British-U.S. team evaluates computational methods vs. fragment screening for drug discovery

Designing new drugs is a challenging but crucial task for preventing and treating human diseases.

The basic method is to search for molecules that will activate or deactivate a biological molecule (e.g. a protein) by binding to it and thereby provide benefit to the patient. An important technique in the drug discovery process is fragment screening, where molecules with low molecular weight are tested to see how they bind to proteins.

This process needs to be high throughput in order to provide the largest possible number of molecular candidates for further testing and development. ITC, SPR, and NMR are all great biophysical techniques that have been used historically for fragment screening, but they are not high-throughput. X-ray crystallography remains the leading approach in the field but it still suffers from a number of issues.

Due to problems with physical techniques, computational modeling has become essential for the whole drug discovery pipeline. 

Now a team of researchers from the United States and the United Kingdom has devised and carried out an evaluation on the use of computational methods. Their work appears in the Journal of Computer-Aided Molecular Design, April 15.

The computational challenge

The Statistical Assessment of Proteins and Ligands, known as SAMPL is an international collaboration in which industrial or academic participants use computational trials to predict the results of experiments in a fully blinded fashion. In this way there can be an objective assessment of the efficacy of computational vs. physical fragment screening methods. The SAMPL participants have been evaluating many physical and chemical properties involved in drug design since 2008.

This is the first SAMPL challenge in which the research team performed fragment screening for a particular protein called PHIP (human pleckstrin homology domain interacting protein) and used their results to evaluate computational methods. The researchers used crystallographic fragment screening to look for binding sites.

PHIP2 is known to be involved with the insulin signaling pathway and other functions, and is potentially involved in causing tumors. The researchers found "52 different fragments bound across four distinct sites, 47 of which were bound to the pharmacologically relevant acetylated lysine [Kac] binding site."

They then set up a three-stage SAMPL challenge for participants around the world to use computational methods to predict the results of the team's screening experiments. This challenge is known as SAMPL7.

In Stage 1 of the challenge, the researchers gave participants one month to categorize binders from non-binders using computational analysis. Laboratories in China, Japan, Spain, and the United Kingdom participated.

Stage 2 followed up the first stage, and involved prediction of binding poses for the known binders. Stage 3 involved "suggestions of follow-up molecules that might improve affinity." This last stage, they note, was interrupted by the COVID-19 pandemic because funds had to be diverted.

Challenge results

The results of SAMPL7 demonstrated that for these types of fragment screening computational methods still have a ways to go.

The researchers report computational methods are "better suited for larger, drug-like molecules and there is a general lack of prospective and validation/evaluation in fragment space."

"All stages of this SAMPL edition were clearly challenging," the researchers write. They note some of the factors that may have accounted for the "relatively poor performance" of the computational analyses, including the small number of submissions and the short time periods for the challenge. They also suggest areas for improvement in computational prediction, including more attention to water molecules and molecular dynamics.

------

H. Grosjean et al., SAMPL7 protein-ligand challenge: A community-wide evaluation of computational methods against fragment screening and pose-prediction, Journal of Computer-Aided Molecular Design, April 15, 2022.

DOI: https://doi.org/10.1007/s10822-022-00452-7