Publications

First broadly generalizable diffusion model for 3D molecular generation. State-of-the-art results for periodic crystals and non-periodic molecular systems through transfer learning. (ICML 2025 & Spotlight at ICLR 2025 Workshop)

RNA-FrameFlow is a generative model for 3D RNA backbone design based on SE(3) flow matching. We evaluated the extent to which protein generative models can be adapted for RNA.

A new generative toolkit for inverse design of 3D RNA structure. gRNAde has been validated in wet labs, achieving expert-level accuracy and outperforming all previous computation methods at designing RNA structure and function. (ICLR 2025 Spotlight & invited book chapter for Methods in Molecular Biology)

Open-source benchmarking framework for evaluating and scaling up protein structure representation learning methods, comparing protein language models and Geometric GNNs trained on AlphaFoldDB, the largest non-redundant protein structure corpus.

A ‘Hitchhiker’s Guide’ introducing deep learning for molecular modelling—written without the goal of publication—if we could travel back in time and learn from it at the start of our PhDs! 🤗

We propose gRNAde, a geometric deep learning pipeline for 3D RNA inverse design from one or more backbone structures. (Invited book chapter for Methods in Molecular Biology)

A unified theory of Graph Neural Networks and Transformers for 3D molecular structure. And mechanistic understanding of how architectural decisions, such as invariance vs. equivariance, impact the class of functions that models can express. (ICML 2023 & Oral at NeurIPS 2022 Workshop)

Strategies for distilling representational knowledge from expressive yet cumbersome-to-deploy GNNs to resource-efficient and lightweight architectures. (IEEE TNNLS)

We propose a point discriminative learning method for unsupervised representation learning on 3D point clouds to learn local and global shape features. (3DV 2022)

HYFA (Hypergraph Factorisation) jointly imputates multi-tissue and cell-type gene expression via a specialised graph neural network operating on a hypergraph of individuals, metagenes, and tissues. (Nature Machine Intelligence, cover article)

We study zero-shot generalization to large-scale instances in neural network-driven solvers for the Travelling Salesman Problem: what architectures, inductive biases and learning paradigms enable better generalization? (Invited submission to the Constraints Journal)

We study zero-shot generalization to large-scale instances in neural network-driven solvers for the Travelling Salesman Problem: what architectures, inductive biases and learning paradigms enable better generalization? (CP 2021)

Representation learning for free-hand drawings using GNNs and Transformers. (IEEE TNNLS)

Open-source benchmarking framework to identify scalable and powerful GNN architectures, and track the progress of graph representation learning research. (500+ citations, 2000+ GitHub stars)

How do learning paradigms impact zero-shot generalization to large-scale instances in learning-driven TSP solvers? (NeurIPS 2019 Workshop)

Deep Graph ConvNets paired with parallelized graph search can learn TSP up to few hundred cities, but fall short of classical solvers.

Feature attribution to interpret decision tree models and study the generational impact of caste on women’s work force participation in India. (ICLR 2019 Workshop)

Combinatorial optimization problems, also called NP-hard problems, are practical constraint satisfaction problems that are impossible …

Learning low-dimensional vector spaces of financial statements and invoices for constructing semantic similarity graphs. (US Patent)

Matching financial statements and invoices via graph combinatorial optimization algorithms on semantic similarity graphs. (US Patent)

Representing sets of financial statements and invoices as bags-of-features. (US Patent)

We discuss the Question-Answering task in NLP, and introduce a class of deep learning models known as memory networks to build QA and chatbot systems.

New datasets and Memory Network architectures for studying personalized dialog systems based on user profiles. (NeurIPS 2017 Workshop)