BioMD: Multimodal AI in Biomedicine

• BioMD is a multifaceted term representing a family of AI systems in biomedicine that integrate molecular graphs, protein sequences, and textual data for cross-modal reasoning.
• It encompasses diverse models such as BioMedGPT for generative multimodal alignment, BioBridge for knowledge graph integration, and all-atom simulation for protein–ligand dynamics.
• BioMD methods offer practical insights into achieving unified biomedical representation, improved retrieval and simulation performance, and accelerated discovery in imaging and materials design.

BioMD is a context-dependent term in recent biomedical AI literature rather than a single universally standardized model name. In one usage, it denotes a biomedical multimodal foundation-model paradigm that attempts to align the “languages of life”—molecular structures, protein sequences, and eventually cellular and omics data—with human natural language, as operationalized by BioMedGPT (Luo et al., 2023). In another usage, BioMD is the explicit name of an all-atom generative model for long-timescale protein–ligand dynamics simulation (Feng et al., 2 Sep 2025). Closely related work extends the same thematic space to synthetic microscopy video generation and to AI-native biomedical materials discovery under uncertainty (Yilmaz et al., 2024, Mei et al., 20 May 2026). The term is also easily confused with adjacent systems that are not called BioMD; notably, BioMedLM explicitly “does not use the term ‘BioMD’,” but instead denotes a 2.7B-parameter biomedical LLM trained on PubMed text (Bolton et al., 2024).

1. Terminology, scope, and recurrent meanings

In the cited literature, “BioMD” functions as an umbrella for AI systems that operate over biomedical representations that are not reducible to ordinary prose. Those representations include molecules as graphs or strings such as SMILES or SELFIES, proteins as amino acid sequences and structures, cell or assay records, and atomistic trajectories. This suggests that BioMD is best understood as a family of computational programs centered on cross-modal reasoning, generative simulation, or closed-loop decision-making in biomedicine.

Usage in the literature	Representative system	Core object
Biomedical multimodal foundation model	BioMedGPT	molecules, proteins, natural language
KG-mediated multimodal bridging	BioBridge	frozen unimodal FMs linked by KG relations
All-atom biomolecular dynamics generation	BioMD	protein–ligand trajectories
Biomedical video diffusion model	BVDM	microscopy videos with pixel-level annotations
AI-enabled biomedical materials discovery	AIMBio-Mat	materials–biomedical closed-loop workflows
Adjacent but distinct textual model	BioMedLM	biomedical text only

A central misconception is that BioMD is interchangeable with any biomedical LLM. The BioMedLM paper explicitly clarifies that the model discussed is BioMedLM, not “BioMD,” and presents it as a GPT-style autoregressive decoder-only Transformer trained exclusively on PubMed abstracts and full articles (Bolton et al., 2024). By contrast, the BioMedGPT and BioBridge lines are explicitly multimodal, the BioMD dynamics line is atomistic and trajectory-generative, and AIMBio-Mat is a governance-aware decision layer rather than a single predictive model (Luo et al., 2023, Wang et al., 2023, Feng et al., 2 Sep 2025, Mei et al., 20 May 2026).

2. Biomedical multimodal foundation models

In the multimodal-foundation-model sense, BioMD is exemplified by BioMedGPT, an open multimodal generative pre-trained transformer designed to bridge the modality gap between biological representations and natural language (Luo et al., 2023). The motivating claim is that general-purpose foundation models often cannot directly interpret molecular graphs or protein sequences, and may produce out-of-order or meaningless outputs on tasks that require reasoning over these representations. BioMedGPT addresses this by combining a domain-specific LLM core with modality encoders and neural alignment into the LLM’s feature space.

The BioMedGPT family contains two principal components. BioMedGPT-LM-7B is a 7-billion-parameter Llama2-Chat-based biomedical LLM obtained via incremental training on 26 billion tokens of curated biomedical literature from S2ORC; the curation pipeline began from 5.5M biomedical papers and yielded 4.2M articles after removing non-full-text items and duplicates, with sentence-level chunking tokenized by the Llama2 tokenizer (Luo et al., 2023). BioMedGPT-10B is the multimodal instance composed of the 7B LLM backbone, a molecule encoder given by a 5-layer GIN pre-trained by GraphMVP with 1.8M parameters, a protein encoder given by ESM-2-3B, and two modality adaptors that map per-atom and per-residue outputs into the LLM token space.

