IA pour données omiques

Focus sur quelques apports

Mahendra Mariadassou

MaIAGE

September 17, 2026

Outline

AI methods have already made groundbreaking advances in biology:

  • Variant calling — pileups as images

  • Protein structure — from AlphaFold2 to protein Language Models and the “protein world”

  • Foundation Models — Evo2, AlphaGenome, and models that combine LLMs with gLM/pLM

  • Next-generation Foundation Models — new biological “tokens” (cells, tissues, microbiomes), and the open challenges

An explosion of AI models…

ChatGPT and LLM (Large Language Models) took the world by storm in 2022

  • In parallel, explosion of foundations models (FM) trained at scale on biological data :

    • 2018: DeepVariant [1]: CV-based variant calling

    • 2021: AlphaFold2 [2]: 3D structures from AA sequences (pre-FM); DNABERT [3]: first genome FM

    • 2023: DNABERT-2 [4], Geneformer [5]: Single-Cell FM

    • 2024: AlphaFold3 [6]: 3D structure of complexes (proteins / DNA / RNA / binders), Evo [7] long context genome FM, DNABERT-S [8], scGPT [9] Single-Cell FM, virtual cell manifesto [10]

    • 2025: Evo2 [11] improvement over Evo, [12] perspective of multimodal single-cell FM, virtual tissues [13]

    • 2026: protein world models (ESM Atlas and ESMC) [14], sequence-to-function model (AlphaGenome) [15], alignment/phylogeny-aware gLM for variant effect prediction (GPN-Star) [16]

    • etc.

Variant calling

Goal of variant calling

All images are copied from DeepVariant blog.

Images from DeepVariant blog

Principle of DeepVariant

Collaboration Google Brain / Verily Life Sciences, images from DeepVariant blog post

Impressive Results (PrecisionFDA)

Confirmed in several studies

Accuracy of SNV (A) and Indel (B) calling. Fig 1 from [17]

Accuracy of SNV (A) and Indel (B) calling. Fig 1 from [17]

Precision-recall curve of several caller on high (top) and low (bottom) datasets. DeepVariant is in blue and red. Adapted from Fig 2 of [18]

Precision-recall curve of several caller on high (top) and low (bottom) datasets. DeepVariant is in blue and red. Adapted from Fig 2 of [18]

Why it matters

Higher accuracy \(\rightarrow\) lower coverage \(\rightarrow\) less money 💰 for GBS

Image from DeepVariant blog

Image from DeepVariant blog

AI for variant calling

Deep neural networks can learn to call variants in pileup images with high accuracy and capture signals that are difficult for humans to identify.

Protein structure

A very short primer on proteins

  • Proteins responsible for functions essential to life.

  • Biological function depends on tertiary / quaternary structure

  • Understanding protein folding is essential

    • Incorrect folding can lead to diseases

    • Protein structure useful to develop drugs

From Wikipedia, created by user Kep17

From Wikipedia, created by user Kep17

Protein structure

Resolving protein structure is (or used to be) hard.

  • The first protein structure (that of myoglobin) was solved in 1958 (by X-ray crystallography) and refined in 1960 by John Kendrew (who received the Nobel Prize in 1962). He started working on this structure in 1949.

  • By 2021, there were 200 000 experimentally resolved proteins (0.1% of the >200M known proteins)

  • Huge need for structure prediction methods: see timeline in [19]

John Kendrew and the myoglobin structure in the PDB.

John Kendrew and the myoglobin structure in the PDB.

Main events in structure prediction. Fig. 1 from [19]

Main events in structure prediction. Fig. 1 from [19]

AlphaFold2 enters the ring

AlphaFold2 enters the ring

  • CASP13 (2018): Warm-up round

  • CASP14 (2020): Win by KO with “transformational” results.

  • AlphaFoldDB (2022): 200M+ protein structure prediction

  • AlphaFold3 [6]: 3D structures of complexes

Results from CASP14 (2020). Higher is better.

Results from CASP14 (2020). Higher is better.

