The ability to efficiently screen large libraries compounds in order to find candidates that will exhibit a specific gene expression profile of interest is practically very useful, and doing so without requiring experimental data is very appealing. Here we propose a new approach that creates shortlists of compounds for a query gene expression signature. This approach uses a previously proposed variant of the k-NN procedure (called variable-NN) where a test signature results from the weighted average from the signatures of surrounding neighbours. However we propose an added modification to this method, where predicted signatures are submitted to a quality (or confidence) criterion based on standard deviation of predictions. Two transcriptomics datasets were used in this work, LINCS and a Novartis internal transcriptomics panel (PANOMICS), and two different similarity measures were tried, calculated from structural and biological (predicted) signatures, in order to explore the feasibility of a recently available in silico biological signature. The signatures for the full dataset, obtained from the variable-NN procedure, were ranked by similarity to target (desired) signatures. This process showed to consistently enrich the top of the ranked compound lists with the compounds that match each of the target signatures. Obtained median ranking percentiles of true compound-signature matches, which were originally consistent with random sorting when confidence is not taken into account, were as high as 90% when confidence was used to correct similarity. This translates into obtaining shortlists formed of the top 10% of ranked compounds that are likely to contain the best compound candidate. This process allows filtering target signature queries according to reliability in such a way that correlates with compound enrichment. The use of confidence criteria to filter signatures was pivotal in the ability to enrich the top of the ranked lists with the compounds correctly matching the target signatures. This approach allows locating candidates for a target gene signature purely from in silico information, thus enabling high-throughput virtual screening of compounds libraries to find compounds associated with a gene expression profile of interest.

Gearing transcriptomics towards high-throughput screening: Compound shortlisting from gene expression using in silico information

256th National Meeting (2018)

The official archive and repository for all ACS Meetings & Expositions abstracts and select posters or presentations.