Sometimes the future arrives unexpectedly. For me, it began on an airplane in 2025, while browsing the in-flight movie options and stumbling across The Thinking Game, a documentary following Demis Hassabis and Google DeepMind’s pursuit of artificial general intelligence. At the center of the film was AlphaFold, the AI system that transformed biology by predicting the three-dimensional structures of proteins from their amino acid sequences, a feat that helped crack the decades-old protein-folding problem and opened new frontiers in drug discovery.
Just a year later, that same technological leap took center stage at the 2026 SOT Annual Meeting, where a new “Global Session” convened scientists from across Asia and the United States to examine what structure-based toxicology might look like in the AlphaFold era.
Led by Dr. Robyn Tanguay of Oregon State University, the session was designed to address a problem that is at once scientific, regulatory, and international. Her chosen topic for this year was structure-based toxicology in the AlphaFold era. The session drew a strong audience and sustained attention throughout, a sign of the field’s growing interest in artificial intelligence as both an opportunity and a challenge.
AlphaFold itself is more than a scientific curiosity. Developed by Google DeepMind, the system predicts 3D protein structures with remarkable accuracy and has rapidly become a powerful in silico new approach methodology, also known as a NAM. In toxicology, those structural predictions can help researchers anticipate chemical interactions, identify potential drug targets, and assess toxicity pathways without relying solely on animal testing. The appeal is obvious: if scientists can better understand how a chemical binds to a protein, they may better predict what that chemical will do in a living system.
Dr. Kazuki Takeda of Kitasato University opened with a sweeping view of cytochrome P450 enzymes, a superfamily central to xenobiotic metabolism and toxicology. Genome sequencing efforts across diverse organisms have now identified more than 15,000 P450 genes, and structural biology has made significant progress since the first x-ray crystal structure in 1987. Yet, even with more than 110 P450 structures deposited in the Protein Data Bank, many important enzymes remain structurally unresolved. Here, AlphaFold offers a major advance.
Dr. Takeda showed that AlphaFold models provide high backbone accuracy and can be particularly useful for already characterized CYP enzymes, but limitations remain significant. Heme, the essential cofactor for P450 function, is not included by default. The models also fail to fully capture the open and closed conformational states that are central to enzyme activity, and confidence scores near active sites are often low. Tools such as AlphaFill can restore cofactors, and the availability of more than a million structural models is undeniably transformative. Even so, Dr. Takeda argued, AlphaFold is not a replacement for structural biology. Rather, it is a catalyst—one that must be paired with simulations and experimental approaches to capture the dynamic behavior of these flexible enzymes.
If Dr. Takeda highlighted structural opportunity, Dr. Yijie Geng of the University of Washington showed how those opportunities can be translated into biological discovery. His talk asked a provocative question: do the toxicological targets researchers already know explain all toxicity? To probe that question, Dr. Geng’s lab built a reverse molecular docking pipeline to screen toxicants against the predicted 3D human proteome. Using the HProteome-BSite database, which maps predicted binding pockets across more than 80 percent of UniProt entries represented in the AlphaFold human proteome database, his team explored how PFAS compounds might interact with human proteins.
The results pointed to folate receptors as mediators of PFAS toxicity. Through a combination of in silico, in vitro, and in vivo validation, the group identified folate receptor beta as a biological target of PFNA, PFOA, and PFOS. Their experimental work extended into zebrafish, where the “Fishbook Social Behavior Assay” provided a quantitative measure of social behavior. Notably, co-exposure to folic acid rescued social deficits caused by PFAS. Given the central role of folate in early neurodevelopment, the findings suggest that interference with folate receptors may contribute to ASD-related neurodevelopmental toxicity associated with PFAS exposure.
Dr. Tanguay’s own presentation brought the conversation back to one of toxicology’s oldest and most difficult questions: what is the structural basis for differential chemical activity? In her framing, researchers often begin with almost no prior knowledge. Any expressed gene product, she noted, could in principle be the target of an unknown chemical. The challenge is to “fish out” the relevant one—a fitting metaphor from a zebrafish biologist.
Her lab’s approach combines high-throughput in vivo screening with molecular docking, generating two complementary streams of evidence: empirical effects in living organisms and computational predictions of target binding. Using this framework, her group has identified and corroborated candidate targets such as NT5E for various alkyl naphthalenes. But Tanguay emphasized restraint. Many docking predictions may not be biologically relevant. Computational data alone are not enough; rigorous validation, especially in complex biological systems, is essential. What these methods do provide is power: the ability to prioritize likely targets, better understand gene-by-environment interactions, and explain species differences when model systems do not align. In that sense, structure-based toxicology may help bridge non-human new approach methodologies with questions of human relevance.
The session’s final talk, from Dr. Hao Zhu of Tulane University, tackled a more fundamental barrier: data harmonization. In the post-AlphaFold era, Dr. Zhu argued, the central challenge of computational toxicology is no longer simply generating predictions—it is making sense of messy, incomplete, and often inconsistent public data.
Dr. Zhu pointed to the uneven progress of deep learning, which has won some modeling challenges and failed others, often depending on data quality rather than algorithmic sophistication. In one hepatotoxicity study, his group worked with 869 compounds, 272 assays, more than 11,000 genes, and 2,751 pathways—a scale that underscores both the promise and the chaos of modern toxicological datasets. Docking can help fill in some of the gaps, but only if the underlying chemical and biological data are trustworthy. Public databases, Dr. Zhu warned, are rife with naming errors and inconsistencies. Lack of concordance across species, endpoints, and assay systems can erase useful information before models ever have a chance to learn from it. Manual curation remains indispensable, even as large language models begin to assist with identifying literature evidence to support model predictions.
Together, the speakers painted a vivid picture of a field in transition. AlphaFold has unquestionably changed the landscape. It has expanded access to protein structures on a scale unimaginable just a few years ago and strengthened the case for in silico approaches as practical, non-animal tools in toxicology. But the SOT session also underscored a harder truth: prediction is not understanding. Structures are powerful, but they are static approximations of dynamic biology. Docking is useful, but it is not proof. AI can accelerate discovery, but only when paired with careful experiments, curated data, and a willingness to test what the models get wrong.
Like others, I am looking forward to the 2027 Global Session—Dr. Tanguay encouraged folks to put in session proposals to bring together scientists for next year to talk about globally relevant issues in toxicology.
This blog reports on the Featured Session titled “Structure-Based Toxicology in the AlphaFold Era: Emerging Opportunities and Unmet Challenges” that was held during the 2026 SOT Annual Meeting and ToxExpo. An on-demand recording of this session is available for meeting registrants on the SOT Online Planner and SOT Event App.
This blog was prepared by an SOT Reporter and represents the views of the author. SOT Reporters are SOT members who volunteer to write about sessions and events in which they participate during the SOT Annual Meeting and ToxExpo. SOT does not propose or endorse any position by posting this article. If you are interested in participating in the SOT Reporter program in the future, please email SOT Headquarters.
#SOTReporter