Abstract: Generating (macro-)molecules that have desirable physicochemical properties holds the key to enabling better drugs, materials, and batteries, etc. However, several challenges must be overcome, e.g., in order to (a) tractably search over an extremely large combinatorial space (e.g., ~10^60 drug-like structures), (b) model the underlying dynamics (interactions) to be able to focus on the right regions of the space, and (c) segregate the representation for undesirable properties (e.g., toxicity) - that can then be suppressed - from the desired part.

I will give an overview of some of our upcoming work [1, 2, 3] that makes significant advances in this quest. In particular, unlike previous approaches, we’re able to generate high quality molecules without resorting to any validity checks or correction.

[1] Yogesh Verma, Samuel Kaski, Markus Heinonen, and Vikas Garg. Modular Flows: Differential Molecular Generation, NeurIPS 2022.
[2] Giangiacomo Mercatali, Andre Freitas, and Vikas Garg. Symmetry-induced Disentanglement on Graphs, NeurIPS 2022.
[3] Amauri Souza, Diego Mesquita, Samuel Kaski, and Vikas Garg. Provably expressive temporal graph networks, NeurIPS 2022.

Speaker: Vikas Garg is an Assistant Professor at Aalto University and FCAI, and Chief Scientist at YaiYai Ltd. He holds a PhD in Computer Science from MIT, and has led various research and engineering efforts during stints at IBM Research, Microsoft Research, and Amazon A9.
His co-authored works have contributed to advancing multiple domains, including the first graph-based deep learning model for generating new proteins, context-aware recommender systems for e-commerce platforms, fast inference on resource-constrained IoT devices such as smartphones, integration of renewable energy into smart grids, and design of next generation wireless systems; as well as exposing the limitations of Generative Adversarial Networks (GANs) and Graph Neural Networks (GNNs).

Vikas has served as an invited Area Chair/Senior Program Committee member/Panelist at premier AI/ML venues. His select honors include a BP Technologies Energy Fellowship and recognition by MIT EECS as one of its strongest incoming students, and highest evaluation scores for a course on Applied Machine Learning that he co-designed and co-instructed at MIT.

Affiliation: Aalto University