– Fall 2007

In a recent study performed by researchers at Eli Lilly, Phase outperformed Catalyst HypoGen in a head-to-head study evaluating the predictive value of each program's 3D QSAR models.¹ Working with data sets that would be appropriate for a lead-optimization project, the authors found that Phase produced superior results across the majority of test cases, and was able to generate a reliably predictive model more frequently than Catalyst HypoGen.

The ability to correctly rank-order compounds is of obvious use in lead optimization, and is significantly more challenging than identifying a compound as being either active or inactive. While there are a variety of techniques that can be used to create QSAR models, the Lilly researchers noted that the 3D QSAR models created by pharmacophore-based approaches lend themselves to easy interpretation. This makes it comparatively easy to suggest new compounds for synthesis, which is particularly beneficial in lead optimization projects where short turnaround times are essential.

With an eye toward this need for practicality, the researchers created 3D QSARs in Phase and Catalyst HypoGen using a variety of automated workflows suitable for an actual production environment. By and large, default parameters were used in both programs, indicating that the published results are comparable to what users might expect to see with out-of-the-box performance.

The authors took similar care to construct meaningful data sets, beginning with their selection of eight pharmaceutically relevant targets. At each target, ligand activity data was assembled using measured values from a single reliable publication, and ligand sets were chosen to resemble those that might be seen in the lead-optimization phase of a project. When constructing QSAR models, an automated procedure was used to select training sets, thereby eliminating subjective bias.

Upon comparing the best hypotheses found by either program at each target, the authors found that Phase was consistently able to match or outperform results from Catalyst HypoGen. Phase produced a model with an equivalent or superior R² at 7 of the 8 targets, and a superior r_p for 7 out of the 8 targets. The authors attributed much of the advantage to Phase's ability to create 3D QSARs using an atom-based scoring grid – while Phase is capable of constructing 3D QSARs using either grid-based or feature-based QSARs, only the latter method is available in Catalyst.

The authors acknowledge that an expert user might obtain superior results by varying training set definitions or the parameters used to construct the QSAR models – indeed, Phase’s tools for identifying multiple binding modes were not used in any of the cases in this study, although there were multiple targets and data sets that may have benefited from their use.

However, this does not detract from the importance of the type of workflow the authors describe – given that lead optimization projects see the frequent addition of new compounds to data sets, there is an obvious need for QSAR tools that inform the discovery process with minimal turnaround time. The authors’ conclusion that “the performance of Phase is better than or equal to that of Catalyst HypoGen” suggests that Phase is a prudent choice for such a tool.

¹ Evans, D.A.; Doman, T.N.; Thorner, D.A.; Bodkin, M.J., “3D QSAR Methods: Phase and Catalyst Compared”, J. Chem. Inf. Model, 2007, 47(3), 1248-1257.