LORE-examples / README.md
jwkova's picture
Modality table: realistic atac example + context column header
9a7de41 verified
|
Raw
History Blame Contribute Delete
8.02 kB
metadata
license: mit
task_categories:
  - feature-extraction
tags:
  - biology
  - genomics
  - proteins
  - multimodal
  - mimic
pretty_name: LORE Examples
thumbnail: >-
  https://raw.githubusercontent.com/PolymathicAI/MIMIC/main/assets/MIMIC_logo.png
configs:
  - config_name: raw
    data_files: raw/examples.jsonl
    default: true
  - config_name: tokenized
    data_files: tokenized/examples.jsonl
default_config_name: raw

MIMIC

LORE Examples

A small set of matched multimodal examples from LORE, for the MIMIC model — enough to try inference, embedding, and generation across DNA, RNA, and protein modalities without wiring up your own data.

Each example is a single biological entity (a transcript and/or its protein) with several co-observed modalities. Rows are drawn from the held-out (validation) split of MIMIC's training data, so they are in-distribution and length-bounded to the model's context window.

Contents

Two parallel views of the same rows, exposed as two dataset configs:

Config Format Use
raw (raw/examples.jsonl) modality name → raw value (nucleotide/AA string, per-position track, …) model.input([row]) — the tokenizers run for you
tokenized (tokenized/examples.jsonl) tok_<modality> → token-id list model.input([row]) — pre-tokenized, so the heavy tokenizers (BioBERT text, ESM3 structure) don't run

raw is the default config, so load_dataset("polymathic-ai/LORE-examples") (no config name) loads it. Every row carries kind (rna / protein / both) plus uniprot_id / genome_feature_id anchors. Both views cover the same rows and the same modalities; tokenized just skips running the tokenizers (and the one-time ESM3 weight download for prot_struct).

Usage

from datasets import load_dataset
from mimic import load_pretrained

model = load_pretrained(version="1.0")

ANCHORS = ("kind", "uniprot_id", "genome_feature_id")

def to_sample(row):
    # drop anchor columns and modalities absent from this row (stored as null)
    return {k: v for k, v in row.items() if k not in ANCHORS and v is not None}

# raw view — modality name -> raw value; tokenizers run inside input()
raw = load_dataset("polymathic-ai/LORE-examples", "raw", split="train")
model.input([to_sample(raw[0])])
reps = model.embed()   # {"full": [B, N, D], "mod_ids": [B, N]}

# tokenized view — tok_ keys with pre-tokenized ids
# (skips running the BioBERT/ESM3 tokenizers; same rows as the raw view)
tok = load_dataset("polymathic-ai/LORE-examples", "tokenized", split="train")
model.input([to_sample(tok[0])])

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). Each row here populates a subset of these under its short name (raw view) or its tok_ key (tokenized view). The authoritative per-checkpoint list is model.modality_info. The Dtype column is the raw/decoded Python type; the tokenized config stores every modality as list[int]. 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.

Modality Track Dtype Description Example Conditioning context (context)
rna_seq nucleic str RNA/DNA nucleotide sequence (unspliced) — the core nucleic input "UUUGGAAACUUU…"
cds_junctions nucleic str Coding-sequence (CDS) exon–exon junction positions, per position "…0001000…"
splice_regions nucleic str Splice-region (exon) annotation, per position "…0011100…"
splice_jctns_5cls nucleic str Per-position 5-class splice-site type: 0=none, 1=acceptor, 2=donor, 3=TSS (first-exon start), 4=TES (last-exon end) "…00020…0100…"
is_coding nucleic list[int] Coding vs. non-coding flag [1]
feature_type nucleic list[str] Genomic feature-type label ['protein_coding']
phylop_human nucleic list[float] phyloP evolutionary-conservation score (human), per position [-0.66, 1.04, …]
phylop_mouse nucleic list[float] phyloP evolutionary-conservation score (mouse), per position [-0.26, -0.92, …]
atac nucleic str ATAC-seq chromatin-accessibility signal, per position (N = unmeasured). Cell-state-conditional "…N888887…" "human, GM23338 lymphoblastoid cell line (EBV-transformed B lymphocyte)"
cage nucleic list[float] CAGE transcription-start signal, per position. Cell-state-conditional [0.001, 0.001, …] "skeletal muscle, human, fetal"
rasp2 nucleic list[float] RASP2 (icSHAPE-style) RNA-structure reactivity, per position (nan where unmeasured). Condition-conditional [nan, 0.42, …] "technology: icSHAPE, reagent: NAI-N3, in vivo, cell line: K562, human"
aa_seq protein str Amino-acid (protein) sequence — the core protein input "MTPPERLFLP…"
rna_codons protein list[str] Codon sequence aligned to the protein (nucleotide content, protein-aligned track) ['AUG', 'ACA', 'CCA', …]
prot_struct protein biotite AtomArray Protein 3D structure as ESM3 VQVAE tokens (decode to a backbone via detokenize_structure) AtomArray (backbone)
dssp protein str DSSP secondary-structure class, per residue "CCXX…HHH…"
sasa protein list[float] Solvent-accessible surface area, per residue [225.1, 128.6, …]
prot_abund protein list[float] Protein abundance (PaxDb ppm), scalar. Cell-state-conditional [385.6] "Leptospira interrogans (bacterium), control"
funcprot_caption protein str Free-text protein functional caption "Catalyzes the hydrolysis of…"
masif_charge protein list[float] MaSIF surface Poisson–Boltzmann charge, per vertex [6.9, -3.3, …]
masif_hbond protein list[float] MaSIF surface hydrogen-bond potential, per vertex [-1.77, -1.64, …]
masif_hydrophobicity protein list[float] MaSIF surface hydrophobicity, per vertex [0.32, -0.31, …]
masif_si_index protein list[float] MaSIF surface shape-index, per vertex [0.34, 0.22, …]
masif_n_vertices protein list[float] MaSIF surface vertex count, per patch [65.0, 42.0, …]
context text str Free-text semantic context (e.g. cell-state / assay) — the conditioning channel itself "HepG2 cell line"
corpus text str Free-text biomedical literature (PubMed abstracts / articles), used as a language corpus "…regulates cell-cycle arrest and apoptosis…"
gene_family_txt text str Organism taxonomic lineage as free text (broad clade → phylum → class → order → family → genus → species) "metazoa chordata mammalia primates hominidae homo homo sapiens"

See also

Provenance & license

Derived from the validation split of MIMIC's training corpus (LORE). Released under the MIT license, matching the model code. See the MIMIC repository for details and citation.