The interface is deliberately generative. Molecules are represented as 2D graphs derived from RDKit processing, with each atom treated as a token; proteins are represented as amino acid sequences with each residue treated as a token. Role-based prompts inject aligned embeddings into the LLM context through tags such as <molecule><moleculeHere></molecule> and <protein><proteinHere></protein>, after which the system answers in free text. In practical terms, the projected per-atom and per-residue embeddings function as “virtual tokens” attended to by the frozen LLM during decoding (Luo et al., 2023).

Alignment is performed by supervised multimodal instruction-style QA training rather than by a contrastive loss. The paper states that the LLM is frozen during alignment, following mPLUG-Owl, in order to avoid catastrophic forgetting and reduce compute. The objective is the standard autoregressive language-modeling loss conditioned on modality features,

$L_{LM} = -\sum_t \log p(x_t \mid x_{<t}, z),$

where $z$ denotes the modality-conditioned context comprising aligned molecule or protein embeddings together with textual prompts and questions (Luo et al., 2023).

The multimodal alignment datasets are PubChemQA and UniProtQA. PubChemQA contains 325,754 unique molecules and 365,129 molecule-text pairs, with an average description length of 17 words and a QA schema of “Please describe the molecule.” UniProtQA contains 569,516 proteins and 1,891,506 QA samples across four categories—functions, official names, protein families, and sub-cellular locations—with an 8:1:1 train/validation/test split (Luo et al., 2023).

Evaluation results show why this line of work became a reference point for BioMD. On biomedical QA, BioMedGPT-10B reaches 51.4 on MedMCQA, 76.1 on PubMedQA, and 50.4 on USMLE; on PubMedQA it is described as on par with human experts at 78.0, and on USMLE it is the only open model in the comparison, besides ChatGPT, above 50% (Luo et al., 2023). On molecule QA using ChEBI-20, BioMedGPT-10B reports BLEU-2 0.234, BLEU-4 0.141, ROUGE-1 0.386, ROUGE-2 0.206, ROUGE-L 0.332, and METEOR 0.308, substantially above unaligned ChatGPT and Llama2-7B-Chat. On UniProtQA test data, BioMedGPT-10B with alignment reaches BLEU-2 0.571, BLEU-4 0.535, ROUGE-1 0.743, ROUGE-2 0.759, ROUGE-L 0.622, and METEOR 0.754, whereas unaligned Llama2-7B-Chat is described as producing essentially unusable outputs (Luo et al., 2023).

The strengths and liabilities are explicit. The strengths are unified cross-modal reasoning, strong biomedical QA, and flexibility obtained by freezing the LLM while fine-tuning modality encoders and adaptors. The limitations are hallucinations, opaque chain-of-thought, data coverage and modality imbalance, and the inadequacy of BLEU, ROUGE, and METEOR as complete measures of scientific correctness; expert evaluation is recommended (Luo et al., 2023).

3. Knowledge-graph bridging of unimodal biomedical foundation models

A second BioMD trajectory is represented by BioBridge, which seeks multimodal behavior without joint end-to-end fine-tuning of multiple large biomedical encoders (Wang et al., 2023). Its starting point is that biomedical foundation models have largely remained unimodal: protein models for sequences, molecular models for small molecules, and text models for biomedical or clinical language. BioBridge uses a biomedical knowledge graph to learn relation-aware transformations between independently trained embedding spaces while keeping the underlying encoders frozen.

The system uses ESM2-3B for proteins, UniMol for molecules, and PubMedBERT for text, with the training knowledge graph drawn from a curated subset of PrimeKG containing 85,647 nodes and 3,912,240 edges (Wang et al., 2023). Entities include proteins, drugs, diseases, and Gene Ontology terms; relations include protein–GO “interacts with,” drug–disease indication or contraindication, protein–disease association, and drug–protein target, enzyme, transporter, or carrier relations.

