Eigen Insights / Emergence
Expert-authored biology tasks for AI training and evaluation.
Scientists turn biological study data into tasks with reference analyses, documented reasoning, and scoring rubrics. Independent experts review the work before release.
How should we interpret conflicting RNA results?
A model and an expert interpret the same illustrative dasatinib results.
Example model response
Human expert reference
1 / Select evidenceUses the stronger correlation as the main evidence.
1 / Check agreementCompares the two pipelines and flags that the correlation changes sign.
2 / Test assumptionsDiscounts the weaker result without checking how it was produced.
2 / Test assumptionsChecks sample pairing and preprocessing, then reruns the analysis on matched samples.
3 / InterpretCalls the signal robust based on one pipeline.
3 / InterpretKeeps the finding provisional pending an independent check.
What the rubric rewards. Reconciling conflicting evidence, checking assumptions, and stating what would change the interpretation.
Check whether the result survives both pipelines.
The chart uses eight synthetic observations to show how the choice of RNA pipeline can change the interpretation.
Decision. Flag the disagreement for dasatinib and investigate it before calling the RNA signal robust.
Assess sensitivity to venetoclax and dasatinib.
The assigned expert turns the research brief into a task and completes the reference analysis. This task uses RNA readouts from two pipelines, plus exome, proteome, phosphoproteome, and measured ex vivo response for both drugs.
- Choose a method that fits the study and apply it consistently.
- Check whether the two RNA pipelines support the same conclusion for each drug.
- Use other molecular evidence to probe disagreements and state what remains uncertain.
Expert deliverables. The authored task, runnable reference code, outputs, and a defensible interpretation for each drug.
Correct calculations can still earn a poor score.
In the evaluation environment, a model completes the reviewed task independently. Expert reviewers compare its methods, outputs, and reasoning with a human expert’s completion of the same task. Here, the example model reproduces the correlations but treats the dasatinib signal as robust.
- Execution · 2/2. Code reproduces the reported correlations.
- Method · 2/2. The submission reports both pipelines’ results for each drug.
- Interpretation · 0/2. The conclusion overlooks the disagreement between pipelines.
Illustrative score · 4/6. The submission needs to revise its conclusion and identify what to check next.
Task catalogue
Task collections.
Expand a collection to see example tasks, study inputs, and what experts deliver.
Omics & Multi-omicsGene expression, cell states, and integration across molecular assays
- Study inputs
- Bulk or single-cell RNA, proteomics, and chromatin accessibility.
- Package contents
- Expert workflows for quality control, cross-modal integration, and resolving disagreement between data layers, with code, outputs, and scoring rubrics.
- Example task types
- Bulk RNA-seq. Separate differential expression from batch or cohort effects.
- Single-cell RNA-seq. Distinguish changes in cell abundance from changes within a cell type.
- RNA–protein integration. Explain discordant RNA and protein signals.
Drug Response & BiomarkersResponse biomarkers, acquired resistance, and agreement across assays
- Study inputs
- Molecular readouts, drug exposure, and measured response.
- Package contents
- Expert analyses of response signals and candidate biomarkers across pipelines or assays, with code, outputs, and evidence-strength rubrics.
- Example task types
- Response biomarkers. Test whether a baseline marker predicts measured drug response.
- Acquired resistance. Use paired samples to distinguish adaptation from a change in cell composition.
- Assay concordance. Decide whether a response signal holds across pipelines or assays.
Perturbation & Functional GenomicsCRISPR screens, genetic dependencies, and drug–gene interactions
- Study inputs
- CRISPR or chemical screens, perturbation metadata, and matched controls.
- Package contents
- Expert analyses for target attribution, dependency ranking, and separating sensitization from baseline growth effects, with code, outputs, and rubrics.
- Example task types
- CRISPR screen interpretation. Separate target-specific effects from guide artifacts or baseline growth.
- Context-specific dependencies. Test whether a dependency holds across models or is confined to a subtype.
- Drug–gene interactions. Distinguish drug sensitization from knockout toxicity.
Spatial TranscriptomicsTissue regions, cellular neighborhoods, and spatial expression
- Study inputs
- Spatial expression profiles, tissue coordinates, and histology where available.
- Package contents
- Expert analyses of tissue neighborhoods and spatial signals, including checks for cell composition and section effects, with code, outputs, and rubrics.
- Example task types
- Tissue domain annotation. Use expression and histology to identify tissue regions.
- Cellular neighborhoods. Test whether cell types co-localize beyond what their abundance would predict.
- Spatial differential expression. Check whether a regional signal persists after accounting for cell composition and tissue section.
Licensed post-training datasets include expert reference work. For held-out evaluation, models perform separate tasks in an evaluation environment. Expert reviewers use a rubric and a protected human reference to judge the model’s work.
Custom projects
Build around your research question.
Commission expert-authored tasks based on your own study or on proprietary or public study data cleared for the project. Define the capability, permitted use, deliverables, and acceptance criteria before production.
Explain acquired drug resistance across timepoints
Use paired pre- and post-treatment RNA data and response metadata. Identify programs that persist after treatment and specify a follow-up experiment to distinguish adaptation from a shift in cell composition.
Separate drug sensitization from knockout toxicity
Use CRISPR × drug perturbation data and matched controls. Expert analyses rank credible sensitizers and explain which hits disappear after accounting for baseline growth effects.
Distinguish a spatial signal from a cell-composition shift
Use spatial expression, tissue coordinates, and available histology to test whether a signal persists after accounting for the mix of cell types and the sampled tissue section.
Evaluate a bioinformatics agent on blinded multi-omics tasks
Experts design held-out tasks using authorized RNA, protein, and chromatin inputs. The model performs them in the evaluation environment. Expert reviewers score reproducibility and biological reasoning against human reference work kept outside the model workspace.
How Emergence works
From expert-authored tasks to model evaluation.
- 1. Assign the research brief. We match a domain expert to the research question, study data, and methods.
- 2. Create and complete the task. The expert writes the task, runs the reference analysis in ORIGIN, and records their code, outputs, and reasoning.
- 3. Review the task and reference. Independent experts check that the task is clear, the analysis reproduces, and the conclusions follow from the evidence. They review the scoring rubric before release.
- 4. Evaluate the model’s reasoning. The model performs the reviewed task in an evaluation environment with the human reference held back. Expert reviewers compare the model’s completed work with the human expert’s completion of the same task and score the quality of its evidence use, methods, and interpretation.