Preserve modality-specific structure.
Respect spatial, temporal, sequence, and experimental relationships before aligning representations.
Case study · AI for science
Multimodal evidence across scales.
Scientific questions rarely live in one format. We explore systems that connect sequences, measurements, images, text, and experimental context while keeping provenance and uncertainty attached to every synthesis.
- Status
- Research direction
- Central question
- How can models connect evidence across modalities without erasing what makes each measurement distinct?
- Evidence standard
- Task-specific validation, strong baselines, provenance, and expert review.
01 · Question
Can one representation respect many kinds of evidence?
A biological observation can look different at the molecular, cellular, tissue, and organism level. Its meaning also depends on experimental design, controls, and measurement technology. Joining those views can make a problem more tractable, but only if the system preserves their context.
The goal is not to flatten diverse data into one confident answer. It is to help organize relationships, identify missing evidence, and generate testable next questions.
02 · Approach
Connect modalities without losing provenance.
Use shared context carefully.
Connect related observations while keeping sample, assay, condition, and source metadata available.
Separate retrieval from inference.
Make clear which conclusion follows from supplied evidence and which is a model-generated hypothesis.
Return questions to experiment.
Prioritize outputs that can be checked against held-out measurements, perturbations, or expert review.
03 · Evidence
Scientific usefulness needs task-specific proof.
This page describes broad research interests and does not announce a validated scientific or clinical result. Evidence should be matched to a clearly scoped question, evaluated on data separated by the relevant biological and experimental boundaries, and compared with meaningful baselines.
Held-out biological context
Split data in ways that test the claimed generalization—not merely random records from the same context.
Value beyond simpler methods
Compare against domain tools, retrieval, and unimodal models before attributing gains to multimodality.
Calibration and abstention
Measure whether confidence tracks reliability and whether the system can defer when evidence is thin.
Traceable inputs and outputs
Keep data sources, transformations, and the basis for a generated hypothesis inspectable.
04 · Limitations
A coherent story can still be wrong.
Scientific datasets contain batch effects, confounding, missing context, and uneven representation. A model can produce plausible connections without causal support. Performance on one dataset does not establish transfer across populations, laboratories, assays, or biological scales.
Use boundary. The work described here is exploratory. It is not medical advice, a diagnostic system, a substitute for qualified scientific judgment, or a claim of validated discovery.