Artificial Intelligence Disrupts Drug Testing for Oncology

Sophisticated computer analytics can be tuned to predict the results of laboratory testing and thus render drug development both faster and cheaper. In pharmaceutical research, the quest for new drugs involved above all a very large number of lab tests. If, for example, basic research has identified a protein that has a role in tumor growth, often tens of thousands of active ingredients need to be tested in hopes that one of them will alter the protein’s function so that tumor growth will stop. At the same time, lab testing reveals some potential harmful side effects on the body. In other words, in order to find viable active substances, researchers need to generate huge quantities of data and review them for certain patterns. This sounds exactly like a computational task – and indeed, pharmaceutical research is about to change fundamentally and undergo unprecedented acceleration with the emergence of AI.

Especially R&D targeting complex systemic diseases stands to benefit. Large-scale assays for drug discovery are expensive and often impractical, particularly for the increasingly important physiologically relevant model systems that require primary cells, organoids, whole organisms, or expensive or rare reagents, as a recent article, Repurposing High-Throughput Image Assay Enables Biological Activity Prediction for Drug Discovery, demonstrates in Cell Chemical Biology. Oncology is morphing increasingly into a subspecialty of bioinformatics.

Time, effort and expense of conventional screening methods in the example described above stems from the fact that researchers need to test the interactions of each molecule under consideration with the protein they seek to influence, which translates into tens or hundreds of thousands of rounds of testing.

Deep Learning has been a prominent tool in toxicity prediction for some time. One hypothesis of the cited study was therefore that the pathway to a viable drug could be shortened significantly by having AI conduct specific preliminary selection. Instead of running a complete set of tests for each molecule, a different source of information might be capable of higher levels of automation: analyzing image data, which computers are already pretty good at interpreting. This led to the idea of a scalable machine-learning-based method predicting compound activity from images. The study hypothesized that data from a single high-throughput imaging assay could be repurposed to predict the biological activity of compounds in other assays, even those targeting alternate pathways or biological processes.

In the process tested here, cell cultures treated with the active substance under review are imaged using high-throughput microscopy. Preparation of cell lines is inexpensive and could also be performed by robots in the future. Cells are then dyed and photographed under the microscope. This is not only vastly more expeditious but also costs just $10-15 per picture.

Pictures are not evaluated by humans but by AI: learning algorithms receive data feeds from databases of already known substances and of their effects at cellular level. With this knowledge in store, software now scours through all new cell images and proposes a handful of the most promising options based strictly upon morphological changes observed in the cell structure. Quantitative information extracted from a three-channel microscopy-based screen for glucocorticoid receptor translocation was able to predict assay-specific biological activity in two ongoing drug discovery projects.

AI is aware that, whenever cells present a certain picture, laboratory tests have also shown a positive result in the past. The human eye is not good at systematically detecting those visual patterns. Which specific criterion AI uses to determine the outcome – whether it is the form of cell nuclei, the positioning of the cells with respect to one another, or other indicative patterns – remains as inscrutable and unpredictable as human intuitive processes, but its efficiency may be measured ex post. The study identified 125 promising molecules from a total of 500,000 candidates for an oncological application: hundreds of assays predicted by one image screen annotating half a million compounds. Subsequent random lab tests confirmed AI’s results in almost all cases. Repurposing increased hit rates 50- to 250-fold over that of the initial project assays while increasing the diversity of chemical structure of the hits. Connections discovered in the process are in many cases very surprising to human researchers as image-based models boosted hit rate and diversity in two drug discovery projects. Even if this method does not produce biological knowledge per se, it nonetheless provides starting points with excellent odds for further research. The study concluded that data from high-content screens are a rich source of information that can be used to predict and replace customized biological assays. It is expected that drug development will accelerate substantially as a result of involving AI-based screening and testing methods such as this, using wide varieties of pattern recognition. The instant study represented just a proof of concept justifying further work on image-based learning for drug discovery. It will also become more accurately targeted because problematic side effects can be predicted earlier. Of course, clinical trials will still not become obsolete until simulation of entire human organisms and their considerable varieties becomes available, which is why it will still take some time until the first AI-based drugs will come to market.