Idea of an artificially recreated human.
Graphic: Gerd Altmann, Pixabay.

In this issue, we focus on Dr. Ignacio Tripodi of Sciome LLC. based in Research Triangle Park, North Carolina. He is a bioinformatician developing computer programs for contemporary risk assessment that will lead to the reduction and, hopefully, long-term replacement of animal testing.

The researcher has developed a computer system to elucidate acute toxicity mechanisms without animal testing. The system is called MechSpy. It generates the most likely toxicity mechanisms by combining data from human biology, toxicology, and biochemistry with those of gene expressions from cells of human tissue. The model can predict simple, lucid toxicity mechanisms for both well-studied compounds and other chemicals for which the toxicity mechanism has not been as well known. The MechSpy computer model can be extended to include additional toxicity mechanisms or other human biology mechanisms.⁽¹⁾

At the EUSAAT congress in Linz, the scientist presented another recent work. For this purpose, he uses information from so-called Adverse Outcome Pathways to let the computer analyze the effects of chemicals on the organism.

In the concept of Adverse Outcome Pathways, biological processes such as cell receptor binding of molecules lead to harmful effects in the organ, organism, or population via several sub-steps.

Schematic representation of an Adverse Outcome Pathway (AOP) with hypothetical events.
Adapted from AOP Wiki, CC BY-SA 3.0.

An AOP describes a sequence of events that begins with an initial interaction of a stressor with a biomolecule in an organism that causes a perturbation in its biology (i.e., a molecular initiating event, MIE). That interaction may progress through a dependent series of intervening key events (KE) and culminate in an adverse outcome (AO). This process is considered relevant to risk assessment or regulatory decision-making. In this process, the biological change has to be measurable and the disturbance must be severe enough in terms of strength, duration, and frequency to follow the path to a harmful outcome. The AOP concept does not describe every detail of the biology, but rather only the critical steps or checkpoints on the path to disruption.

The concept can be used to identify the so-called molecular initiating events (MIE) and key events (KE) in cells, tissues, or organs to understand the response of chemicals, for example. A MIE is a specific type of KE that is the starting point of an interaction between a chemical and a stressor at the molecular level within the organism that results in a perturbation that initiates an AOP. The KE, on the other hand, is a change in the biological or physiological state in the cell or tissue that can be measured and is essential for the progression of a defined biological perturbation leading to a specific adverse outcome.

The identification of AOPs is based on an international collaboration of scientists, hereafter the pathways are subject to a revision by the OECD Extended Advisory Group on Molecular Screening and Toxicogenomics (EAGMST). The data already explored are freely available via AOP-Wiki.⁽²⁾

The AOP concept is also an important basis for the new approach to risk assessment currently being developed and it is considered one of the major scientific advances of the last 10 years. It is expected to lead to greater use of non-animal methods such as cell cultures, organoids, organ-on-a-chip models, or in silico computer processes.

In silico
In silico refers to processes that take place in the computer. Computer programs are used to elucidate the biochemical processes of living organisms, especially humans. Such methods are also increasingly suitable for reducing or even replacing animal experiments. While in the past the computer was only a data storage device, nowadays new insights into the functionality of nature, and the physiology of cells, organs, and living systems can be gained, which could not be observed or discovered so easily. Computer-aided simulations are carried out to find processes taking place in cells, for example, which can then be reproduced in vitro and, if possible, confirmed. Conversely, findings in cells are either confirmed or refuted with the help of computers.

Meanwhile, there is a multitude of different computer models available. The classic case in bioinformatics is a sequence analysis, where a DNA sequence, for example, can first be identified by comparison using sequence databases, e.g. BLAST.⁽³⁾ Computational chemistry, for example, aims to use computers to calculate the properties of molecules. Software programs are being developed for this purpose.⁽⁴⁾ The goal is to use computer models to search for a treatment
method or the like. With the help of this identification and others, such as the so-called QSAR Toolbox⁽⁵⁾, the properties of similar chemicals can be found without having to resort to animal testing. The QSAR Toolbox is a freely available software that supports the identification and filling of data gaps in (eco)toxicity for hazard assessment of chemicals. In this process, chemicals are classified. Chemicals can be identified based on user-selected interaction mechanisms or molecular characteristics.

Bioinformatician Dr. Ignacio Tripodi uses such databases and, to ensure that the computer draws the right conclusions, makes the computer process the natural language of humans using the so-called Natural Language Processing (NLP). NLP is a branch of artificial intelligence. In this case, it enables the computer to convert all Adverse Outcome Pathways that have already been researched and confirmed as correct into testable hypotheses. The goal is to maximize the information potential of AOPs.

 Graphic: Gerd Altmann, Pixabay.

This is a realistic perspective to deal with a large number of substances to be retested under the European "Green Deal". This could never be done on animals due to time resources, cost, scale, and ethical reasons.

European Green Deal
The European Commission adopted its Chemicals Strategy for Sustainability on October 14, 2020. The strategy is part of the EU's zero-pollutant target - a key commitment of the European Green Deal. It aims to better protect citizens and the environment from harmful chemicals and drive innovation by promoting the use of safer and more sustainable chemicals.

