Paolo Tosco Abstract

Design in 2D, model in 3D: live 3D pose generation from 2D sketches

Paolo Tosco1, Mark Mackey1

How nice would it be to be able to draw a molecule in your favourite 2D sketcher, and see in real time how its 3D electrostatic potential looks like, what interactions it makes in the active site of its protein, how its shape and electrostatics compare to a known active for the same target? We have thought of this many times, as have most medicinal chemists; this is why we decided to build from ground up a new molecular design platform around this engaging idea.

However, as we started fleshing out the details of the algorithm, a number of obvious and more subtle gotchas emerged. While a ligand is being sketched on a 2D canvas, its 3D counterpart which is growing inside the protein active site should not raise steric clashes with the target itself. On top of this, a 3D conformation has far more degrees of freedom than a 2D graph; ideally, while making its way into the binding pocket, the 3D ligand should automatically adjust its torsion angles to pick up the most favourable contacts available, with as little intervention from the user as possible.

Where no crystal structures are available for the target of interest, there might be reference ligands with known affinity for the target to compare to in terms of electrostatic and shape properties. When the chemist decorates a certain aromatic ring with an ortho or meta substituent, the algorithm should put it on the right side of the 3D structure, in order to either best fit the pocket (if there is one), or to mimic a certain electrostatic feature in the reference ligand. Where relevant, the same paradigm should apply to the cis/trans geometry of double bonds and to the stereochemistry of tetrahedral centres, in case the user does not enforce a specific stereochemistry.

A single 2D change such as switching a bond from wedged to dashed or converting a carbonyl group into a sulfoxide in the middle of the molecule may have a dramatic leverage on the overall 3D arrangement of the molecule, requiring a complete reassessment of its pose. Adding or removing a single bond may break a molecule into two separate fragments, or suddenly attach a large, previously disconnected moiety to the current design. Similar disruptive actions include the cyclization of a linear chain, the breakage of a ring, the aromatization of a piperidine into a pyridine or the hydrogenation of benzene to cyclohexane.

While drawing a molecule in 2D, we often go through a number of chemically invalid states: we are thinking of drawing a sulfonyl, but what we actually do is to draw a gem-diol, then turn the diol into a gem-dicarbonyl monster, and finally convert the carbon atom into a sulfur. Likewise, we most often draw a tetravalent nitrogen before adding a positive formal charge. The algorithm must cope with these workflows and not choke on the invalid intermediates.

In this talk we will illustrate how we addressed the challenge of providing immediate 3D poses from 2D sketches in our Grow3D algorithm. We will also give an overview of the integration of Grow3D in the context of our molecular design platform, showing how effective it is at bridging the gap between 2D and 3D methods.