-C contact matrices compressed into 1D arcs as in HiView and MHiC. It also records which knowledge-graph edges each rendered element witnesses.

The third part is a selector. It chooses either one view or a contrastive pair by the information-gain criterion. It updates $p(z \mid q)$ from clicks, zooms and track toggles over a multi-turn session. That multi-turn tracking also addresses CLAIM's single-turn limitation.

Our latency target is one LLM call per query plus cheap graph and rendering computation.

There is no benchmark for this task, so we will build one. Variants come from the GWAS Catalog, gene models from RefSeq and Ensembl, and contact data from publicly available Hi-C maps across several cell types, following the HiView corpus design. For each variant we sample a hidden interpretation. We then generate a query that omits a controlled subset of slots, labelled by CLAMBER's taxonomy: missing cell type and window are aleatoric omissions, while gene-alias collisions are linguistic. A simulated user picks the view consistent with the hidden interpretation, with noise. This follows Hier-SUCB, which validated its simulator against 51 human participants; we will run a similar fidelity check.

Baselines

The naive reference is the do-nothing browser. It runs exact-identifier lookup that centers on the variant with a fixed default track set, as HiView does. It establishes how often users land on the wrong interpretation without any help. A second naive reference is a "show everything" view that stacks all cell types' contact arcs over a wide window in one dbEmbryo-style dashboard. This baseline matters because it resolves ambiguity by never committing to an interpretation.

The strongest existing competitor is a CLAIM-style text clarifier. We sample multiple LLM readings of the query, measure their disagreement, ask the information-gain-maximizing clarifying question in text, and then render the resulting view with an LLM4Vis-style generated rationale. It is the strongest competitor because it already uses the same information-gain principle and the best available ambiguity detector. The only difference from our method is the medium of the question and the source of the explanation. It gets the same LLM, the same knowledge graph, the same track catalog and the same number of user turns. Prompts are tuned on a held-out split with the same budget we use for our parser. We also run a Hier-SUCB-style semi-bandit over factored track configurations with the same interaction budget, as the representative of unexplained online adaptation.

No reported result exists on Lacuna for any of these methods on genomic view selection, since the task is new. The nearest reported figures are on textual tasks and serve only as calibration. ChatGPT reaches 54.25% accuracy and 52.77% F1 on CLAMBER ambiguity identification (CLAMBER). CLAIM reaches 81.85% accuracy on ClariLM-test (CLAIM). ToC reaches 33.7 Disambig-F1 on ASQA (Tree of Clarifications). We will report our parser's ambiguity-identification accuracy on CLAMBER as a sanity check that it is not weaker than an LLM detector on text.

Two ablations isolate the key ingredients. The first replaces contrastive pairs with a single top-ranked view plus rationale, testing whether showing the ambiguity is what helps. The second replaces path-constrained rationales with LLM-generated free-text rationales over the same view, testing faithfulness.

Beating the CLAIM-style baseline means three things. Interpretation recovery must be higher at equal turns, or equal recovery must come with fewer turns and fewer LLM calls. The rationale-grounding rate must be higher. The nearest-gene error rate on looped-target cases must be lower.

Why not just…?

Why not just ask a clarifying question in text, as CLAIM does? This is the real threat, and it is our main baseline. Our hypothesis is that some ambiguities are hard to phrase but easy to see. "Do you mean the gene it contacts in 3D in prostate cells?" requires the user to already know a contact exists, whereas a pair of views shows the arc to MYC. If the text clarifier matches us on interpretation recovery and nearest-gene error at equal turns, the contrastive-view idea fails, and we will report that. The faithfulness advantage would then remain as a separate, smaller contribution.

Why not just show all tracks at once? Stacking avoids committing, but the interpretation space multiplies across cell types, modalities and window scales. A megabase window and a kilobase window cannot coexist on one axis. We test this directly with the show-everything baseline, measuring recovery, task time and NASA Task Load Index. BioGraphia's drop in that workload score from 63 to 23 under assisted annotation (BioGraphia) suggests workload is the axis where clutter costs. If show-everything ties on accuracy and workload, the selector is unnecessary.

Why not just prompt a strong LLM to pick the view and explain it? CLAMBER reports that such models guess instead of clarifying, missing over 80% of blatant contradictions, and that chain-of-thought often does not help. The citation-reasoning work found general LLMs below a knowledge-grounded model on triplet extraction (73.55% F1 for the grounded model). Our free-text-rationale ablation measures this directly, as the fraction of rationale claims with no rendered witness.

Why not just expand the query with retrieval? Expansion, in the style of pseudo-relevance feedback, enriches a single reading and so collapses the ambiguity rather than exposing it. ASQA's gap between retrieval-augmented models (32.1 DR) and humans (61.8 DR) (ASQA) reflects exactly this conflation of divergent intents. We include expansion-then-single-view as an additional arm.

Proposed Experiments

Four checks are decisive. The first is simulated-user sessions on the constructed benchmark, stratified by ambiguity type. Here we measure interpretation recovery at one, two and three turns, LLM calls per query, and nearest-gene error on variants whose hidden target is a contacted gene rather than an overlapping one. The second is automatic faithfulness scoring: the fraction of rationale claims witnessed by a rendered element, plus the rate at which a tissue mismatch is surfaced when the hidden cell type has no matching data. The third is the simulator fidelity check against human choices. The fourth is a small study with genomics researchers on held-out variants, comparing our system with the text clarifier and the show-everything view on correct target identification, time and NASA Task Load Index.

The proposal stands if contrastive views beat text clarification on recovery or turns on the ambiguity types where the readings differ visibly, and if path-constrained rationales stay near full grounding while free-text rationales do not. If neither holds, the honest conclusion is that a good text clarifier is enough for genome browsers.