The switch to chemicals and production technologies that require less energy should lead to a limitation of emissions. This means that the Green Deal needs the "right" chemistry.⁽⁶⁾ One of its ambitious chemical strategy goals is to increase the amount of information available on potentially hazardous chemicals while making the best use of innovative methods that do not involve animal testing. These include techniques that use human cells, computer models, and organ-on-chip devices.⁽⁷⁾

 
The next few years will be dominated by computational toxicology. Dr. Tripodi is not the only one who is convinced of this; many researchers worldwide agree. First, however, there are still numerous challenges to be overcome regarding the quality of the developed models and the availability of information in databases. For example, most of the data in databases traditionally come from previous animal experiments.

Based on the molecular triggering events, researchers are developing platforms to share and quality check with other scientists. There are also databases available in the USA, such as the CompTox Chemicals Dashboard. Their developers are confident that they can phase out animal toxicity testing by using the CompTox Chemicals Dashboard instead.⁽⁸⁾ Together, they have accumulated enough data, according to a webinar held on the topic in the fall. CompTox provides easy access to chemistry, toxicity, and exposure information for more than 900,000 chemicals. Using the dashboard, users can access high-value chemical structures, experimental and predicted physicochemical properties of chemicals, environmental fate and transport, and toxicity data. The data comes from both traditional animal studies and high-throughput screening. The latter information is based on studies using live cells or proteins that had been subjected to chemical tests. The dashboard can be searched by chemical name or other identifiers, consumer product categories, or even high-throughput screening data. CompTox is intended for chemical toxicity information, not drug efficacy information.

The more intensive work is done with expert systems, databases, or Artificial Intelligence (AI), the more necessary it is that young scientists are also trained in computer science and, if possible, learn to master this discipline. Therefore, developers like the young scientist Ignacio Tripodi are urgently needed. A big problem at the moment is still explaining to laymen how to use such programs.

 "... AI can infer relationships today with impressive precision and recall..."
InVitro+Jobs interviewed Dr. Ignacio Tripodi about his work and perspectives on computational toxicology.

InVitro+Jobs: What experiences led you to want to become a bioinformatician?

Dr. Tripodi: I started with a computer science career where I was lucky to work on increasingly challenging and complex problems, but despite being in a very comfortable position, at the end of the day I felt that I was not using everything I learned to solve large, global problems. I was also curious about biology but after an established career in computer science, I thought that ship had sailed. However, after becoming increasingly involved in local humane societies and animal welfare issues in general (for example, the volume and type of animal use in research), I discovered the world of bioinformatics and computational modeling to answer questions in biology and chemistry, and my career took a detour towards science to help make a difference in the reduction and replacement of animal models. I've been very fortunate to be admitted to the Interdisciplinary Quantitative Biology Ph.D. program at the University of Colorado where, besides my computer science research and requirements, I also had to sit in classes like molecular biology, genetics, or microscopy, do a 10-week hands-on "wet lab" rotation, etc. All of that gave me the opportunity to become "bilingual" in both the computational and biology worlds, gave me the tools to learn about toxicology, and empowered me to understand the nuances of the problems I'm trying to solve far beyond the computational details.

InVitro+Jobs: What research question are you currently working on?

Dr. Tripodi: There are multiple questions I'm currently addressing, but I could say the all-encompassing one would be: how far can we push the automatic generation of a mechanistic hypothesis from a combination of experimental results and our existing body of curated knowledge? How far can a combination of cutting-edge machine learning techniques and more traditional semantic artificial intelligence take us, to help explain our observations?

InVitro+Jobs: Which regulatory processes are supported by your development, and which endpoints are addressed?

Dr. Tripodi: For regulatory applications, I'm addressing both the development of novel adverse outcome pathways (AOPs) and the verification/use of these AOPs. We're giving toxicology researchers a tool to obtain a quantitative measure of enrichment of an AOP (and individual AOP events) in a reproducible manner and to provide a putative explanation for that result based on their experimental data. These tools aid both the process of hypothesis generation and of exploratory research for new small molecules, to better understand why a specific type of exposure leads to an adverse outcome in certain tissues. This mechanistic inference framework is not limited to a specific endpoint, and it's generalizable to other pathway definitions beyond AOPs.

InVitro+Jobs: Automated workflows combined with human control are more like expert systems than AI, or how do you see this?

Dr. Tripodi: These days "AI" is being used a lot in very different contexts as a buzzword, with a very wide range for interpretation. Expert systems are also a type of AI, as are the machine learning methods I use to extract a computer-friendly representation of concepts from massive semantic knowledge graphs^(a), or the computational linguistics methods to map AOP events to these concepts. We could say it's a combination of different types of AI techniques, ranging from traditional knowledge representation frameworks to language models and deep learning. Human control will always be present, this is not a tool intended to replace the researcher but rather a "virtual assistant" to collaborate with.

InVitro+Jobs: What do you mean by a knowledge graph?

Dr. Tripodi: I would define a knowledge graph as an integration of multiple, heterogeneous sources of knowledge (ontologies, databases, etc) into a directed graph. This network-like representation allows us to connect the dots between concepts in ways that were not obvious by looking at each database individually.