Mathematically, each unimodal embedding is projected to a common dimensionality by a modality-specific head $p_c$, and a learned relation-aware residual transformation $\psi$ maps the head embedding into the target modality space:

$\hat{h}_i = z_i + \psi(z_i, c_i, c_j, r_{ij}).$

Training uses InfoNCE over knowledge-graph triples,

$$\mathcal{L}_{ij} = - \log \frac{\exp(\hat{h}_i \cdot z_j / \tau)}{\exp(\hat{h}_i \cdot z_j / \tau) + \sum_{j^\prime \neq j}^M \exp(\hat{h}_i \cdot z_{j^\prime} / \tau)},$$

with negative tails sampled from the same modality type (Wang et al., 2023). The bridge module $\psi$ is a 6-layer transformer over the token stack $[z_i, c_i, c_j, r_{ij}]$, and only $\psi$ and the projection heads are trained.

The reported empirical effect is strong cross-modal retrieval. BioBridge ranks first across MRR and Hits@K on the reported tasks, and the paper states an average improvement over the best KG embedding baselines of around 76.3% (Wang et al., 2023). Representative MRR values are 0.136 for Protein→BP, 0.326 for Protein→MF, 0.319 for Protein→CC, 0.189 for Drug→Disease indication, and 0.172 for Protein→Drug target. It also reports semantic-similarity Spearman correlations of 0.91 for MF, 0.80 for BP, and 0.73 for CC, and demonstrates out-of-domain generalization in cross-species protein–phenotype retrieval and in PPI settings with relation types not present in the training KG (Wang et al., 2023).

Within the BioMD frame, the importance of BioBridge is architectural rather than nomenclatural. It shows that multimodal biomedical behavior can emerge from parameter-efficient bridging over frozen unimodal encoders, with the KG serving as weak supervision. This suggests a compositional route to BioMD that differs from the single-core generative alignment strategy of BioMedGPT.

4. All-atom generative BioMD for biomolecular dynamics

In a distinct and narrower sense, BioMD is the proper name of an all-atom generative model for biomolecular dynamics simulation, introduced as “the first all-atom generative model to simulate long-timescale protein-ligand dynamics using a hierarchical framework of forecasting and interpolation” (Feng et al., 2 Sep 2025). The survey “Learning Structure, Energy, and Dynamics” situates this model within a broader AI-for-protein-dynamics landscape organized around learning from structural ensembles and trajectories, learning from physical energy signals, and learning to accelerate simulations (Tang et al., 28 Apr 2026).

BioMD operates directly on all-atom Cartesian coordinates. For each frame $t$, the system uses $z$0 together with static conditioning $z$1 comprising the initial coordinates, amino acid sequence, and ligand atom types (Feng et al., 2 Sep 2025). The architecture combines an SE(3) Graph Transformer encoder over the initial conformation with a FlowTrajectoryTransformer that performs local-global-local attention over atoms and time. The reported key dimensions are $z$2, $z$3, $z$4, $z$5, $z$6, and $z$7 (Feng et al., 2 Sep 2025).

The generative mechanism uses “noise as mask.” For each frame,

$z$8

with ground-truth velocity

$z$9

Training minimizes

$p_c$0

augmented by geometric regularizers for ligand-center preservation, collision avoidance, and bond preservation (Feng et al., 2 Sep 2025).

The hierarchy is temporal. Forecasting subsamples the original trajectory every $p_c$1 frames to create a coarse sequence, and interpolation then reconstructs the frames between adjacent anchors. At inference, the coarse sequence can be generated either all-at-once or auto-regressively in blocks, such as AR-5; the interpolation stage conditions on clean anchors and generates intermediate frames from noise (Feng et al., 2 Sep 2025). The stated rationale is to avoid both averaging over multiple futures and error runaway during long rollouts.

The principal datasets are MISATO and DD-13M. MISATO contains almost 20,000 protein–ligand interaction trajectories, each with 100 frames from an 8 ns MD simulation. DD-13M contains 26,612 dissociation trajectories across 565 protein–ligand complexes, with average 480 frames per trajectory (Feng et al., 2 Sep 2025). Evaluation includes bond and angle MAE or MSE, steric clashes, RMSF correlations, unbinding path RMSD, and unbinding success.

