A Scalable Approach to Combinatorial Library Design for Drug Discovery

Abstract

In this paper, we propose an algorithm for the design of lead generation libraries required in combinatorial drug discovery. This algorithm addresses simultaneously the two key criteria of diversity and representativeness of compounds in the resulting library and is computationally efficient when applied to a large class of lead generation design problems. At the same time, additional constraints on experimental resources are also incorporated in the framework presented in this paper. A computationally efficient scalable algorithm is developed, where the ability of the deterministic annealing algorithm to identify clusters is exploited to truncate computations over the entire data set to computations over individual clusters. An analysis of this algorithm quantifies the tradeoff between the error due to truncation and computational effort. Results applied on test data sets corroborate the analysis and show improvement by factors as large as 10 or more, depending on the data sets.

On reduction of graphs and Markov chain models

“…squared Euclidean and K L divergence used in Sections III and IV) and therefore accommodates many application areas (discussed in Section I). Also, the algorithm presented in this paper is for a simple graph-reduction problem; however the approach can eas ily incorporate dynamic, communication, and computational constraints by adapting the tools that we have developed for clustering/classification problems in our previous work [7]- [10]. For instance, we have applied this approach to simplification of influence diagrams obtained from neuro scientific community with good results, where we could account for the distance functions and constraints that are unique to that particular application area.…”

“…However, the phase-transition property makes this algorithm progressively localized; which when exploited makes it com putationally efficient [7], [9], [24]. Thus this algorithm makes use of the global information in its initial steps to avoid local minima while the localization of the latter iterations for reducing computational expense.…”

“…We adopt a soft partitioning approach from the DA al gorithm [7]- [9], [20] (in contrast to the hard partitioning approach, where each node uniquely and determinately be longs to one partition), which allows each node i to associate with more than one cell j through a (nonnegative) weight ing parameter p(ZjIXi) . More specifically, the dissimilarity…”

“…The resulting cost functions are non-convex and typically exhibit multiple local minima. In this aspect, graph reduction closely resembles widely studied data-clustering [xu27, salapaka, beck3]@illinois.edu problems (such as resource allocation) [7]-[11] and therefore can avail tools from the latter. A critical difference lies in the difficulty of formulating a cost function since this requires defining a metric that compares two graphs (the original and the reduced) that are of diff erent dimensions.…”

Dynamic Maximum Entropy algorithms for clustering and coverage control

“…They are all in the University of Illinois at Urbana-Champaign, Urbana, Illinois, USA. Email: [xu27, salapaka, beck3]@illinois.edu Dynamic coverage problems inherit the comuptational complexity of facility location problems that arise in a variety of static applications such as locational optimization, facility location, optimal coding, pattern recognition and learning, and data clustering and classification [5]- [8]. The static coverage problems are known to be NP-hard [9]- [12], where the cost functions are non-convex and are typically riddled with multiple local minima.…”