Drug target identification & biomarker discovery from multi-omics (incl. single-cell foundation models)

Most drug programmes fail because the chosen target is wrong, not because the molecule is poor. AI platforms combine omics data, human genetics, knowledge graphs and literature to rank candidate targets, predict how cells respond when a gene is perturbed (virtual-cell and single-cell foundation models) and propose biomarkers that identify the patients most likely to respond. The output is a prioritised hypothesis list that scientists then test in the lab, before any chemistry (SCI-002) starts.

Problem & value

• Today: target selection relies on expert curation of papers, public databases (Open Targets, GWAS catalogues) and in-house omics, analysed by bioinformaticians with R/Python pipelines (Seurat, Scanpy, DESeq2). It is slow, biased towards well-studied genes and hard to reproduce. Biomarker work suffers from small cohorts, batch effects and overfitting.
• AI value: integrate many evidence types at once, surface under-studied targets, simulate perturbations cheaply before costly CRISPR screens, and prioritise which experiments to run.
• Evidence it can work: Insilico Medicine's PandaOmics identified TNIK as a target in idiopathic pulmonary fibrosis; the resulting molecule rentosertib reported Phase 2a results in Nature Medicine (2025), an often-cited proof of concept for AI-nominated target plus AI-designed molecule [1]. BenevolentAI's knowledge-graph analysis proposed baricitinib for COVID-19, later supported by clinical trials [2]. Human-genetic support for a target is associated with higher clinical success in published analyses [3]; treat headline "AI speed-up" claims from vendors with caution, since the decisive validation remains wet-lab and clinical.

Variants

• Target identification and prioritisation: multi-evidence ranking (expression, genetics, pathway, literature, tractability, safety liabilities).
• Single-cell and spatial analysis with cell foundation models (main method variant): models such as scGPT, Geneformer, scFoundation, UCE and Arc Institute's STATE embed cells, annotate cell types, integrate batches and predict perturbation responses ("virtual cell") [4][5].
• Drug repurposing: knowledge-graph reasoning over drug-target-disease links.
• Biomarker and patient stratification: signatures from tumour, blood or tissue omics for companion diagnostics and trial enrichment (see HLTH-009).
• Academic / public research: the same open tooling on consortium data, without proprietary platforms.

How it works

data intake (omics, Perturb-seq, GWAS, literature, KG)
  -> QC, normalisation, batch integration (scVI, Harmony)
  -> representation learning (cell foundation models, gene embeddings)
  -> evidence integration (knowledge graph + genetics + expression + LLM literature mining)
  -> target scoring & ranking (tractability, safety, novelty)
  -> in silico perturbation / virtual-cell prediction + causal checks
  -> biomarker signature discovery with nested cross-validation
  -> expert review (target dossier) -> wet-lab validation (CRISPR, organoids)

• Cell foundation models are pretrained on tens of millions of single-cell profiles and fine-tuned for annotation or perturbation prediction. Independent benchmarks show that for perturbation prediction they often do not clearly beat simple baselines (e.g. mean or linear models), so a baseline comparison is mandatory [6].
• Knowledge-graph reasoning (graph neural networks, link prediction) proposes target-disease and drug-disease links; LLMs extract and summarise evidence from papers with citations.
• Human-in-the-loop: biologists review every ranked target, inspect the evidence trail and decide which hypotheses enter experimental validation. No result is a finding until validated in the lab.

Data preparation

Data sit in many places: omics matrices and raw FASTQ files on institute file shares or HPC storage, clinical metadata in LIMS, ELN or study databases, literature in PDFs and reference managers, and legacy results in Excel. Typical gaps are inconsistent gene and sample identifiers, missing or free-text clinical annotation, batch effects between cohorts, mixed protocols, scattered ELN notes and unclear consent per sample.

Preparation steps: inventory datasets and owners (PREP-001, PREP-004); map sample IDs to a common ontology and harmonise metadata (PREP-010, PREP-012); export clinical and ERP-style data cleanly (PREP-005); pseudonymise patient-level data before it leaves the secure zone (PREP-022); check purpose and legal basis for reuse (PREP-025); build held-out evaluation sets with domain experts (PREP-018). Tools: Nextflow/nf-core for reproducible QC, AnnData/h5ad as common format, scVI or Harmony for batch integration, and an ontology service for cell types and diseases. See DATA-003 for data cataloguing and GUIDE-003 for a readiness check.

Expect 50-70 percent of total effort in curation, harmonisation and governance, far more than model work.

Implementation paths

Which path for whom

• Enterprise: B - managed EU-region compute with in-house pipelines keeps proprietary patient and target data under own control while scaling elastically; platform licences (A) are a fallback for teams without data scientists.
• Developer: C - open single-cell models and scverse tooling are mature and allow baseline comparison and fine-tuning.
• Public: C - academic and consortium work uses open data and open tooling, which also satisfies reproducibility and procurement requirements.

