AI in Genomics

advanced

AI Tools in Genomics

AI/ML Applications in Genomics 2026

🤖AlphaMissense71M missense variants classified; 90% sensitivity; integrated in VarSome

🔮AlphaFold3Protein + nucleic acid + small molecule structure prediction; atomic accuracy

💡ESM-215B parameter protein language model; zero-shot functional prediction

✂️SpliceAICNN splice prediction; delta>0.5 = strong splicing evidence for ACMG PP3

📊CADD v1.7Ensemble 100+ features; PHRED score; >20 = top 1% deleterious variants

🧬Nucleotide Transformer500M-2.5B params; genomic foundation model; variant effect fine-tuning

Foundational Models

AlphaFold2 (DeepMind, 2021): predicted structures for 200M+ proteins with atomic accuracy; revolutionised structural biology
AlphaFold3 (2024): extends to nucleic acids, small molecules, and protein complexes; critical for drug design
ESM-2 (Meta, 2022): 15B parameter protein language model trained on 250M sequences; zero-shot functional prediction
Nucleotide Transformer (InstaDeep, 2023): 500M–2.5B parameter genomic foundation model; fine-tuned for variant effects
GeneFormer (Harvard, 2023): trained on 30M single cells; predicts gene network dynamics; drug target discovery
scGPT (2024): single-cell foundation model; cell type annotation, perturbation prediction

Variant Effect Prediction

AlphaMissense (Google DeepMind, 2023): classifies 71M missense variants; 90% sensitivity at 90% specificity; integrated in VarSome/ClinGen
CADD v1.7: combines 100+ annotations; PHRED-scaled score; >20 = top 1% deleterious
REVEL: ensemble of 13 tools; best AUC for distinguishing P/B variants in ClinVar benchmarks
SpliceAI (Illumina): deep learning CNN; predicts splice donor/acceptor changes within 50 bp
AbSplice: tissue-specific splicing predictions; integrates SpliceAI + tissue expression data
PrimateAI-3D: uses primate evolution + 3D protein structure; strong for missense interpretation

Cancer AI Applications

DITTO: pan-cancer driver mutation prediction; outperforms COSMIC census for novel genes
Trained on TCGA: 33 cancer types, 10,000+ WES tumours; mutation signatures + driver genes
Foundation models for pathology (CONCH, UNI): predict genomic alterations from H&E slides
DeepMEL: melanoma survival prediction from mutation profile
MutSig2CV: identifies significantly mutated genes beyond background mutation rate
MMSig: mutational signature deconvolution; identifies APOBEC, UV, MMR deficiency signatures

Using AI Tools in Practice

code

# Query AlphaFold structure database
import requests

def get_alphafold_structure(uniprot_id):
    url = f"https://alphafold.ebi.ac.uk/files/AF-{uniprot_id}-F1-model_v4.pdb"
    r = requests.get(url)
    if r.status_code == 200:
        with open(f"{uniprot_id}.pdb", "w") as f:
            f.write(r.text)
        return f"{uniprot_id}.pdb"
    return None

# BRCA1 structure
get_alphafold_structure("P38398")

# Query AlphaMissense for a specific variant
# Via Ensembl REST API (includes AlphaMissense)
import requests
server = "https://rest.ensembl.org"
ext = "/vep/human/hgvs/ENST00000357654.9:c.5266dupC"
response = requests.get(server+ext,
  headers={"Content-Type":"application/json"})
print(response.json()[0]["most_severe_consequence"])

Clinical Genomics Pipeline Learning Roadmap