TLDR: This paper explores how self-supervised learning (SSL) can improve high-throughput toxicity testing for new chemicals and materials. It addresses challenges like limited labeled data and the need to identify subtle, continuous toxicant-induced changes. Using the EmbryoNet dataset of zebrafish embryo phenotypes, the researchers demonstrate that SSL can effectively learn representations that distinguish between different modes-of-action. The paper also discusses integrating these machine learning models into a physical testing device called TOXBOX, highlighting SSL’s potential for more efficient and ethical toxicity assessment.
The world is constantly introducing new chemicals and materials, and ensuring their safety is paramount. Regulations like REACH in the European Union mandate extensive toxicity testing for compounds entering the market. However, traditional methods, often involving animal testing, are costly, time-consuming, and raise ethical concerns. This has spurred a search for faster, more efficient, and ethical alternatives, known as New Approach Methodologies (NAMs).
NAMs include techniques like using zebrafish embryos or Daphnia magna, which are cheaper to maintain and reproduce quickly. In vitro tests using cell-based assays or organ models are also gaining traction. These high-throughput screening (HTS) methods generate vast amounts of data, making automated evaluation via machine learning models essential.
While machine learning has been applied to toxicology, particularly in silico models, there’s a growing need for automated evaluation of experimental HTS data. Deep Learning (DL) models are particularly well-suited for this due to their ability to handle high-dimensional data, such as microscopic images and time-series data.
The Power of Self-Supervised Learning
A significant challenge in developing robust machine learning models for toxicity testing is the scarcity of labeled data. This is where Self-Supervised Learning (SSL) offers a powerful solution. SSL methods learn useful representations from the data itself, without requiring explicit human-provided labels. This means models can be pre-trained on large amounts of unlabeled data, making them more label-efficient for specific downstream tasks like classification.
Beyond label efficiency, SSL provides several advantages for toxicity assessment:
- Continuous Representations: Toxicant-induced changes are often continuous, not discrete. Traditional classification models struggle with this, often relying on subjective cutoff points. SSL maps samples into a latent space that allows for continuous representations of morphological changes, effectively modeling concentration-dependent gradients of toxicity. This means that as toxicant concentrations increase, the learned representations move further away from those of healthy phenotypes.
- Identification of Similar Modes-of-Action: SSL inherently clusters similar inputs together. This property can be leveraged to identify compounds that cause similar biological effects or ‘modes-of-action’. If a new compound induces a phenotype similar to known toxicants, its representation will cluster with those known compounds, providing valuable insights.
- Detecting Unknown Effects: Unlike classifiers that force unknown changes into predefined categories, SSL can map representations of unknown morphological changes away from known classes, making it apparent that a novel effect has been observed, which warrants further investigation.
Proof-of-Concept with Zebrafish Embryos
To demonstrate these capabilities, researchers conducted a proof-of-concept using SimCLR, a popular SSL method, and the publicly available EmbryoNet dataset. This dataset contains images of ten different zebrafish embryo phenotypes, including normal development and various toxicant-induced changes affecting major signaling pathways.
The SimCLR model was trained to learn meaningful representations of these images. After training, a linear classifier built on top of these learned representations achieved an accuracy of 79.9% in classifying the different phenotypes. While slightly lower than a fully supervised baseline, this performance is still acceptable and highlights the quality of the representations learned without explicit labels.
Visualizations of the learned representations showed clear clustering, particularly for distinct phenotypes like ‘Dead’ embryos, which were mapped far from other classes. This indicates that the model successfully learned to distinguish between different toxicant-induced changes and group similar modes-of-action.
Also Read:
- Advancing Sleep Monitoring: How Self-Supervised Learning Makes Wearable EEG More Efficient
- Ensuring Safe AI: A Look at World Model Pathologies in Embodied Agents
Integrating AI into TOXBOX
The ultimate goal is to integrate these advanced machine learning models into physical toxicity testing devices. The TOXBOX project, for instance, aims to create an all-in-one platform for reliable toxicity testing, featuring in vitro organ models and a zebrafish embryo module. The strategies discussed in this paper, especially SSL pre-training followed by fine-tuning, are considered the most viable options for TOXBOX.
This approach is particularly beneficial when data and labels generated by the TOXBOX device are scarce. Furthermore, the latent space learned by SSL models can provide deeper insights into tested compounds, helping to determine if a compound has a known mode-of-action or an entirely new, unknown effect that requires more thorough investigation.
However, integrating ML models into real-world devices also presents challenges, such as ‘concept drift’ – gradual changes in underlying data over time that can degrade model performance. Continuous monitoring and strategic retraining of models will be crucial to ensure reliable predictions.
In conclusion, self-supervised learning offers a promising pathway to address critical challenges in toxicity testing, enabling more efficient, ethical, and insightful evaluation of new chemicals and materials. For more details, you can refer to the full research paper here.