Path A - Buy

• Target discovery platforms: Insilico Medicine PandaOmics, Open Targets Platform (free, EMBL-EBI, EU-based), Recursion platform, Owkin (France, EU; federated learning, biomarker work), Relation Therapeutics (UK), Turbine (Hungary/US, simulated cell models), Elucidata Polly, Genedata (Switzerland; Genedata Selector/Profiler), QIAGEN Ingenuity Pathway Analysis (Germany/NL), Clarivate/Cortellis, Ochre Bio (UK).
• Single-cell workbenches: 10x Genomics Cellranger/Loupe, Parse Biosciences, Nygen, BioTuring.
• Choose A when: no in-house machine-learning team; accept data-sharing terms with the vendor. Check IP ownership of nominated targets and where omics data are stored.

Path B - Build on cloud APIs

• Managed compute: AWS HealthOmics, Google Cloud Batch, Azure Batch with EU regions; NVIDIA BioNeMo (hosted/NIM microservices) for biology foundation models.
• LLMs for literature and evidence synthesis: frontier reasoning models (Claude, Gemini, GPT) or Mistral via EU endpoints with retrieval over PubMed/Europe PMC, patents and trial registries (see KNOW-008). Use structured outputs and mandatory citations.
• Data and orchestration: Nextflow/nf-core pipelines, Snakemake, Terra/Seven Bridges-style workspaces; keep patient-level data in EU regions.

Path C - Self-host (open source)

• Single-cell stack: Scanpy, scverse (AnnData, scvi-tools, Squidpy, Muon), Seurat; DESeq2/edgeR.
• Open models: scGPT, Geneformer, UCE, scFoundation, Arc STATE (open weights), CellFM-style models; embedding access via CZ CELLxGENE Census [4][5][7].
• Knowledge and genetics: Open Targets (downloadable data) [11], PrimeKG, Hetionet, Neo4j/other graph DB, PyKEEN or PyG for link prediction; CausalNex/DoWhy for causal checks.
• Open LLMs (Qwen, Llama, Mistral families) for literature extraction on a local server.
• Hardware: HW2 for fine-tuning small cell models on a few million cells; HW3 (80-141 GB) for larger pretrained models and atlas-scale inference; HW4 for pretraining from scratch.

Path D - Custom training

Worth it when you own large proprietary perturbation or patient datasets. Fine-tune a cell foundation model on in-house Perturb-seq or longitudinal data; train organism- or tissue-specific models. Needs hundreds of thousands to millions of cells with consistent protocols and a held-out perturbation split to test generalisation to unseen genes and cell types. Invest in the data-generation loop (lab-in-the-loop, see SCI-007), which matters more than model size.

State of the art (2026)

• Virtual-cell race: Arc Institute released STATE and ran the Virtual Cell Challenge; CZI, Genentech, Recursion, NVIDIA and others build large perturbation datasets and models [4][8]. Larger datasets from genome-scale Perturb-seq are the main driver of progress.
• Reality check: peer-reviewed benchmarks found that several single-cell foundation models fail to outperform simple baselines on perturbation prediction and sometimes in zero-shot annotation [6]. Gains are clearest for batch integration, annotation and transfer to new datasets.
• Agentic analysis: LLM agents that run bioinformatics pipelines, query databases and draft target dossiers are in pilot use (see SCI-011); output needs auditing for fabricated references.
• Genetics-first prioritisation (Open Targets scoring, human-genetic support) remains the most robust evidence layer.
• Limits: batch effects, cell-line to patient transfer, confounding in observational cohorts, train-test leakage between related datasets, poor calibration, and publication bias in knowledge graphs. Biomarker signatures often fail to replicate in independent cohorts.

Shared resources & combinations

• Same stack as SCI-009 (genomics pipelines, ClinVar/gnomAD) and SCI-001 (structure of the nominated target for tractability); downstream SCI-002 (molecule generation) and SCI-015 (in-silico toxicology for safety liabilities).
• SCI-011 (AI scientist agents) for hypothesis generation; SCI-007 for lab-in-the-loop validation.
• KNOW-008 (literature assistant) and the knowledge-graph/vector-DB capabilities are shared; HLTH-002 (digital pathology) supplies spatial and histology features; HLTH-009 uses biomarkers for trial enrichment.