Large protein models

Drawbacks: AlphaFold[2-3] can’t produce novel structures and requires MSA.

Protein Language Models (pLM) have been proposed to tackle this problem:

  • ESM2 [20], based on masked transformers for proteins.

    • attention patterns \(\leftrightarrow\) residue contact map
    • lower TM (Template Model) scores than AlphaFold2 but better scores than AlphaFold2 when using the AA sequence and not MSA
    • Similar accuracy (median all-atom RMSD of 1.91Å and a backbone RMSD of 1.33Å) for high confidence proteins

  • EMBER2 [21, 22], based on embeddings from a pLM (protT5) to predict 2D (inter residue distance) or 3D structure from a PLM trained on AA sequences rather than MSA.

    • Less accurate than AlphaFold2 but orders of magnitude faster
    • Outperforms AlphaFold2 for deep mutational scanning (impact of single AA change on structure)

Scaling up to the “Protein World”

Move over, AlphaFold: open-source model predicts shape of 1 billion proteins [23]

  • ESM Cambrian (ESMC) [14] trained on 2.8 billion protein sequences
    • use the latent space to predict folding (ESMFold2)
    • build an Atlas of 6.8 billion sequences and 1.1 billion predicted structures (ESM Atlas)
    • use SAE to extract \(\sim\) 16 000 interpretable features from representations

Protein world and atlas (bi[o]hub)

Compositional grammar for protein biology

Compositional grammar for protein biology

Highlighting one feature of the protein

Highlighting one feature of the protein

Highlighting another feature

Highlighting another feature

Filter using fine scaled features

Filter using fine scaled features

ESM Atlas

Use case: designing antibodies

Fraction of designed proteins that successfully bound their targets in the lab

Fraction of designed proteins that successfully bound their targets in the lab

AI for protein structure

Structure-prediction models (AlphaFold family) and protein language models (ESM, EMBER2) now cover both folding and design, moving us from “one structure at a time” to atlases of billions of predicted proteins.

Foundation Models

Motivations

From many tasks to one foundation model

  • Classical ML: one task / one dataset / one model

  • Foundation models (FM): a single pretrained model learns a generative grammar of biology from large-scale unlabeled sequences (DNA, RNA, protein), then adapts to many downstream tasks.

Why it matters ?

  • Transfer: knowledge learned from large datasets transfers to small (e.g. rare-disease) and unseen tasks — few-shot / zero-shot.

  • Generalization: one model spans many modalities & tasks (e.g. AlphaGenome [15]: expression, splicing, chromatin, 3D).

  • Variant effect prediction: models score coding and non-coding variants (e.g. AlphaGenome [15], GPN-Star [16]).

  • Efficiency & scalability: automated analysis of large-scale omics data, sparing experimental budgets.

  • Interpretability: learned representations expose biological structure (e.g. Evo 2’s prophage and secondary-structure detectors).

Evo 2: the new kid on the block

See Guillaume’s and Pierre’s talks

AlphaGenome

One model, many molecular outputs

  • Single sequence-to-function model from DeepMind [15]

  • Input: up to 1 Mb of raw DNA sequence

  • Output: 5,930 molecular tracks at (near) bp resolution over 11 modalities (gene expression, splicing, chromatin accessibility, contact maps, etc)

  • Variant Effect Prediction

Adapted from figure 1 of [15]

Unlike Evo2 (next-token prediction), AlphaGenome is supervised and trained on functional genomics tracks (regulatory-code specialist versus general sequence model).

AlphaGenome Atlas

Launched September 8, 2026 [24] precomputes AlphaGenome predictions genome-wide so nobody has to run the model themselves.

  • Every possible single-letter change (~9 billion SNVs, hg38) + >100M observed indels

  • ~1 petabyte of predictions (30\(\times\) bigger than AlphaFold DB)

  • AVI score (AlphaGenome Variant Impact):
    • one number per variant,
    • feature attribution: expression, splicing, chromatin…
    • use AlphaMissense for coding effects

AlphaGenome Atlas principle

Combining LLM and gLM / pLM

ChatNT: a multimodal model [25]

BioReason: “Chain of Thought” with DNA-LLM [26]

Foundation models are converging

FMs are no longer single-modality:

  • ChatNT and BioReason reason across DNA, RNA, protein and natural language.

