MIMIC: A Generative Multimodal Foundation Model for Biomolecules
Siavash Golkar, Jake Kovalic, Irina Espejo Morales, Samuel Sledzieski, Minhuan Li, Ksenia Sokolova, Geraud Krawezik, Alberto Bietti
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.
AI Impact Assessments
(3 models)Scientific Impact Assessment: MIMIC – A Generative Multimodal Foundation Model for Biomolecules
1. Core Contribution
MIMIC introduces a unified generative foundation model that jointly represents and reasons across heterogeneous molecular biology data spanning the central dogma — from DNA/RNA sequences through protein structures. The key intellectual contribution is the reframing of biological foundation modeling from siloed, forward-prediction tasks toward approximating the *joint distribution* of molecular states across modalities. This is accompanied by LORE, a curated multimodal dataset aligning ~15.5M proteins and ~13M RNA transcripts with evolutionary, structural, regulatory, and semantic/contextual annotations.
The split-track encoder-decoder architecture is the central architectural innovation: co-aligned features (e.g., phyloP conservation aligned to nucleotide positions) are summed within tracks, while distinct molecular entities are concatenated across tracks. This keeps sequence length tractable while preserving positional alignment. The asymmetric design (10K encoder, 1K decoder context) allows conditioning on broad molecular context while generating focused outputs.
2. Methodological Rigor
Architecture and training: The split-track summation, RoPE with group-reset, heuristic pathway sampling (~25 pathways), and staged curriculum (1K→10K tokens) are well-motivated engineering choices. The register token mechanism with random dropout for self-supervision is clever, though the paper could better analyze what information registers actually encode.
Evaluation breadth: The paper evaluates across an impressive range of tasks — sequence completion, protein benchmarks (PFMBench, 11 tasks), RNA benchmarks (mRNABench, 7 tasks), variant effect prediction, splice prediction, RNA design, protein design, and context-conditioned chemical probing. State-of-the-art claims are supported by comparison to strong baselines (ESM3, ESM-C, Evo2, AlphaGenome, SpliceAI, Orthrus).
Potential concerns: Several evaluation design choices warrant scrutiny. For splice prediction, all models are evaluated on MIMIC's test set, not necessarily held-out for other models — this could slightly disadvantage competitors. The protein design evaluation relies heavily on in silico metrics (AlphaFold2 pLDDT, AlphaFold3 iPTM) without experimental validation. For hACE2, only 2/20 high-confidence designs achieved iPTM > 0.75, which is modest. The RNA design experiment (HBB splice correction) uses SpliceAI as oracle but lacks experimental validation. The ablation study (Figure S3-S4) nicely demonstrates that multimodal conditioning helps most when auxiliary modalities are "ambiguous" — partially predictable from sequence — which is an insightful mechanistic finding.
3. Potential Impact
Unification paradigm: If MIMIC's approach scales, it could fundamentally change how the field approaches biological foundation models — moving from task-specific architectures to unified multimodal generative models. The ability to condition on arbitrary subsets of modalities and generate missing ones is a powerful abstraction.
Dataset contribution: LORE addresses a genuine infrastructure gap. Aligned multimodal biological datasets are scarce, and the community would significantly benefit from its release (currently pending).
Design applications: The constrained design demonstrations — RNA splice correction and protein binder design — showcase the inverse-problem capability that distinguishes this work from forward-prediction models. The HBB splice correction example, where MIMIC identifies compensatory edits without reverting the pathogenic mutation, is particularly compelling conceptually.
Variant effect prediction: The phyloP VEP approach, using a two-dimensional feature space (wild-type constraint + mutation-induced perturbation), dramatically outperforms embedding-based approaches (+82% on complex variants). This interpretable, low-dimensional approach could have immediate practical utility.
4. Timeliness & Relevance
This work addresses a clear bottleneck: biological AI remains fragmented across modalities, with most models treating biology as a collection of independent prediction problems. The multimodal approach aligns with the biological reality that function emerges from coupled constraints. The paper arrives at a moment when the field is saturated with single-modality models and recognizing their limitations. The semantic conditioning — using natural language descriptions of experimental context — is particularly timely given the growing recognition that biological measurements are context-dependent.
5. Strengths & Limitations
Key strengths:
Notable limitations:
Scalability questions: At 1B parameters, MIMIC is relatively modest. Whether the multimodal approach continues to outperform scaling-focused single-modality models at larger sizes remains an open question. The authors argue for "diverse data over pure scaling," but this hypothesis needs testing at scale.
Summary
MIMIC represents an ambitious and largely successful attempt to unify biological foundation modeling across the central dogma. Its strongest contributions are conceptual (the joint distribution framing), architectural (split-track encoder-decoder), and empirical (consistent SOTA across diverse benchmarks). The lack of experimental validation for design tasks and the pending release of code/data are significant limitations for assessing real-world impact. Nevertheless, the work establishes a compelling paradigm that could reshape how the field builds and evaluates biological AI systems.