Effort, cost & KPIs

• Path A: weeks to onboard; licences are typically enterprise subscriptions or project-based.
• Path B: 3-6 months to a working pipeline with governed data access.
• Path C: 6-12 months; needs bioinformatics and machine-learning engineers plus GPU server.
• Path D: 12+ months plus experimental data-generation costs, which dominate the budget.
• KPIs: share of nominated targets passing wet-lab validation, benchmark performance versus simple baselines on held-out perturbations, biomarker replication in independent cohorts, time from hypothesis to first experiment, novelty of validated targets, reproducibility of pipelines.

People & change

Involve the department lead and bioinformatics staff from the first scoping workshop, because they define which evidence counts. Management sets budget and IP rules (who owns AI-nominated targets). IT and IT security provide GPU or cloud access and protect patient-level data; the data protection officer (DPO) must be involved before any omics cohort is touched, with a DPIA (Art. 35 GDPR) likely for genetic and health data and a processor agreement (Art. 28 GDPR) for every cloud or platform vendor. Legal and compliance cover consent, EHDS and IP clauses.

Works council: depends. The use case analyses samples, not staff, but platforms with user logs can fall under co-determination (section 87(1) no. 6 BetrVG, objective suitability suffices), and the works council must be informed when work processes are planned with AI (section 90(1) no. 3). Clarify early; public institutions involve the staff council (Personalrat).

AI literacy: the AI Act Art. 4 asks employers to take measures that support staff AI literacy; train scientists on baseline comparison, leakage and over-reliance on vendor scores. Pick a pilot group of two or three biologists with one disease area. Typical concern: "the model replaces my judgement" - answer with the rule that every target passes expert review and wet-lab validation.

Guides: GUIDE-005, GUIDE-009, GUIDE-006.

Risks, regulation & compliance

• EU AI Act: research use for target discovery is minimal risk (research exemption, Art. 2(6) covers AI developed solely for scientific research). A biomarker used as a companion diagnostic or clinical decision aid can become a medical device or in-vitro diagnostic under MDR/IVDR and then a high-risk AI system under the AI Act (Art. 6, Annex I) [9].
• GDPR: patient omics are health and genetic data (Art. 9); single-cell and genomic profiles are hard to anonymise. Need a legal basis (explicit consent or research provisions of Art. 89), DPIA, data-processing agreements, EU hosting and access controls. The European Health Data Space regulation adds rules for secondary use of health data [10].
• Sector rules: ICH/GCP and EMA expectations when biomarkers enter trials; EMA reflection paper on AI across the medicinal-product lifecycle; Nagoya Protocol for biological material origin.
• Scientific risks: spurious targets from confounding or leakage; hallucinated citations; over-reliance on vendor scores. Mitigate with baselines, pre-registered validation plans, independent cohorts, provenance tracking and expert sign-off.
• IP: clarify ownership of AI-nominated targets and training-data licences.

Getting started

1. Pick one disease area and assemble a curated dataset (public: CELLxGENE, GEO, Open Targets) with a documented data-governance plan.
2. Reproduce a baseline: Open Targets ranking plus a Scanpy/scvi-tools pipeline on your data.
3. Benchmark a cell foundation model against simple baselines on a held-out perturbation split.
4. Run a retrospective test: can the pipeline rediscover a known, validated target from data before its discovery date?
5. Select 5-10 targets for a small CRISPR or organoid validation round and track the hit rate.

Sources

1. Xu Z. et al. - Rentosertib (TNIK inhibitor) Phase 2a in IPF, Nature Medicine 2025 - https://www.nature.com/articles/s41591-025-03743-2
2. Richardson P. et al. - Baricitinib as potential treatment for 2019-nCoV acute respiratory disease, The Lancet 2020 - https://www.thelancet.com/journals/lancet/article/PIIS0140-6736(20)30304-4/fulltext
3. Minikel E. et al. - Refining the impact of genetic evidence on clinical success, Nature 2024 - https://www.nature.com/articles/s41586-024-07316-0
4. Arc Institute - STATE virtual cell model - https://github.com/ArcInstitute/state
5. Cui H. et al. - scGPT: foundation model for single-cell multi-omics, Nature Methods 2024 - https://www.nature.com/articles/s41592-024-02201-0
6. Ahlmann-Eltze C. et al. - Deep-learning-based gene perturbation effect prediction does not yet outperform simple linear baselines, Nature Methods 2025 - https://www.nature.com/articles/s41592-025-02772-6
7. CZ CELLxGENE Census - https://chanzuckerberg.github.io/cellxgene-census/
8. Arc Institute - Virtual Cell Challenge - https://virtualcellchallenge.org/
9. Regulation (EU) 2024/1689 (AI Act) - https://eur-lex.europa.eu/eli/reg/2024/1689/oj
10. Regulation (EU) 2025/327 (European Health Data Space) - https://eur-lex.europa.eu/eli/reg/2025/327/oj
11. Open Targets Platform - https://platform.opentargets.org/