AI meets zebrafish: is this the future of drug discovery?

New drug candidates must undergo lengthy testing using animal models before being given to participants in human trials. This means many active ingredients must already be sorted out in test series because they either do not have the desired or even have an adverse effect. These test series require many test animals. The use of zebrafish larvae could lead to a reduced demand in this setting.

In this MEDICA-tradefair.com interview, Dr. Jonas Baumann talks about his work with the novel zebrafish larvae model (ZLM). He describes the challenges this entails and explains how artificial intelligence can support the task.

Dr. Baumann, are zebrafish replacing mice in research?

Dr. Jonas Baumann: The ZLM cannot replace rodent testing although it can significantly reduce the latter’s share in preclinical drug research and testing. Newly developed active ingredients can be tested with the ZLM for toxicity, desired mechanisms of action and their effect on target pathogens. This makes it possible to identify promising drug candidates or eliminate less effective candidates before they are being tested on rodents.

The ZLM is not regarded as animal experimentation if the larvae are below 120 hpf (hours post fertilization) according to EU Directive 2010/63/EU. Zebrafish begin independent feeding around 5 days post fertilization. Prior to this stage, they feed on the energy reserves from their yolk sac. This makes the ZLM an animal experiment replacement method as it pertains to the goals of the 3Rs of animal research that seek to Reduce, Replace and Refine. It has enormous potential to reduce the number of rodent tests required in the early drug discovery process in the future.

How can active ingredient testing on fish larvae be scaled to the human organism?

Baumann: Studies have shown that zebrafish share around 70% of orthologous genes (orthologs) with humans. When it comes to genes known to be associated with human disease, 84 percent of human genes have a counterpart in zebrafish. That’s why the zebrafish are genetically far more similar to humans than one might think.

What is the current problem with drug testing/approval in general and with tests using zebrafish larvae in particular?

Baumann: It goes without saying that the ZLM also has its limitations. The larva is tiny, measuring only a few millimeters in length and a few hundred micrometers in diameter. All observations and interventions must take place under the microscope, which requires a lot of finesse and skill. Having said that, thanks to their small size, zebrafish larvae subsequently also need very little space. Like conventional in vitro tests, this makes it possible to perform tests in microtiter plates (multi-well plates) to help save space.

Since their eggs are fertilized externally, it is relatively easy to collect zebrafish embryos. Zebrafish embryos have a transparent chorion, permitting direct observation during embryonic development. Depending on the objective, the embryo must be removed from its chorion. This can be done via mechanical manipulation using a needle, for example, or via protease digestion. Both methods pose a risk of injury to the developing embryo if performed improperly.

Exposure usually occurs by immersing the fertilized egg, embryo, or larva in the test solution. Uptake takes place primarily through absorption through the body surface. Oral uptake is restricted but can be achieved in older larvae by using a gavage. However, this is very complex, making the procedure not well suited for routine testing. Since the larvae have not yet developed gills, the respiratory exposure route is also not an option. One can also administer the active ingredient that’s meant to be tested directly via microinjection into the yolk sac or caudal vein, for example. This approach is primarily used to test the efficacy of antibiotics where the larva has been infected with a target pathogen beforehand. However, the complexity of this makes this method better suited for low to medium throughput rates rather than for high-throughput screenings.

Regulations for animal welfare limits testing to the first five days post fertilization of the embryo or larva, respectively. While this may sound like a very short timeframe, the zebrafish has already formed a more or less fully developed organ and nervous system after the 24 hpf time point. That is why in vivo processes can still be reproduced in an in vitro experiment (the ZLM is appreciated as an advanced cell experiment).

The brief five-day window period only facilitates short-term tests as chronic effects cannot be mapped using the ZLM. That being said, the first few days of life are the most sensitive in a fish's life. Hence, acute toxicity tests always represent a worst-case scenario. If the ZLM could not ascertain a negative effect of a drug candidate, acute toxicity in more fully developed fish is very rare. However, animal experiments are still necessary to entirely rule out the risk of chronic effects from a drug candidate.

How can AI/machine learning play an important role in this setting?

Baumann: Due to the small size of the larvae, even distinct phenotypic effects can only be observed under the microscope. This makes the monitoring and assessment of substance effects via manual analysis very time-consuming, which also limits the number of tests you can run in parallel. Thanks to an automated classification of the larvae tests using a computer-assisted microscopy system and the effect analysis by AI and machine learning this process can essentially run without manual intervention. This makes it possible to identify or exclude more substances as potential drug candidates in less time. The automated test analysis also facilitates classification at any time during the experiment, which lasts up to 120 hours, supporting continuous effect monitoring with no staff. This is designed to promote the early detection of potential adverse substance effects in the embryo or in the larva, respectively. If the AI detects this type of effect, the test can be ended immediately, which would cut short test time significantly.

Additionally, AI is designed to achieve a purely objective effect assessment to prevent human experimenter bias.

What goals are the Helmholtz Institute for Pharmaceutical Research Saarland (HIPS) and the CISPA Helmholtz Center for Information Security trying to achieve in the ImageTox project?

Baumann: At the end of the project, HIPS researchers should have a tool that allows them to rapidly test a multitude of new drug candidates in the larvae using an automated high-throughput screening method to identify promising drug candidates at an early stage.

ImageTox is developing a new methodology for image analysis and the analysis of time series, in particular, which is where modern AI methods come into play. To a large extent, the methodology is not specific to the explicit application. Current planning intends the development and evaluation of the methodology for the larvae. We are also interested in exploring applications in other fields.