In an article published in Current Opinion in Structural Biology, we review state of the art in generative models that operate on molecular graphs. We also shed light on some limitations of the existing methodology and sketch directions to harness the potential of AI for drug design tasks going forward. 

In our new paper, we seek to better understand this phenomenon from both theoretical and empirical perspectives.  Since most of these models apply graph neural networks (GNNs), one may suspect that they inherit the representational limitations of GNNs. We analyze this aspect, establishing the first such results for protein-ligand complexes. A plausible counterview may attribute the underperformance of these models to their excessive parameterizations, inducing expressivity at the expense of generalization. 

We also investigate this possibility with a simple metric-aware approach that learns an economical surrogate for affinity to infer an unlabelled molecular graph and optimizes for labels conditioned on this graph and molecular properties. Our model SimpleSBDD achieves state-of-the-art results using 100x fewer trainable parameters and affords up to 1000x speedup. 

In one of our ICML 2023 papers, we show how generative AI can help design new antibodies from scratch given information about the specific antigens (such as virus, bacteria). 

wise interactions, enabling richer representations than GNNs. GNNs also cannot compute important properties such as number and length of cycles (e.g., rings in the molecules), and have been augmented with persistence homology (PH) to mitigate this issue. 

In an ICML 2024 paper,  we design TopNets – a new class of deep learning models that is strictly more powerful than TNNs/GNNs and PH. TopNets can also be readily adapted to handle (symmetries in) geometric complexes, extending the scope of TNNs and PH to spatial settings (such as generating molecular conformations). TopNets achieve strong performance across diverse drug discovery tasks, including antibody design, molecular dynamics simulation, and drug property prediction. 

complex interactions between latent generative factors pertaining to different molecular properties of interest (e.g., ADMET properties). For example, we might want to improve the metabolism property (‘M’ in ADMET) of a molecule without significantly altering its other properties. 

However, not all properties can be disentangled – e.g., usually it’s hard to optimize or improve bioactivity of a candidate molecule without enhancing its toxicity as well. In a NeurIPS 2022 paper, we address this issue with a novel concept of conditional disentanglement: our method  segregates the latent space into uncoupled and entangled parts. This allows us to exercise finer control over molecule generation and optimization. Experimental results on molecular data strongly corroborate the interpretability of our method. Interestingly, this interpretability is accompanied with improved molecular generation performance, as evaluated across multiple standard metrics.

In a popular article published in Helsingin Sanomat (the largest subscription newspaper in the Nordic countries), we discuss how advances enabled by AI/ML, e.g., in protein folding and inverse folding or protein design, have opened tremendous opportunities for bioengineering including designing better drugs, materials, biofertilizers, and batteries. The article focuses on foundations works in this space such as AlphaFold by DeepMind and protein/antibody design by our team in collaboration with MIT and Aalto.

Read more from the Helsingin Sanomat article here.