MIMIC: A Generative Multimodal Foundation Model for Biomolecules

Abstract: Biological function emerges from coupled constraints across sequence, structure, regulation, evolution, and cellular context, yet most foundation models in biology are trained within one modality or for a fixed forward task. We present MIMIC, a generative multimodal foundation model trained on our newly curated and aligned dataset, LORE, linking nucleic acid, protein, evolutionary, structural, regulatory, and semantic/contextual modalities within partially observed biomolecular states. MIMIC uses a split-track encoder-decoder architecture to condition on arbitrary subsets of observed modalities and reconstruct or generate missing components of molecular state across the genome, transcriptome, and proteome. Multimodal conditioning consistently improves MIMIC's sequence reconstruction relative to sequence-only inputs, while its learned representations enable state-of-the-art performance on RNA and protein downstream tasks. MIMIC achieves state-of-the-art splicing prediction, and its joint generative formulation enables isoform-aware inference that further improves performance. Beyond prediction, the same generative framework supports constrained design. For RNA, MIMIC identifies corrective edits in a clinically relevant HBB splice-disrupting mutation without reverting it by using evolutionary and structural signals. For proteins, jointly conditioning on shape and surface chemistry of PD-L1 and hACE2 binding sites produces diverse, high-confidence sequences with strong in silico support for target binding. Finally, MIMIC uses experimental context as semantic conditioning to model assay-dependent RNA chemical probing, rather than treating context as a fixed output. Together, these results position MIMIC's aligned multimodal generative modeling as a strong foundation for unifying representation learning, conditional prediction, and constrained biomolecular design within a single model.

• The paper introduces a split-track encoder-decoder transformer that integrates diverse biological data for joint inference and conditional modeling.
• It leverages the comprehensive LORE dataset, containing over 15M proteins and 13M RNA transcripts, to achieve superior sequence reconstruction and property prediction.
• MIMIC demonstrates versatile applications, including RNA splicing correction, variant effect prediction, and protein binder design, enhancing practical biomolecular engineering.

MIMIC: A Generative Multimodal Foundation Model for Biomolecules

Introduction: Multimodal Representation in Biological Modeling

The MIMIC framework addresses the longstanding fragmentation in computational biology by unifying sequence, structure, regulation, evolution, and contextual modalities into a single generative modeling paradigm. Biological function arises from intricate interdependencies, yet most foundation models in molecular biology are siloed within specific modalities or trained for fixed forward tasks. MIMIC overcomes these constraints by leveraging the newly curated LORE dataset, which aligns heterogeneous genomic, transcriptomic, proteomic, evolutionary, and regulatory measurements, enabling joint inference and conditional modeling across partially observed biomolecular states (Figure 1).

Figure 1: Overview of MIMIC architecture, LORE dataset construction, and multimodal capabilities across molecular biology.

Model Architecture: Split-Track Encoder-Decoder Design

MIMIC utilizes a split-track encoder-decoder transformer capable of ingesting arbitrary subsets of observed modalities. Biological signals sharing a coordinate system are embedded and summed into a unified track representation, while unaligned modalities—including semantic and experimental context—are tokenized separately. Rotary Position Embeddings with local group-reset ensure within-track attention and tractable sequence length for very large contexts. A cross-attention decoder reconstructs or generates target modalities conditioned on observed context, supporting tasks such as protein design from surface geometry, isoform-aware splicing prediction, regulatory inference, and context-conditioned RNA structure modeling (Figure 2).

Figure 2: Split-track input construction and encoder-decoder backbone for MIMIC, efficiently managing multimodal biological heterogeneity.

Dynamic workload balancing addresses the heavy-tailed biological sequence distribution, optimizing memory efficiency via bucketed batching without truncation or entity mixing (Figure 3).

Figure 3: Computational workload balancing strategies for variable-length biological sequence data.

LORE Dataset: Massive, Aligned Multimodal Corpus

LORE comprises ~15.5M proteins and 13M RNA transcripts across over 6,000 organisms, integrating >4B tokens of biomedical text. Each row captures unified, partially observed molecular states, preserving biological alignment amidst inherent measurement sparsity. Protein tracks include sequence, structure (AlphaFoldDB), surface chemistry (MaSIF), functional annotations (UniProt), and abundance (PaxDB). RNA tracks consolidate sequence, splicing, chemical probing (RASP2), evolutionary conservation (phyloP), regulatory assays (ATAC/CAGE), and semantic descriptions (PubMed, tissue/cell line metadata). This comprehensive multimodal substrate is foundational for training models on joint molecular distributions rather than isolated observations (Figure 1B).

Sequence Completion: Multimodal Conditioning and Reconstruction

MIMIC demonstrates consistently superior sequence reconstruction accuracy across masked inpainting tasks for both protein and RNA, outperforming strong sequence-only baselines such as ESM3, ProtBERT, NTv3, and Evo2 when conditioned on structural and auxiliary modalities (Figure 4). Ablations reveal that each modality introduces non-redundant information; performance uplifts are greatest where modality predictability is ambiguous, indicating effective conditional narrowing of the solution space (Figures 8, 9).

