Scientific AI Flagship · Human Technopole


Projects

12 sub-projects, 5 work packages, 2 aims. Aim 1 (WP1–WP3) develops the methodological backbone; Aim 2 (WP4–WP5) puts those methods to work across HT through consultancy and graduating collaborations.

WP1

Combining Biological Scales from Molecules to Populations

Coordinators: Jug, Funke

Bridge the full continuum of biological organization — molecules, cells, tissues, organs, populations — with scalable AI methods that integrate and reason across these levels. The ambition is to turn fragmented data streams into coherent, predictive representations of biological systems.

P1.1

Bridging structure and -omics at subcellular resolution

PIs: Pigino (lead), Funke, Jug

Integrate cryo-EM, expansion microscopy, volumetric CLEM, and spatial omics to correlate ciliary structural features with molecular states at subcellular resolution. Connects to the Ciliopathies RFP and produces a general toolkit for nanoscale structure-to-molecule analysis.

P1.2

From FIB-SEM to live-cell imaging

PIs: Zerial (lead), Funke, Jug

A correlative pipeline linking volumetric FIB-SEM with fluorescence imaging of fixed and live cells. Deep-learning segmentation, efficient annotation, and HPC produce multi-scale maps from organelles to tissue. Initial focus: liver tissue; eventual delivery via HT’s National Facilities.

P1.3

AI for molecular knowledge at population scale

PIs: Soranzo (lead), Ieva, new HDS GL

Integrate population-scale single-cell genomics with electronic health records to derive precise inferences of disease trajectories. Informed by Soranzo’s ERC IMPACT grant and the UK Biobank / Genes and Health cohorts. Connects to the Cardiometabolic RFP.

WP2

Combining Data Modalities into Multimodal Solutions

Coordinators: Glastonbury, Jug

Develop AI strategies that integrate heterogeneous data streams into unified, interpretable representations — revealing relationships invisible within any single modality and supporting predictive, mechanistic understanding of disease.

P2.1

Multimodal, explainable breast cancer risk prediction

PIs: Jug (lead), Di Angelantonio

A device-agnostic AI framework integrating 2D mammography, 3D digital breast tomosynthesis, and longitudinal EHR data for individualized, time-specific risk scores. Focus on out-of-distribution robustness, demographic fairness, and human-centred explainability co-designed with radiologists.

P2.2

Exposome intelligence and digital twins

PIs: Ieva (lead), new HDS GL

AI methods for clinical complexity that go beyond medical records to include the exposome — environmental, socio-demographic, economic, and behavioural factors. Combined with Medical Digital Twins, this enables truly personalized health policies.

P2.3

A multimodal spatial pathology atlas of Alzheimer’s disease

PIs: Glastonbury (lead)

With King’s College London Neurodegenerative Disease Biobank: ~30,000 whole-slide images from 640 donors plus Visium HD spatial transcriptomics, WGS, and pathology reports. Goal: the most detailed multimodal map of Alzheimer’s pathology to date.

WP3

Synthetic Data Generation at Scale

Coordinators: Ieva

Generate realistic synthetic biological and clinical data that protects privacy while expanding analytical reach — supporting research that would otherwise be blocked by data scarcity or sensitivity.

P3.1

Virtual Patient

PIs: Ieva (lead), new HDS GL

Holistic patient representation through deep learning and LLMs. Conditional GANs and language models for structured EHRs; LLM pipelines with RAG and knowledge-graph integration for clinical narratives; anatomically guided medical-image synthesis. Connects to the Cardiometabolic RFP.

P3.2

Failure-aware agentic AI critics from ELN data

PIs: Sottoriva (lead), Jug

Train critic models to evaluate plans proposed by agentic AI in the lab — learning from the silent knowledge buried in failed experiments and protocol deviations recorded in ELN systems. The critics flag risky actions and propose safer revisions.

WP4

Consultancy and Proof-of-concept Solutions

Coordinators: Funke (co-lead Jug)

The dynamic side of the Hub: short, well-scoped proof-of-concept work for HT colleagues. The four projects below are a snapshot — new ideas can come in throughout the Flagship’s lifetime.

P4.1

AI prediction of B-cell differentiation

PIs: Soskic (lead), Funke, Jug

Predict B-cell differentiation outcomes from immunofluorescence and high-content microscopy. Use explainability methods (e.g. QuAC) to identify the visual features driving the predictions, then validate them experimentally.

P4.2

Patient-guided AI for cancer cell vulnerabilities

PIs: Iorio (lead), Jug

Predict drug responses from large-scale functional genomics (DepMap), with patient-guided feature prioritization using TCGA, ICGC, and PCAWG cohorts. Open-source pipelines and patient-level projections.

P4.3

3D segmentation of stem cells in tissue and organoids

PIs: Kalebic (lead), Funke

Robust deep-learning pipeline for segmenting individual cells in 3D brain organoids and tumor organoids — cells with intricate, overlapping morphologies that defeat conventional algorithms. Builds on Funke’s connectomics work.

P4.4

Omics2EM: bridging -omics and cryo-EM

PIs: Calviello (lead), Erdmann, Funke, Jug

Use molecular heterogeneity from transcriptomics and proteomics as a predictor of alternative complexes in cryo-EM — and conversely, use cryo-EM latent representations to identify heterogeneous molecular complexes. Initial focus: the ribosome.

WP5

Collaborative Science Facilitation Projects

Coordinators: Jug (co-lead Funke)

Where successful proof-of-concept projects from WP4 graduate into sustained, co-developed research. Two projects have already reached this level of maturity.

P5.1

Codon- and structure-aware RNA language model

PIs: Legnini (lead), Jug

An RNA language model that integrates codon semantics, regulatory motifs, and secondary structure to predict mRNA half-life. Hybrid tokenization (codons for coding regions, sub-words plus structural annotations for UTRs) and physics-informed regularization.

P5.2

AI-enabled cryo-FIB and lift-out pipeline

PIs: Erdmann (lead), Pigino, Jug

An automated pipeline that integrates light microscopy, SEM, and FIB to guide cryo-FIB lift-out and lamella preparation. Deep-learning feature detection enables automated, reproducible sample production from organoids and tissues.