Network Planners

Building a drug candidate ranking is essentially a graph or network construction problem. A connected network, with candidates as nodes, can describe all the relationships among each candidate. The relationship between two molecules corresponds to an edge in this network, which translates to one Relative Binding Free Energy (RBFE) calculation. In practice, computational chemists conduct numerous RBFE calculations to construct networks that illustrate the relationships of all drug candidates to one another.

It’s important to note that free energy is a thermodynamic state function, making it path independent. This means that a network only needs to be connected to provide insights into all relationships, provided all RBFE calculations are of high quality. For example, if molecule A has a direct relation x with molecule C, and they are connected through molecule B, using relations y and z, then the sum of y + z will still yield x.

So, what can we do with this? As mentioned, the free energy network should be calculated efficiently. Keep in mind that computational chemists utilize high-performance computing resources to perform RBFE calculations, which can take several hours. Therefore, RBFE calculations are costly, and it is typically avoided to conduct too many.

This consideration directly impacts the design of the free energy network. To orchestrate the calculations, a plan for which connections to compute is generated before the simulations. This is where Konnektor comes into play.

Konnektor is a package that assists in generating the free energy network calculation plan. It implements multiple network layouts, each with its own advantages and disadvantages. How is such a network plan generated? In our RBFE calculation example, each edge can be represented as an AtomMapping, indicating a common substructure between the two molecules to be compared. This AtomMapping can then be scored using an AtomMappingScorer, which indicates the expected difficulty of the transformation in terms of convergence or accuracy. But you can also look at this more abstract, as an ComponentMapping is describing any relation between two Component`s and the difficulty is estimated with a `ComponentMappingScorer.

A network planning algorithm NetworkPlanner can now use these scores along with graph construction algorithms to identify the best calculation paths.

Checkout the network tools to see what you can additionally do with networks.

Network Generators

Network Generators are planners that construct networks from a set of components. They are usually the starting point for any network planning efforts and come in a wide variety of layouts:

The minimal number of edges for ranking can be achieved with the Star Network and the Minimal Spanning Tree (MST) Network (N-1). More redundant layouts include the Twin Star Network, Redundant MST Network, N-Edge Node Network, and Cyclic Graph Network.

The Maximal Network method generates all possible edges and is typically used as an initial solution, which then gets reduced to a more efficent layout. Additionally, there is the Heuristic Maximal Network, which aims to produce an edge-reduced version of the Maximal Network.

Network Concatenators

NetworkConcatenators address the challenge of connecting two networks that do not share any edges. This essentially involves solving a bipartite graph matching problem. These planners generate edges between two unconnected, non-overlapping networks to create a connected network.

Currently, there are two Network Concatenators in Konnektor: the Maximal Concatenator, which yields all possible edges, and the Minimal Spanning Tree Concatenator, which utilizes the best-performing edge scores.