De Novo Drug Design – Do We Really Want to Be “Original”?

An Attention-basesd Conditional Variational Autoencoder (ACoVAE) architecture was employed to generate novel putative colchicine site binders of tubulin. Generation relied on commercially available “seed” compounds, predicted to be active according to a QSAR model calibrated on the pertinent structure-activity data for this key anti-cancer target. This QSAR model followed by docking served to score the generated compounds in terms of their propensity to bind tubulin. Generated compounds included many molecules matching the structural patterns seen in the seed compounds and closely related to other commercially available compounds. Four of the latter were used for a soaking experiment of tubulin crystals, with no success. Genuinely original and intriguing molecules were also generated by the ACoVAE, some containing the “iconic” seven-membered ring typical of colchicine, albeit this was not seen in the seed compounds. However, these compounds are difficult to synthesize, and project partners considered the trustworthiness of the models that selected them to be insufficient to justify the high synthesis costs. This highlights the actual bottleneck limiting the breakthrough of de novo compound design: the in silico scoring and assessment of the candidate compounds.