InVitro+Jobs: Who checked the quality of the data in the databases? Can the AI do this itself in comparison?

Dr. Tripodi: Different public resources have different annotation criteria.^(b) Many of the open biomedical ontologies^(c) we use, for example, rely on agreement from multiple human curators to annotate links between proteins and the concepts represented in their ontology. Many also use computationally-inferred relations derived from a wide array of methods (e.g. computational predictions of interactions between proteins, relations between genes and phenotypes derived from parsing^(d) the literature using natural language processing, etc). While it will never be possible to ensure perfect accuracy and lack of bias from AI-derived relations (the same can be said of the manual annotations, though!), AI can do this today at impressive measures of precision and recall. For the knowledge representation systems we are building, I'm only using annotations that have been vetted by human eyes, and refining them with our own data-driven methods only at a very high confidence level, statistically speaking.

InVitro+Jobs: How can it be verified that the AI is not off in the end?

Dr. Tripodi: It always helps to work with human validators that can at least spot-check the results as much as the volume of data allows. The computational methods themselves also have ways to ensure we prevent "overfitting" the model to a given training dataset, and make it as generalizable as possible.

InVitro+Jobs: According to my information, animal data is still used in the meantime because there is not enough human data to build the models. How do you deal with the fact that most of the data are still from animals (species differences)?

Dr. Tripodi: As statistician George Box wrote, "all models are wrong, but some are useful". We simply make the most of the existing nonhuman animal data by recovering what we can from homology. The assumption that our observations from animal models in 'omics data will directly apply to humans by homology is a big one, given our many differences in physiology. We always lose information when projecting those results into human biology. However, there is also a lot of human data available for a wide variety of tissue types in public repositories, and my entire mechanistic inference^(e) work for years has been centered around human-specific data. In toxicology, for example, a lot of the exposure data has traditionally come from in vivo animal models, however, the number of in vitro data available using human tissue increases every year. The quality of organoid and "tissue-on-a-chip" models has improved tremendously in just a few years, and they produce data that is reliable, physiologically relevant, and directly applicable to humans.

InVitro+Jobs: What are the upcoming challenges in computational toxicology?

Dr. Tripodi: I think the most important next step is perhaps not technical but rather a paradigm change: For any kind of computational model, one of the most important components is a "gold standard" dataset that you can rely on, to assess the performance of your model and compare against different ones. For years, animal model data has been used as the gold standard despite translational issues to human biology and the fact that it often cannot even reproduce itself between different labs. Therefore, we have the need for the curation of a good-quality, public, human-centered 'omics gold standard dataset for different types of perturbations using the proper tissues, that can be used by computer scientists to create and evaluate models. In an ideal scenario, these experiments would also be conducted as a time series to aid the mechanistic inference process and use organoids.

InVitro+Jobs: Thank you very much for the interview.

Dr. Tripodi: Thank you for the opportunity to talk about this research for a wider audience!

Literature:
⁽¹⁾ Tripodi, I.J., Callahan, T.J., Westfall, J.T., Meitzer, N.S., Dowell, R.D. & Hunter, L.E. (2020). Applying knowledge-driven mechanistic inference to toxicogenomics. Toxicol In Vitro 66:104877. doi:10.1016/j.tiv.2020.104877. Epub 2020 May 6. PMID: 32387679; PMCID: PMC7306473.
⁽²⁾ https://aopwiki.org/
⁽³⁾ Dandekar, T. & Kunz, M. (2021). Bioinformatik. 2. Auflage, Springer Verlag.
⁽⁴⁾ http://www.chemgapedia.de/vsengine/glossary/de/computerchemie.glos.html
⁽⁵⁾ https://www.oecd.org/chemicalsafety/risk-assessment/oecd-qsar-toolbox.htm
⁽⁶⁾ Europäische Chemikalienagentur (ECHA). Chemikalienstrategie für Nachhaltigkeit.
https://echa.europa.eu/de/hot-topics/chemicals-strategy-for-sustainability
⁽⁷⁾ https://joint-research-centre.ec.europa.eu/jrc-news/jrc-supporting-alternatives-animal-testing-
2022-04-07_en
⁽⁸⁾ https://www.epa.gov/chemical-research/comptox-chemicals-dashboard

Notes:
^(a) A knowledge graph is, at its core, a knowledge-based system consisting of a knowledge base and a machine that can perform deductive logic over the knowledge base to infer knowledge that would otherwise be hidden (https://blog.vaticle.com/what-is-a-knowledge-graph-5234363bf7f5).
^(b) In an annotation process, some upper meanings are represented by decision trees. On one hand, these decision trees are used to arrive at specific sub-meanings by applying individual criteria, but by making these decisions, they also make it possible to determine the super-meaning in particular. (From: Kiss, Tibor et al. 2020. A handbook for the determination and annotation of prepositional meanings in German. ISSN: 2700-8975)
^(c) An ontology in data management is the usually linguistic and formally ordered representation of a set of terms and the relationships between them.
^(d) Parsing: syntax analysis.
^(e) Inference: If, with the help of already known facts, something is said about still unknown facts, this is called inference.