Accurate prediction of the clinical effects of compounds in preclinical development would considerably improve the efficiency of drug discovery and reduce costly late-stage failures. Writing in Nature Chemical Biology, Fliri and colleagues show that the side effects of known drugs can be linked to their effects in in vitro protein-inhibition assays, providing a potential strategy for predicting the clinical effects of novel compounds early in preclinical testing.

The current study builds on earlier work by the same group in which they analysed the in vitro interactions of 1,567 structurally diverse compounds with a panel of 92 proteins representative of the 'druggable' proteome. Compounds were screened at a single high concentration in ligand binding assays so that the percent inhibition values obtained provided an indication of the potential of each compound to interact not just with the proteins in the panel, but with those in the same family.

These percent inhibition values were then translated into biological activity spectra — biospectra — that express the probability of a particular compound inducing a certain pattern of protein-network perturbations. Comparison of the biospectra of different compounds showed that compounds with similar structures and pharmacological effects clustered together, and that biospectra of novel compounds could be predicted from their molecular structure.

Encouraged by these results, the authors set out to investigate how predictive biospectra are of the clinical effects of drugs. First, using information from the labels of 1,045 drugs, the authors derived 'side-effect spectra' based on the classification of the drugs according to 591 types of side effect that they can show, such as nausea. Analysis of these spectra revealed that compounds with similar structures or protein-network perturbations clustered with each other.

Then, taking a subset of the drugs for which the data necessary to create biospectra were available, Fliri et al. demonstrated that there was a meaningful relationship between the way that drugs clustered according to their biospectra and the way that they clustered according to their side-effect spectra. So, this approach provides a mechanism for linking the molecular structure of compounds, their properties in simple biological assays and their clinical effects.

The key to improving the quality of predictions of the clinical effects of novel compounds from their structure and effects in in vitro assays will be to improve the amount and quality of data on known compounds used in the method. So, as the authors highlight, the potential to reduce the risk of failure in clinical development using this method could provide incentives to create public databases containing preclinical, clinical and safety information not just on marketed medicines but on those that fail in clinical trials.