Abstract:
Artificial intelligence is reshaping how we discover and design in science. This talk introduces a deployment-centric perspective on building trustworthy AI systems that connect generative creativity with real-world impact. I will present recent work from our group on MapDiff, a diffusion-based generative framework that models protein sequences from 3D structures, and DrugBAN, an interpretable bilinear attention network for drug–target prediction. These examples illustrate how generative and multimodal learning can accelerate discovery when guided by deployment-aware principles. By coupling generative capability with interpretability, multimodal integration, and domain knowledge, we move towards AI systems that advance both molecular research and the broader scientific enterprise.
Bio:
Haiping Lu is Professor of Machine Learning at the University of Sheffield, UK, where he leads AI Research Engineering at the Centre for Machine Intelligence. He is also Director of the UK Open Multimodal AI Network (UKOMAIN), funded by the UK Engineering and Physical Sciences Research Council (EPSRC). His research focuses on deployment-centric multimodal AI, integrating diverse types of data to tackle challenges in healthcare and scientific discovery, with methodological interests spanning foundation models, generative AI, domain adaptation, and transfer learning. His recent work spans brain and cardiac imaging, cancer diagnosis, protein design, and drug and materials discovery. He leads the open-source project PyKale for knowledge-aware machine learning and serves as an Associate Editor for IEEE Transactions on Neural Networks and Learning Systems and IEEE Transactions on Cognitive and Developmental Systems. He has received awards from the Alan Turing Institute, Amazon, the Wellcome Trust, and the UK's National Institute for Health and Care Research.
The deployment-centric concept in the title builds on our recent Nature Machine Intelligence Perspective, available here: https://www.nature.com/articles/s42256-025-01116-5
Summary:
Aim: AI for Scientific Discovery with a focus on deployability
Observation
Multi-modal ML models are becoming increasingly more common
Language and vision are much more common than other modalities (time series, sensor, audio, tabular)
Motivates work to incorporate the other modalities
E.g. discriminative and generative models
Generative modeling is very capable because it can reconstruct one modality from others
Bridge between modalities
Align correspondence between them
Cross-modality representations, which are complementary
Novel multi-modal privacy risk because one can’t anonymize modalities in isolation
Models are increasingly maturing; data is the major differentiator that enables discovery
Deployment-centric AI
Consider deployment constraints and user needs
E.g. ethical considerations
Flow:
Planning: Problem definition, deployment constraints, task formulatoin
Multimodal AI system development
Real-world deployment
DrugBAN: Drug-target interaction prediction
Graph convolutional network to encode drug’s chemical structure
1D convolutional network encodes sequences of proteins the drug binds with
Multi-head Bilinear interaction model + pooling integrates the two modalities into a joint representation
Focus on deployability: drug companies specifically care about discovery of new drugs
Cluster data according to drug type
Train on some categories of drugs, predict on different ones
Same approach applied to cell line-drug response, mutation-drug association
MapDiff: Inverse protein folding
Design protein with a particular function/structure
Mask prior-guided denoising Diffusion
Mark-prior pretraining:
Protein 3D structure modality
Mark part of the sequence, predict it again
Model learns a good embedding for 3D structure for amino acid sequences
Mask-guided denoising network
Equivariant graph neural network: generates 3D sequence based on structure
Entropy-based mask (more uncertain sequence components are masked and predicted)
Diffusion model to predict sequence
Using Alphafold to validate the consistency of the sequence’s foldability
Recommendations for deployment-centric development
Safety, reliability, interpretability
Scalability and resource efficiency
Ethical compliance and user preparedness
Recommendations for data- and model-centric development
Data scarcity and access
Balanced data representation
Multimodal fusion and modality selection
Foundation models
Recommendations for stakeholder engagement and collaboration
Stakeholder inclusion and alignment
Cross-disciplinary standards and communication
Intellectual property and workflow adaptation
PyKale: pykale.github.io
Knowledge aware machine learning from multiple sources
UKOMAIN: UK Open Multimodal AI Network: https://multimodalai.github.io/
Builds diverse interdisciplinary network
Create knowledge exchange platform
Idenrtify & fund key focus areas of high impact
Engage industry & policymakers
Promote sustainability & responsibility
Enhance research capability & training
OMAIB: Open Multimodal AI Benchmark (funding call)
CD3: Cancer Data Driven Detection:
Advance ability to prevent, diagnose and detect cancer
Enhance equity
Research community that federates resources
Health records, cancer multi-omics
Advanced analytics
Partnerships
Conclusion: multimodal data + GenAI = Impact