BioVERSE

BioVERSE tackles a structural problem in biomedical AI: large language models and specialized biomedical foundation models (BioFMs) both encode powerful knowledge, but they live in disjoint embedding spaces. A single-cell model, a protein model, and a text LLM each represent their inputs differently, so a system cannot easily reason jointly across a cell's expression profile, a protein, and natural-language questions about them. BioVERSE bridges these silos by aligning the representations of frozen BioFM encoders into the embedding space of a language model, so that heterogeneous biological data can be reasoned over through a common interface.

The approach is deliberately lightweight. Each modality is handled by a pretrained BioFM used as a frozen encoder, and a small modality-specific projection layer maps that encoder's outputs into the LLM's input space. These projections are trained independently to align each modality to the shared space, after which standard instruction tuning on multimodal data brings the aligned modalities together for downstream reasoning. Because the interface is embedding-based, BioVERSE is LLM-agnostic — the authors demonstrate it with IBM's open-weights Granite-3.3-8B-Instruct, but any model that accepts embedding inputs can be used without architectural change. BioVERSE was developed by IBM Research and released as a preprint in October 2025.

The result is a modular way to give a language model native access to raw biomedical modalities, enabling zero-shot annotation, cross-modal question answering, and interactive, explainable dialogue over biological data.

#Key Features

• Alignment of frozen BioFM encoders: Reuses existing biomedical foundation models as frozen encoders, adding only small projection layers rather than retraining large models.
• Two-stage training: First aligns each modality to the LLM space through independently trained projections, then unifies them with multimodal instruction tuning.
• LLM-agnostic interface: Works with any LLM that accepts embedding inputs, demonstrated on IBM Granite-3.3-8B-Instruct with no architectural modification.
• Zero-shot cross-modal reasoning: Supports zero-shot cell annotation, cross-modal question answering, and interactive, explainable dialogue across single-cell, molecular, and protein inputs.

#Technical Details

BioVERSE — Biomedical Vector Embedding Realignment for Semantic Engagement — adapts pretrained BioFMs as modality encoders and connects them to an LLM through lightweight, modality-specific projection layers. Training proceeds in two stages: per-modality alignment, in which each projection is trained to map its encoder's embeddings into the LLM space, followed by standard instruction tuning on multimodal data that lets the aligned modalities interoperate for reasoning. The encoders remain frozen, so the trainable footprint is confined largely to the projections and tuning, keeping the method modular and efficient. The reference implementation uses the 8B-parameter Granite-3.3-8B-Instruct LLM and spans single-cell, molecular, and protein modalities. The model is released under a CC BY-SA 4.0 license, and the work is a preprint awaiting peer review.

#Applications

BioVERSE is intended for researchers who want to query heterogeneous biological data through natural language — for example asking a language model to annotate cell types from an expression profile, answer questions that span a protein and its context, or hold an explainable dialogue that draws on multiple modalities at once. Because it wraps existing BioFMs rather than replacing them, groups can extend the framework to new modalities by training an additional projection, making it a flexible substrate for building biomedical assistants and analysis tools.

#Impact

By showing that frozen biomedical encoders can be aligned to an LLM through small projections and light instruction tuning, BioVERSE offers a scalable, modular path to multimodal biomedical reasoning that avoids retraining large models from scratch. Its LLM-agnostic, encoder-agnostic design positions it as connective infrastructure between the growing ecosystem of specialized BioFMs and general-purpose language models. As a recent preprint, its downstream adoption and independent benchmarking are still developing.