Publications

gRNAde is a geometric deep learning pipeline for 3D RNA inverse design, analogous to ProteinMPNN. gRNAde outperforms physically-based Rosetta and is significantly faster. (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’ for getting started with deep neural networks for 3D structural biology & chemistry – we wrote it without the goal of publishing at particular venues – as if we could travel back in time and learn from it if we were starting our PhDs today! 🤗

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)

We propose a Geometric Weisfeiler-Leman test to study the expressive power of Geometric Graph Neural Networks and demonstrate the advantages of equivariant GNNs over invariant GNNs. (Full paper at ICML 2023, and presented as an 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)