  • Next: the “tokens” themselves are changing.

Next-generation Foundation models

Moving beyond sequences

1st gen: DNA, AA

  • gLM: DNABERTx, Evox, Nucleotide Transformer, AlphaGenome
  • pLM: ESMx

2nd gen:

  • Single-cell transcriptomics
  • Spatial omics
  • Multimodal data

Credits to OmicsML/awesome-foundation-model-single-cell-papers

Key evolutions (II)

Biological entities become tokens:

  • nucleotides 🧬
  • genes
  • cells 🦠
  • tissues 🫀
  • pangenome
    • genes = tokens
    • genomes = sentences
  • microbiome:
    • ASV/OTU/taxa = tokens
    • communities = sentences
    • ecosystems = corpus
      • MGM [27]
      • BiomeGPT [28]
      • Compass [29]

Foundation models are not magic 🪄

Challenges 💪

  • Experimental validation:
    • Real held-out XP data
    • Not (only) in silico metrics
  • Data leakage / contamination:
    • How to find held out data ?
    • High scores need checking
  • Interpretability:
    • Learned representations are opaque
    • Attribution (attention, SAE features, etc) must be sanity-checked against biology
  • Causal inference:
    • Correlative predictions
    • High score \(\neq\) mechanistic explanation

Challenges 💪

  • Reproducibility:
    • Not only weights
    • Data, seeds, and hyperparameters
  • Cost & availability:
    • Serious compute for huge models
    • Licensing can restrict use.
  • Dual Use / Biosecurity:

New actors 🏦

  • Alphabet DeepMind
  • CZ Bi[o]hub
  • Arc Institute

Conclusion

Main messages

Take Home Messages

  • AI methods have already changed some fields

    • variant calling
    • protein folding
    • variant effect prediction

  • FM for sequences 🧬 are being pumped out at a breakneck pace

    • New tool 🔧 in the toolbox 🧰
    • Huge potential for our applications
    • Heavy dose of skepticism is required
  • FM are
    • moving from sequences to cells, tissues (and other biological entities)
    • becoming multimodal

Practical advice 🔬

  • Treat FM predictions as hypotheses to test, not ground truth.

  • Report calibration: models can be confidently wrong.

  • Watch for mode collapse / memorization in generative tasks.

Additional slides: DeepVariant

Principle (I)

  • Consider pileup as an image

From DeepVariant blog post

From DeepVariant blog post

Principle (II)

  • works on images of size 100 \(\times\) 221 (bases)
    • 5 rows for the reference
    • (up to 95) reads
  • Use 6 “channels” (think RGB)
    • read-level features (MQ, strand, supports candidate)
    • base-level features
  • 2 additional channels (maps to alt. allele and haplotype)
  • In the original publication: ConvNetJuly2015v2 CNN with 9 partitions

Image from DeepVariant blog post

Image from DeepVariant blog post

Test it for yourself

Images from DeepVariant blog post

Homozygous alternate allele

Heterozygous

Harder isn’t it 🤔

Additional slides: Broken promise

The zero-shot benchmark illusion

Promise: Zero-shot scores \(\leftrightarrow\) measured fitness:

  • Good for improving proteins
  • High correlation in benchmarks [30]

Reality: check from protein engineers: best model, best mutations, … no improvement

Why? Benchmarks mix millions of mutations, mostly viable vs. dead. Zero shot =

  • good coarse filter, screens out dead mutations
  • bad ranker, can’t find the best mutation
  • True across architectures, across modalities

Fitness may be a bad predictor for some properties.

Figures source: Wilke, Jul 2026 and [30].

