Block 1: Lecture 10: Molecular modeling in drug discovery

How do you choose a method?

The pharmacophore

A pharmacophore is the ensemble of steric and electronic features that is necessary to ensure the optimal supramolecular interactions with a specific biological target structure and to trigger (or to block) its biological response.

A pharmacophore does not represent a real molecule or a real association of functional groups, but a purely abstract concept that accounts for the common molecular interaction capacities of a group of compounds towards their target structure. The pharmacophore can be considered as the largest common denominator shared by a set of active molecules.

IUPAC

International union of pure and applied chemistry

Ligand-based design (shaep and volume)

Two molecules that fill the same space, with their polar groups in the same places, can bind the same site even when the structures look nothing alike. We would know that the ligands are active, but we do not know the difference in binding points in each molecule (e.g. some can flip).

QSAR

Quantitative structure-activity relationship

What are the 3 steps of QSAR?

1. Describe each molecule as numbers: logP, polar surface area, counts of donors and acceptors, shape and electronic terms, or a fingerprint.

2. Fit a model on compounds whose activity you know. Use it to rank compounds you have not made.

3. Modern QSAR is machine learning. The equation became a random forest or a neural network, but the logic is unchanged

What makes QSAR trustworthy?

1. Train and test must be separate

• A model scored on the data it was fitted to tells you nothing.

2. Applicability domain

• The model is only valid for molecules resembling the training set. Ask it about something else and it answers confidently and wrongly.

3. Correlation is not mechanism

• A descriptor can predict well and mean nothing physical.

What elementswill get predicted?

• logP, logD and aqueous solubility

• Permeability (Caco-2, PAMPA)

• Plasma protein binding

• Metabolic stability and CYP inhibition

• hERG liability

• A druglikeness summary, of varying quality

What is reliable to measure?

logP, solubility, simple physicochemical properties. Good enough to rank and to filter

Which elements should be used with care?

Metabolic stability, CYP inhibition, hERG. Right often enough to prioritise experiments, not to replace them.

Which elements are not there yet?

Human pharmacokinetics from structure alone.

What is docking?

IUPAC: ”Docking studies are computational techniques for the exploration of the possible binding modes of a substrate to a given receptor, enzyme or other binding site.”

In docking you try the ligand in many positions and conformations inside the site, then score each one. If you test different molecules/sites in docking, you will find new docking points (3-9 sets of assumptions).

What is the aim of docking?

 For each ligand, find the protein-ligand complex with the lowest energy and estimate energy of binding.

• Molecule conformationally explored in binding site → binding pose

• Scoring of binding pose → estimated binding energy

What do you use docking for?

Binding mode investigation and design

• Structure-Activity Relationship

• Molecular design

Virtual screening

• Can provide new chemical starting points (not novel high affinity ligands)

• Can be seen as a dataset filter before experimental evaluation

What are challenges in virtual screening?

1. Non-optimal binding pocket for all ligands/induced fit van der Waals radii are scaled down to fit binding site

2. Crystallized waters

• Some ligands have bridging waters to target, some do not

• Remove non-structural waters

• WaterMap

• Molecular dynamics

• Multiple crystal structures

3. Rigid target structure in an MM force field is quite far from reality, but still useful!

Do you want to keep or remove waters?

A water in the pocket is either part of the binding site or something the ligand should displace. Keep it or remove it and you get different answers, both of them plausible. You would ideally want to remove bridging water, because if water is incorporated into the molecule, the oxygen molecule will interact with protons. You get a better docking score for compounds that are lipophilic (stickier).

Free energy calculations

Docking scores rank. Free energy calculations try to compute the number itself. Use it late, on a few compounds, to decide which analogue to make next. It is the opposite end of the axis from virtual screening.

What do free energy calculations compute?

•  The difference in binding free energy between two related ligands,

by simulating the physical path from one to the other.

What do free energy calculations cost?

Hours to days of GPU time per pair. Tens of compounds, not millions.

How good are free energy calculations?

Around 1 kcal/mol in favourable, closely related series. That is roughly a factor of five in affinity, which is often enough to choose.

Where do free energy calculations break?

 Large structural changes, flexible or water-filled sites, and any case where the starting structure is wrong.

Density functional theory (DFT)

A common method for chemical reaction modelling

• Calculates the electronic structure of molecules

• No problems to break and form bonds, e.g. calculate transition states, using DFT

• Accurate (relative)

• Fast (relative)

• Great price/performance ratio

What is the definition of a mechanism?

A chain of discrete intermediates joined by transition states. That is exactly what makes it something you can calculate. You know the reagents, the catalyst and the conditions. What you want to know is why this product and not one of the others.

A DFT study, step by step

The same cycle, now with the actual compound in place of R and X. Every intermediate is a structure to optimize, and every arrow hides a transition state to find. The palladium cycle is the energy profile for that cycle. It tells you where the bottleneck sits and why one product wins over another.

What is the idea of machine-learned force fields?

Train a neural network on quantum-chemical energies and forces, then use the network in place of the quantum calculation.

What does machine-learned force fields buy?

Close to quantum accuracy at close to force-field cost. Systems and timescales that were simply out of reach.

Where is machine-learned force fields going?

Reaction modelling, conformational sampling, and binding calculations on systems too large for DFT.

What is the open question in machine-learned force fields?

Transferability. A network is reliable on chemistry that resembles its training set, and confidently wrong outside it.

Why does testing one factor at a time fail?

• Only without interactions: and you rarely know in advance whether you have them.

• Unclear when you are done: the number of runs grows with the variables, and nothing tells you it is enough.

• It answers anyway: a different starting point gives a different optimum. You get a result, and it looks fine.

Full factorial designs

Two levels per variable, a high and a low. You are looking for direction, not curvature. The pattern is identical whatever the number of variables, so you read the design off rather than think it out.

Why do you add center points to a full factorial design?

• Complement the design with center points.

• Possible to discover any curvature in the model → need another type of design.

• Replicates gives information about the reproducibility of the experiments.

How do you select a method?

• Maximise diversity

Factorial or D-optimal designs across the whole descriptor space. Spread the compounds out. You are mapping unknown territory.

• Similarity-based selection

Cluster around a known active, then sample within the cluster. Concentrate them. You are refining a lead you already trust

Diversity is never a property of the molecules. It is diversity with respect to the descriptors you chose, and choosing different ones changes which compounds look different.

It is not only how many compounds you make. It is how well they represent the space you are trying to learn about.

What is generative design?

Models that propose new structures conditioned on a target, a scaffold or a property profile, rather than picking from a catalogue.

What has happened within generative design?

Molecules from AI-driven pipelines have entered clinical trials. The most cited case is rentosertib for idiopathic pulmonary fibrosis, where both the target and the molecule came out of a computational pipeline.

What has not happened within generative design?

No AI-designed drug has been approved. Several early candidates from this route have been discontinued, at ordinary rates, for ordinary reasons.