MIMIC 1.0

MIMIC is a multimodal encoder–decoder foundation model of the central dogma, trained jointly over DNA, RNA, and protein together with a range of structural and functional tracks. A single model embeds any subset of modalities into a shared representation space and generates any modality conditioned on the others.

• Architecture: 20-layer encoder × 1536-d, 12-layer decoder × 1536-d, rotary position embeddings, mixed (uni-/bi-directional) attention, 5 register tokens.
• Parameters: ~1.25B.
• Weights: bfloat16, safetensors.
• Modalities (26): nucleotide/amino-acid sequence, codons, splice junctions/regions, coding annotation, conservation (phyloP human/mouse), protein structure tokens, DSSP, SASA, MaSIF surface features, protein abundance, RASP2 reactivity, and free-text functional/context channels.

Benchmarks & Transfer Learning

MIMIC's learned representations transfer to RNA and protein property prediction, where a single multimodal model is competitive with or ahead of strong single-modality foundation models:

• Protein — PFMBench (function, structure, interaction, and developability): MIMIC matches or exceeds protein-only baselines including ESM3, ESM-C, ProTrek, and SaProt, leading on at least 7 of 11 tasks — with especially strong protein–ligand binding (BindingDB, PDBbind) and the best results across all developability tasks.
• mRNA — mRNABench (function, localization, transcriptional regulation, and variant-effect prediction): MIMIC outperforms Evo 2 and Orthrus on 4 of 7 tasks and a dilated-ResNet baseline on 6 of 7, leading overall.

Beyond property prediction, the same model supports conditioned design — e.g. identifying corrective edits for a clinically relevant HBB splice-disrupting mutation, and generating diverse, high-confidence binder sequences by jointly conditioning on backbone shape and surface chemistry (PD-L1, hACE2).

See the paper for the full benchmark tables and comparisons.

Usage

from mimic import load_pretrained

# Downloads config.json + model.safetensors pinned to git tag v1.0.
model = load_pretrained(version="1.0")

# Real matched examples from LORE (raw view = modality name -> raw value).
from datasets import load_dataset
ex = load_dataset("polymathic-ai/LORE-examples", split="train")   # default config = raw
prot = next(r for r in ex if r["kind"] == "protein")              # has aa_seq + sasa
rna  = next(r for r in ex if r["kind"] == "rna")                  # has rna_seq + splice_jctns_5cls

# --- Embed ---
model.input([{"rna_seq": rna["rna_seq"]}])
reps = model.embed()                      # {"full": [B, N, 1536], "mod_ids": [B, N] per-token group ids}
#   reps["full"] -> (batch, num_tokens, 1536); pass return_register=True / return_modality=True for more

# --- Generate demo 1 (cross-modal): protein sequence -> per-residue solvent accessibility ---
# The target length is inferred from the co-grouped aa_seq; the default strategy is an
# Ensemble soft-vote (deterministic at low temp).
model.input([{"aa_seq": prot["aa_seq"]}])
out = model.generate("sasa")
print(out["sasa"])                                  # per-residue SASA (float array)

# --- Generate demo 2 (cross-modal): RNA sequence -> per-position splice-site classes ---
model.input([{"rna_seq": rna["rna_seq"]}])
out = model.generate("splice_jctns_5cls")
print(out["splice_jctns_5cls"])                     # one 5-class site label per position

# --- Richer output: raw arrays alongside the detokenized prediction ---
# By default generate() returns just the detokenized prediction. Set any of
# return_tokens / return_logits / return_probs / return_sampling_probs and each
# value becomes a dict with "preds" plus the extras you asked for (numpy arrays).
model.input([{"rna_seq": rna["rna_seq"]}])
out = model.generate("splice_jctns_5cls", return_probs=True, return_tokens=True)
print(out["splice_jctns_5cls"]["preds"])            # per-position class labels (as above)
print(out["splice_jctns_5cls"]["probs"].shape)      # (num_positions, num_classes)
print(out["splice_jctns_5cls"]["tokens"].shape)     # (num_positions,) predicted token ids

load_pretrained fetches only config.json + model.safetensors; tokenizers ship inside the mimic package.

Install: pip install git+https://github.com/PolymathicAI/MIMIC.git (imports as mimic). Source and docs: github.com/PolymathicAI/MIMIC.

Files

File	Description
config.json	Architecture + modality configuration consumed by load_pretrained.
model.safetensors	Model weights (bf16).

Example data

Ready-to-run examples (linked in the sidebar under Datasets):

• polymathic-ai/LORE-examples — a small set of matched multimodal rows (25 modalities) in both raw and tokenized form, for quickstart / embedding / generation.

from datasets import load_dataset
ds = load_dataset("polymathic-ai/LORE-examples")

Modalities

MIMIC represents each molecule as a set of co-observed modalities grouped into three tracks: nucleic (RNA/DNA and its per-position annotations), protein (amino-acid sequence, structure, and derived features), and text (free-text / categorical context). Pass any modality to model.input() under its short name (e.g. rna_seq) or pre-tokenized under its tok_ key. The authoritative, per-checkpoint list is model.modality_info; the Track column drives pathway gating (generation is limited to within-track and text-association pathways). A few assay tracks (atac, cage, rasp2, prot_abund) are context-conditional: pass a free-text context alongside them to condition on cell-state / assay metadata — the Conditioning context column shows a real example for each.

See the full modality table (with per-modality examples and conditioning contexts) on the LORE-examples dataset card.

Citation

If you use this work, please cite:

@misc{golkar2026mimicgenerativemultimodalfoundation,
      title={MIMIC: A Generative Multimodal Foundation Model for Biomolecules}, 
      author={Siavash Golkar and Jake Kovalic and Irina Espejo Morales and Samuel Sledzieski and Minhuan Li and Ksenia Sokolova and Geraud Krawezik and Alberto Bietti and Claudia Skok Gibbs and Roman Klypa and Shengwei Xiong and Francois Lanusse and Liam Parker and Kyunghyun Cho and Miles Cranmer and Tom Hehir and Michael McCabe and Lucas Meyer and Rudy Morel and Payel Mukhopadhyay and Mariel Pettee and Helen Qu and Jeff Shen and David Fouhey and Hadi Sotoudeh and Vikram Mulligan and Pilar Cossio and Sonya M. Hanson and Alisha N. Jones and Olga G. Troyanskaya and Shirley Ho},
      year={2026},
      eprint={2604.24506},
      archivePrefix={arXiv},
      primaryClass={cs.AI},
      url={https://arxiv.org/abs/2604.24506}, 
}

License

This model and its accompanying source code are released under the MIT License.