Why ? Lost in context

Problem

  • pLMs trained to fit sequence context \(\rightarrow\) penalize context-breaking mutations

  • Novel functions need destabilizing mutations \(\rightarrow\) rejected by pLM

  • PSSM affected by context \(\rightarrow\) favor frequent mutations

Solution ?

Combine PSSM minus pLM score [31]: \[ PSSM(sequence) \mathbf{-} pLM(sequence) \]

  • High pLM score \(\rightarrow\) stable, unlikely to have novel function

Context-free vs. context-aware scoring

Context-free vs. context-aware scoring

DraNramp:

  • pLM picks Mn2+ importers only
  • PSSM–pLM finds Mg2+ importers

Source: Wilke, Aug 2026, on [31]

References

1. Poplin R, Chang P-C, Alexander D, Schwartz S, Colthurst T, Ku A, et al. A universal SNP and small-indel variant caller using deep neural networks. Nature Biotechnology. 2018;36:983–7. doi:10.1038/nbt.4235.
2. Jumper J, Evans R, Pritzel A, Green T, Figurnov M, Ronneberger O, et al. Highly accurate protein structure prediction with AlphaFold. Nature. 2021;596:583–9. doi:10.1038/s41586-021-03819-2.
3. Ji Y, Zhou Z, Liu H, Davuluri RV. DNABERT: Pre-trained bidirectional encoder representations from transformers model for DNA-language in genome. Bioinformatics. 2021;37:2112–20. doi:10.1093/bioinformatics/btab083.
4. Zhou Z, Ji Y, Li W, Dutta P, Davuluri R, Liu H. DNABERT-2: Efficient foundation model and benchmark for multi-species genome. 2023. doi:10.48550/ARXIV.2306.15006.
5. Theodoris CV, Xiao L, Chopra A, Chaffin MD, Al Sayed ZR, Hill MC, et al. Transfer learning enables predictions in network biology. Nature. 2023;618:616–24. doi:10.1038/s41586-023-06139-9.
6. Abramson J, Adler J, Dunger J, Evans R, Green T, Pritzel A, et al. Accurate structure prediction of biomolecular interactions with AlphaFold 3. Nature. 2024;630:493–500. doi:10.1038/s41586-024-07487-w.
7. Nguyen E, Poli M, Durrant MG, Kang B, Katrekar D, Li DB, et al. Sequence modeling and design from molecular to genome scale with evo. Science. 2024;386. doi:10.1126/science.ado9336.
8. Zhou Z, Wu W, Ho H, Wang J, Shi L, Davuluri RV, et al. DNABERT-s: Pioneering species differentiation with species-aware DNA embeddings. 2024. doi:10.48550/ARXIV.2402.08777.
9. Cui H, Wang C, Maan H, Pang K, Luo F, Duan N, et al. scGPT: Toward building a foundation model for single-cell multi-omics using generative AI. Nature Methods. 2024;21:1470–80. doi:10.1038/s41592-024-02201-0.
10. Bunne C, Roohani Y, Rosen Y, Gupta A, Zhang X, Roed M, et al. How to build the virtual cell with artificial intelligence: Priorities and opportunities. Cell. 2024;187:7045–63. doi:10.1016/j.cell.2024.11.015.
11. Brixi G, Durrant MG, Ku J, Poli M, Brockman G, Chang D, et al. Genome modeling and design across all domains of life with evo 2. 2025. http://dx.doi.org/10.1101/2025.02.18.638918.
12. Cui H, Tejada-Lapuerta A, Brbić M, Saez-Rodriguez J, Cristea S, Goodarzi H, et al. Towards multimodal foundation models in molecular cell biology. Nature. 2025;640:623–33. doi:10.1038/s41586-025-08710-y.
13. Wenckstern J, Jain E, Vasilev K, Pariset M, Wicki A, Gut G, et al. AI-powered virtual tissues from spatial proteomics for clinical diagnostics and biomedical discovery. 2025. doi:10.48550/ARXIV.2501.06039.
