Venkat Venkatasubramanian

Vipul Mann, Venkat Venkatasubramanian

Rapid discovery of new reactions and molecules in recent years has been facilitated by the advancements in high throughput screening, accessibility to a much more complex chemical design space, and the development of accurate molecular modeling frameworks. A holistic study of the growing chemistry literature is, therefore, required that focuses on understanding the recent trends and extrapolating them into possible future trajectories. To this end, several network theory-based studies have been reported that use a directed graph representation of chemical reactions. Here, we perform a study based on representing chemical reactions as hypergraphs where the hyperedges represent chemical reactions and nodes represent the participating molecules. We use a standard reactions dataset to construct a hypernetwork and report its statistics such as degree distributions, average path length, assortativity or degree correlations, PageRank centrality, and graph-based clusters (or communities). We also compute each statistic for an equivalent directed graph representation of reactions to draw parallels and highlight differences between the two. To demonstrate the AI applicability of hypergraph reaction representation, we generate dense hypergraph embeddings and use them in the reaction classification problem. We conclude that the hypernetwork representation is flexible, preserves reaction context, and uncovers hidden insights that are otherwise not apparent in a traditional directed graph representation of chemical reactions.

Venkat Venkatasubramanian, N. Sanjeevrajan, Manasi Khandekar

We present a novel theory of the microstructure of deep neural networks. Using a theoretical framework called statistical teleodynamics, which is a conceptual synthesis of statistical thermodynamics and potential game theory, we predict that all highly connected layers of deep neural networks have a universal microstructure of connection strengths that is distributed lognormally ($LN(μ, σ)$). Furthermore, under ideal conditions, the theory predicts that $μ$ and $σ$ are the same for all layers in all networks. This is shown to be the result of an arbitrage equilibrium where all connections compete and contribute the same effective utility towards the minimization of the overall loss function. These surprising predictions are shown to be supported by empirical data from six large-scale deep neural networks in real life. We also discuss how these results can be exploited to reduce the amount of data, time, and computational resources needed to train large deep neural networks.

Venkat Venkatasubramanian

Large Language Models (LLMs) are often criticized for lacking true "understanding" and the ability to "reason" with their knowledge, being seen merely as autocomplete systems. We believe that this assessment might be missing a nuanced insight. We suggest that LLMs do develop a kind of empirical "understanding" that is "geometry"-like, which seems adequate for a range of applications in NLP, computer vision, coding assistance, etc. However, this "geometric" understanding, built from incomplete and noisy data, makes them unreliable, difficult to generalize, and lacking in inference capabilities and explanations, similar to the challenges faced by heuristics-based expert systems decades ago. To overcome these limitations, we suggest that LLMs should be integrated with an "algebraic" representation of knowledge that includes symbolic AI elements used in expert systems. This integration aims to create large knowledge models (LKMs) that not only possess "deep" knowledge grounded in first principles, but also have the ability to reason and explain, mimicking human expert capabilities. To harness the full potential of generative AI safely and effectively, a paradigm shift is needed from LLM to more comprehensive LKM.

Kevin Zhang, Vipul Mann, Venkat Venkatasubramanian

Various template-based and template-free approaches have been proposed for single-step retrosynthesis prediction in recent years. While these approaches demonstrate strong performance from a data-driven metrics standpoint, many model architectures do not incorporate underlying chemistry principles. Here, we propose a novel chemistry-aware retrosynthesis prediction framework that combines powerful data-driven models with prior domain knowledge. We present a tree-to-sequence transformer architecture that utilizes hierarchical SMILES grammar-based trees, incorporating crucial chemistry information that is often overlooked by SMILES text-based representations, such as local structures and functional groups. The proposed framework, grammar-based molecular attention tree transformer (G-MATT), achieves significant performance improvements compared to baseline retrosynthesis models. G-MATT achieves a promising top-1 accuracy of 51% (top-10 accuracy of 79.1%), invalid rate of 1.5%, and bioactive similarity rate of 74.8% on the USPTO- 50K dataset. Additional analyses of G-MATT attention maps demonstrate the ability to retain chemistry knowledge without relying on excessively complex model architectures.

Vipul Mann, Abhishek Sivaram, Laya Das et al.

Swarm Intelligence-based optimization techniques combine systematic exploration of the search space with information available from neighbors and rely strongly on communication among agents. These algorithms are typically employed to solve problems where the function landscape is not adequately known and there are multiple local optima that could result in premature convergence for other algorithms. Applications of such algorithms can be found in communication systems involving design of networks for efficient information dissemination to a target group, targeted drug-delivery where drug molecules search for the affected site before diffusing, and high-value target localization with a network of drones. In several of such applications, the agents face a hostile environment that can result in loss of agents during the search. Such a loss changes the communication topology of the agents and hence the information available to agents, ultimately influencing the performance of the algorithm. In this paper, we present a study of the impact of loss of agents on the performance of such algorithms as a function of the initial network configuration. We use particle swarm optimization to optimize an objective function with multiple sub-optimal regions in a hostile environment and study its performance for a range of network topologies with loss of agents. The results reveal interesting trade-offs between efficiency, robustness, and performance for different topologies that are subsequently leveraged to discover general properties of networks that maximize performance. Moreover, networks with small-world properties are seen to maximize performance under hostile conditions.