On MISATO, BioMD-rel reports bond MAE 0.0395 Å, angle MAE 0.0655 rad, intra-ligand clashes 0.0003, protein–ligand clashes 0.0006, ligand RMSF correlation 0.4861, and protein RMSF correlation 0.5945; BioMD-abs reports stronger protein RMSF correlation at 0.6854 but higher clash and local-geometry error (Feng et al., 2 Sep 2025). On DD-13M, BioMD-rel (AR-5) reaches Success@1 70.88%, Success@5 92.95%, and Success@10 97.06%, while BioMD-abs (AR-5) reaches 56.76%, 74.19%, and 79.41%, respectively (Feng et al., 2 Sep 2025). The paper emphasizes that BioMD-rel generates complete ligand unbinding paths for 97.1% of systems within ten attempts, and reports a 6EY8 case in which BioMD produced a complete path in under 10 seconds whereas metadynamics required about one hour on a single GPU (Feng et al., 2 Sep 2025).

The broader survey highlights the unresolved criteria that govern this subfield: scalability, thermodynamic consistency, kinetic fidelity, and integration with experimental constraints (Tang et al., 28 Apr 2026). In that sense, BioMD is not merely a faster simulator; it is part of an ongoing attempt to learn structure, energy, and dynamics jointly.

5. Diffusion-based synthetic biomedical imaging and video

Another BioMD-adjacent usage concerns synthetic biomedical data generation for microscopy. The Biomedical Video Diffusion Model, or BVDM, is presented as a DDPM-plus-flow-field system that generates arbitrarily long synthetic microscopy videos with pixel-level annotations, trained only on a single real video (Yilmaz et al., 2024). This is not named BioMD in the paper title, but the provided description explicitly frames it as fitting the BioMD theme of biomedical diffusion-model-based data generation.

BVDM combines a denoising diffusion probabilistic model with a Flow Prediction Model implemented with VoxelMorph. The DDPM synthesizes realistic microscopy frames, while the flow component predicts dense non-rigid transformations between consecutive masks to preserve temporal consistency. The core equations are the standard DDPM forward and reverse processes,

$p_c$2

$p_c$3

together with the warping relation

$p_c$4

where $p_c$5 is the dense flow field between consecutive masks (Yilmaz et al., 2024).

The training data are HeLa cell microscopy sequences from the Cell Tracking Challenge: two grayscale videos of size 700×1100 with 92 frames each, one used for training and the other for validation, with silver-truth annotations (Yilmaz et al., 2024). Individual cells are extracted into 96×96 crops. The DDPM is trained with batch size 32, learning rate $p_c$6, and 10,000 iterations; the FPM uses batch size 32, learning rate $p_c$7, and 2,000 iterations on an NVIDIA GeForce RTX 3090 GPU (Yilmaz et al., 2024).

Quantitatively, BVDM reports SEG 0.829, TRA 0.979, FVD 79.4, and FID 29.6, compared with real-data training at SEG 0.823 and TRA 0.976, MitoGen at SEG 0.822, TRA 0.972, FVD 84.6, FID 49.6, and CellCycleGAN at SEG 0.781, TRA 0.953, FVD 350.4, FID 70.7 (Yilmaz et al., 2024). The recommended trade-off is $p_c$8 and $p_c$9. The reported interpretation is that sufficiently large synthetic datasets can improve downstream segmentation and tracking performance relative to limited real data.

Within the broader encyclopedia of BioMD, this line of work broadens the domain beyond molecules and sequences. It treats the biomedical object not as a linguistic or atomistic trajectory alone, but as a temporally coherent image process for which fully annotated synthetic data are themselves a central modeling target.

6. AI-native biomedical materials discovery and translation

A further extension of BioMD appears in AIMBio-Mat, which maps BioMD to AI-enabled biomedical materials discovery requiring simultaneous reasoning across composition, processing, structure, biological response, manufacturability, safety, and governance (Mei et al., 20 May 2026). The core proposition is not a single network architecture but an AI-native, FAIR, governance-aware decision layer linking existing materials and biomedical ecosystems.