14. Candido S, Hayes T, Derry A, Rao R, Lin Z, Verkuil R, et al. Language modeling materializes a world model of protein biology. 2026. http://dx.doi.org/10.64898/2026.06.03.729735.
15. Avsec Ž, Latysheva N, Cheng J, Novati G, Taylor KR, Ward T, et al. Advancing regulatory variant effect prediction with AlphaGenome. Nature. 2026;649:1206–18. doi:10.1038/s41586-025-10014-0.
16. Ye C, Benegas G, Albors C, Li JC, Prillo S, Fields PD, et al. Predicting genome-wide functional constraints with GPN-star. Nature. 2026. doi:10.1038/s41586-026-11005-5.
17. Barbitoff YA, Abasov R, Tvorogova VE, Glotov AS, Predeus AV. Systematic benchmark of state-of-the-art variant calling pipelines identifies major factors affecting accuracy of coding sequence variant discovery. BMC Genomics. 2022;23. doi:10.1186/s12864-022-08365-3.
18. Chen N-C, Kolesnikov A, Goel S, Yun T, Chang P-C, Carroll A. Improving variant calling using population data and deep learning. BMC Bioinformatics. 2023;24. doi:10.1186/s12859-023-05294-0.
19. Bertoline LMF, Lima AN, Krieger JE, Teixeira SK. Before and after AlphaFold2: An overview of protein structure prediction. Frontiers in Bioinformatics. 2023;3. doi:10.3389/fbinf.2023.1120370.
20. Lin Z, Akin H, Rao R, Hie B, Zhu Z, Lu W, et al. Evolutionary-scale prediction of atomic-level protein structure with a language model. Science. 2023;379:1123–30. doi:10.1126/science.ade2574.
21. Weissenow K, Heinzinger M, Rost B. Protein language-model embeddings for fast, accurate, and alignment-free protein structure prediction. Structure. 2022;30:1169–1177.e4. doi:10.1016/j.str.2022.05.001.
22. Weissenow K, Heinzinger M, Steinegger M, Rost B. Ultra-fast protein structure prediction to capture effects of sequence variation in mutation movies. 2022. http://dx.doi.org/10.1101/2022.11.14.516473.
23. Callaway E, Naddaf M. Move over, AlphaFold: Open-source model predicts shape of 1 billion proteins. Nature. 2026;654:13–4. doi:10.1038/d41586-026-01686-3.
24. DeepMind’s new genome ’atlas’ charts effects of all nine billion possible mutations. Nature. 2026. doi:10.1038/d41586-026-02835-4.
25. Almeida BP de, Richard G, Dalla-Torre H, Blum C, Hexemer L, Pandey P, et al. A multimodal conversational agent for DNA, RNA and protein tasks. Nature Machine Intelligence. 2025. doi:10.1038/s42256-025-01047-1.
26. Fallahpour A, Magnuson A, Gupta P, Ma S, Naimer J, Shah A, et al. BioReason: Incentivizing multimodal biological reasoning within a DNA-LLM model. 2025. https://arxiv.org/abs/2505.23579.
27. Zhang H, Zhang Y, Kang Z, Xiong J, Yang R, Ning K. MGM as a large-scale pretrained foundation model for microbiome analyses in diverse contexts. Advanced Science. 2026;13. doi:10.1002/advs.202513333.
28. Medearis NA, Zhu S, Zomorrodi AR. BiomeGPT: A foundation model for the human gut microbiome. 2026. http://dx.doi.org/10.64898/2026.01.05.697599.
29. Treloar NJ, Ur-Rehman S, Yang J. Learning the language of the microbiome with transformers. 2026. http://dx.doi.org/10.64898/2026.05.02.722381.
30. Woolley PR, Feller AL, Ellington AD, Wilke CO. Overestimating zero-shot fitness prediction: Broad benchmarks mask local failures and practical limitations. 2026. doi:10.64898/2026.06.04.730121.
31. Berry SP, Gaudet R, Marks DS. Differences between protein fitness models can be used to design variants of altered specificity. 2026. doi:10.64898/2026.06.10.731299.