Figure 4: State-of-the-art performance for property prediction benchmarks; MIMIC outperforms other models across diverse RNA and protein tasks.

Figure 5: Ablation analyses show substantial accuracy gains from multimodal context in sequence completion tasks.

Figure 6: Uplift in inpainting accuracy correlates with partial predictability of auxiliary modalities, confirming their non-redundant constraint.

Downstream Applications: RNA and Protein Property Prediction

Using embedding probes, MIMIC achieves state-of-the-art results on PFMBench (function, structure, interaction, developability) and mRNABench (function, localization, translation-related tasks). Notably, MIMIC excels on protein-ligand binding and translation prediction for endogenous human isoforms, outperforming specialized models despite its parameter economy (Figure 2D, 2E).

Variant Effect Prediction: Context-Aware Conservation Modeling

MIMIC integrates phyloP conservation as an explicit output, enabling interpretable and highly predictive variant effect benchmarks. Conditioning with phyloP substantially improves wild-type allele recovery, and mutation-induced conservation perturbation (phyloP VEP) provides robust discriminative signal especially where background conservation is weak—complementing traditional evolutionary proxies (Figure 7).

Figure 7: PhyloP VEP scores from MIMIC offer strong, complementary signal to raw conservation for variant effect prediction.

Splicing Prediction and RNA Design: From Forward Modeling to Generative Correction

MIMIC advances splice site prediction by outperforming SpliceAI, AlphaGenome, and NTv3 in both gene-level and transcript-level tasks, with larger gains in non-coding regions. Conditioning on transcript boundaries (TSS/TES) further boosts accuracy and specificity in isoform selection. The generative formulation enables design of corrective RNA edits; for the HBB IVS-II-654 C>T pathogenic variant, MIMIC designs sequence windows that suppress cryptic exon inclusion by strategic rewiring of splicing patterns and conservation profiles (Figure 8).

Figure 8: Splice site prediction and corrective RNA design; MIMIC re-designs sequences to abrogate inclusion of pathogenic cryptic exons.

Figure 9: Generative splicing design with customizable nucleotide window sizes.

Protein Binder Design: Multimodal Conditioning for Diverse, Functional Sequences

MIMIC enables constrained protein sequence generation, targeting therapeutically relevant complexes (e.g., PD-L1, hACE2). Joint conditioning on backbone and MaSIF surface features yields diverse sequences that recover wild-type fold, surface chemistry, and binding competence as measured by TM-score and AlphaFold iPTM. Notably, generated sequences retain high structural fidelity but only ~50% sequence identity to wild-type, reflecting efficient exploration of non-trivial fitness landscapes (Figure 10).

Figure 10: Protein binder designs demonstrate high confidence, structural fidelity, and sequence diversity across conditioning schemes.

Contextual RNA Reactivity and Structure Modeling

MIMIC ingests semantic experimental context (natural language descriptions of assay, cell line, reagent) to predict RNA chemical reactivity (RASP2), adapting output to in vivo vs. in vitro conditions and probe type. Context-aware modeling improves RNA secondary structure prediction—incorporating MIMIC predictions as SHAPE constraints yields folds nearly identical to experimentally-guided references, outperforming sequence-alone baselines by significant $F_1$ margins (Figure 11).

Figure 11: Condition-specific RNA reactivity prediction enables accurate context-dependent RNA structure modeling.

Discussion and Implications

MIMIC delivers notable advances by aligning multimodal context within a generative encoder-decoder architecture, demonstrating that heterogeneous biological measurement integration drives consistent uplift in performance and enables versatile applications beyond forward prediction—including conditional inference, variant effect interpretation, and inverse design for both nucleic acids and proteins. The explicit modeling of semantic context and experimental state moves beyond rigid output schemas, facilitating practical translation to RNA therapeutics and protein engineering. These findings indicate that multimodal context can serve as stabilizing constraints, reducing the search space for design and enhancing model interpretability.

Limitations remain regarding incomplete modality coverage (e.g., chromatin, interaction networks) and decoder context scale. Future developments will entail broadening modality representation, improving alignment quality, expanding context lengths, and refining semantic representations to capture richer experimental states. Longer, multiscale architectures and structured semantic conditioning are anticipated to further enhance conditional generative capacity and downstream biological insight. The paradigm established by MIMIC positions multimodal joint modeling as a robust foundation for unified cellular biomolecule representation—a critical step toward truly generalist biological AI.

Conclusion

MIMIC exemplifies a coherent multimodal generative modeling strategy for molecular biology, delivering robust performance across diverse benchmarks, supporting flexible conditional inference, and enabling constrained design in both RNA and protein contexts. The combination of a split-track architecture, dynamic context conditioning, and rigorous dataset alignment offers a template for future advances in biological foundation modeling, expanding the practical and theoretical scope of AI-driven discovery in life sciences.