The system is organized into five modular layers: ingest and validate; data and knowledge representation; AI and simulation; closed-loop discovery; and governance or user interaction (Mei et al., 20 May 2026). Its knowledge graph is formalized as

$\psi$0

and an AI-ready scientific record is represented as

$\psi$1

for composition or structure, processing, surface or microstructure, context, outcomes, uncertainty, governance, and provenance (Mei et al., 20 May 2026).

The optimization problem is explicitly multi-objective and constrained:

$\psi$2

Safe exploration is enforced through chance constraints,

$\psi$3

and the recommended acquisition rule is a constrained, cost- and safety-weighted EHVI:

$\psi$4

The platform is explicitly scoped to exploratory and preclinical decision support rather than clinical decision-support software (Mei et al., 20 May 2026).

The paper specifies a minimum viable prototype with AI-ready schemas, adapters for ELN or LIMS exports and instrument files, a knowledge layer, a versioned model registry, human-in-the-loop active learning, and immutable decision logs (Mei et al., 20 May 2026). Its pilot case concerns polymeric and hybrid nanoparticles for drug delivery, using an initial dataset of 40–60 batches with DLS size or PDI, zeta potential, morphology, drug loading, release kinetics, uptake, and viability. Success criteria include a reduction in experiments-to-target relative to expert-only and random or Latin-hypercube baselines, with example targets of 30–50% fewer experiments, metadata completeness of at least 90% required fields, improved calibration, and fewer safety violations (Mei et al., 20 May 2026).

This usage of BioMD shifts the emphasis from predictive models toward auditable, experimentally actionable infrastructure. It embeds uncertainty, negative results, access control, and governance fields as first-class objects rather than afterthoughts.

7. Limitations, controversies, and open technical problems

Across its meanings, BioMD is defined as much by unresolved constraints as by current capability. In multimodal language-of-life models, the central problems are hallucinations, interpretability, coverage gaps, and the fact that automatic text-generation metrics may not measure scientific correctness adequately (Luo et al., 2023). In KG-bridged systems, performance depends on KG coverage, relation balance, and the compatibility of frozen embedding geometries; sparse or biased graphs can skew alignment (Wang et al., 2023). In biomolecular dynamics generation, the outstanding challenges identified by the survey are scalability, thermodynamic consistency, kinetic fidelity, and principled integration with experimental constraints such as NMR, SAXS, or cryo-EM (Tang et al., 28 Apr 2026).

The same pattern appears in synthetic microscopy generation and biomedical materials discovery. BVDM learns texture and motion statistics from a single video, which constrains diversity and may underrepresent rare events or different imaging modalities (Yilmaz et al., 2024). AIMBio-Mat explicitly separates exploratory and preclinical prioritization from clinical or regulated-device use, and requires separate validation, change control, monitoring, and regulatory review for any clinical deployment (Mei et al., 20 May 2026).

A second misconception is terminological. BioMD is not coextensive with biomedical NLP, and not every biomedical model using “Bio” in its name belongs to the same lineage. BioMedLM, for example, is a 2.7B-parameter decoder-only Transformer trained exclusively on biomedical text, with strong results such as 57.3% on MedMCQA, 74.4% on PubMedQA, and 95.7% on BioASQ, but the paper explicitly frames it as BioMedLM rather than BioMD (Bolton et al., 2024). This distinction matters because text-only domain adaptation, multimodal alignment, atomistic dynamics generation, and governance-aware optimization solve different technical problems even when they inhabit the same biomedical AI ecosystem.

Taken together, the literature portrays BioMD not as a settled artifact but as a convergent research direction. One branch aligns molecules, proteins, and text in a shared generative interface; another learns relation-aware transformations across frozen biomedical encoders; another replaces expensive atomistic simulation with hierarchical generative dynamics; another generates annotated biomedical video from minimal supervision; and another formalizes biomedical materials discovery as constrained multi-objective optimization under uncertainty. This suggests that the most stable meaning of BioMD is methodological: AI systems that treat biomedical structure, dynamics, and decision processes as intrinsically multimodal, uncertainty-bearing, and experimentally consequential.