Generated Apr 28, 2026
Comparison History (191)
MIMIC introduces a novel multimodal generative foundation model unifying representation learning, prediction, and constrained biomolecular design across genome, transcriptome, and proteome. It demonstrates state-of-the-art results and concrete therapeutic applications (splice correction, protein binder design), promising broad cross-field impact in biology and medicine. Paper 2 addresses an important security concern with a useful analysis tool (HalfLife), but is narrower in scope and primarily demonstrates feasibility of a threat vector. MIMIC's methodological breadth, real-world applications, and transformative potential in life sciences outweigh Paper 2's more focused contribution.
Paper 1 presents a unified multimodal foundation model bridging genomics, transcriptomics, and proteomics. Its ability to perform constrained design for clinical mutations and therapeutic protein binding offers profound real-world applications in drug discovery and molecular biology. While Paper 2 provides a rigorous, valuable systems-engineering solution for LLM agent reliability, Paper 1 demonstrates broader interdisciplinary impact, higher novelty in unifying disjoint biological modalities, and greater potential to catalyze fundamental scientific and medical breakthroughs.
MIMIC addresses a foundational challenge in biology by unifying multiple biomolecular modalities into a single generative model with demonstrated applications spanning representation learning, prediction, and design across genome, transcriptome, and proteome. Its clinical relevance (HBB correction, PD-L1/hACE2 binding design) and breadth of impact across biological subfields suggest transformative potential. Paper 1 offers a solid, well-motivated methodological contribution to asynchronous RL for LLM training, but its impact is narrower and incremental within an active but crowded optimization niche, whereas MIMIC could catalyze new directions in computational biology and therapeutic design.
Paper 2 has higher potential impact because it introduces a foundational multimodal generative model unifying sequence, structure, and evolution for biomolecules. This approach broadly advances computational biology, genomics, and drug discovery. While Paper 1 offers a rigorous, timely benchmark for LLM agents, its impact is primarily confined to AI evaluation and operations research. In contrast, Paper 2's capacity for clinical mutation correction and novel therapeutic protein design targets profound real-world medical applications, representing a paradigm shift in biomolecular design with a significantly wider scientific footprint.
MIMIC presents a generative multimodal foundation model unifying sequence, structure, regulatory, evolutionary, and contextual modalities for biomolecules—a significant architectural and methodological advance with broad applications in drug design, RNA editing, protein engineering, and splicing prediction. It achieves state-of-the-art on multiple benchmarks and demonstrates practical design capabilities (HBB mutation correction, PD-L1/hACE2 binder design). While Paper 1 offers valuable theoretical insights into RLHF limitations via rational inattention, its impact is more narrowly focused on AI alignment methodology. Paper 2's breadth across biology and medicine, combined with its novel multimodal framework, suggests higher overall scientific impact.
While Paper 1 offers a fascinating theoretical diagnostic for AI agents, Paper 2 demonstrates profound real-world applications in biomedicine and drug discovery. By unifying multiple biological modalities into a single generative foundation model, MIMIC achieves state-of-the-art performance in clinically relevant tasks, such as correcting splice-disrupting mutations and designing protein binders. Its breadth of impact across genomics, proteomics, AI, and medicine, combined with its ability to perform both predictive and generative biomolecular design, gives it a significantly higher potential for transformative scientific and societal impact.
Paper 1 (MIMIC) unifies biological modalities into a single generative foundation model, demonstrating profound real-world applications in clinical genetics and drug discovery. Its ability to identify corrective RNA edits and design protein binders for critical targets provides immediate, transformative value across the life sciences. While Paper 2 offers an innovative multi-agent framework for autonomous physical discovery, Paper 1's broader applicability across biology, medicine, and bioinformatics, coupled with its direct implications for human health, gives it a significantly higher potential for broad scientific and societal impact.
Paper 1 likely has higher impact due to greater breadth and immediacy of real-world applications: a multimodal generative foundation model spanning RNA/protein sequence, structure, regulation, evolution, and context, enabling SOTA prediction plus constrained biomolecular design (clinically relevant splicing repair, binding-site–conditioned protein design). Its novelty is amplified by a curated aligned dataset and arbitrary-subset conditioning, positioning it as reusable infrastructure across genomics, proteomics, and drug discovery. Paper 2 is methodologically strong and timely for inverse chaotic dynamics, but its cross-domain uptake is narrower and may remain specialized.
Paper 1 has higher potential scientific impact as it introduces a unifying multimodal foundation model for biology, a field where such models consistently drive massive paradigm shifts. By integrating sequence, structure, and evolutionary data, MIMIC enables both state-of-the-art prediction and constrained generative design for direct clinical applications, such as correcting RNA mutations and designing target-binding proteins. While Paper 2 offers a valuable methodological advance for AI computer-use agents, Paper 1's capacity to accelerate drug discovery and fundamentally advance computational life sciences presents a much more profound real-world impact across medicine and biology.
Paper 2 (MIMIC) is likely higher impact: it introduces a genuinely multimodal generative foundation model spanning sequence, structure, regulation, evolution, and context, enabling both prediction and constrained biomolecular design—clear real-world relevance to therapeutics and functional genomics. The aligned LORE dataset and split-track architecture broaden applicability across biology and ML, and reported state-of-the-art results plus concrete clinical/design case studies suggest strong utility. Paper 1 is valuable and rigorous for open agent data curation, but its contributions are more incremental and narrower to agent benchmark performance.