# Curated Digest: Extracting Biological Knowledge via Mechanistic Interpretability > Coverage of lessw-blog **Published:** March 14, 2026 **Author:** PSEEDR Editorial **Category:** platforms **Tags:** Mechanistic Interpretability, Computational Biology, Foundation Models, Genomics, Single-Cell Data **Canonical URL:** https://pseedr.com/platforms/curated-digest-extracting-biological-knowledge-via-mechanistic-interpretability --- lessw-blog explores how mechanistic interpretability can extract complex, emergent biological structures-like the evolutionary tree of life-hidden within the internal activations of foundation models. In a recent post, lessw-blog discusses the fascinating intersection of artificial intelligence and computational biology, specifically focusing on how mechanistic interpretability can be used to extract structured biological knowledge from foundation models. As biological foundation models grow in scale and capability, they are increasingly treated as black boxes. While these models excel at tasks like predicting DNA sequences or gene expression profiles, the internal mechanisms driving these predictions remain largely opaque. This opacity presents a significant missed opportunity for the scientific community. If an artificial neural network can accurately predict complex biological phenomena, it has likely learned fundamental, underlying rules about biology to achieve that performance. Mechanistic interpretability provides the necessary framework to reverse-engineer these opaque systems. By analyzing the weights and activations of these models, researchers can translate high-dimensional, mathematical representations into human-readable scientific insights, turning predictive engines into discovery engines. The lessw-blog post highlights a compelling case study involving Evo 2, a state-of-the-art genomic foundation model. Researchers discovered that Evo 2 implicitly learned the evolutionary tree of life simply by being trained on the fundamental task of predicting DNA sequences. Remarkably, the model received no explicit training signals regarding evolutionary biology or taxonomy. Instead, the necessity of predicting DNA accurately forced the model to encode phylogenetic relationships as a curved manifold within its internal activations. Because this learned representation is highly structured and geometric, researchers were able to mathematically extract it and meaningfully compare it to established, ground-truth evolutionary trees, proving that the model had independently deduced core evolutionary principles. Building on this foundational success, the author outlines an ambitious plan to apply similar mechanistic interpretability techniques to single-cell data foundation models, such as scGPT. By examining the internal representations of models trained on complex gene expression profiles, the author's goal is to uncover hidden, dynamic pathways of human cell development. If successful, this could map out previously unknown cellular trajectories and differentiation processes. This analysis demonstrates the profound emergent capabilities of biological foundation models to spontaneously organize raw data into structured, scientifically valid knowledge. By making these models transparent, researchers can accelerate biological discovery and validate the internal logic of AI systems. For a deeper examination of the geometric structures found in Evo 2 and the future of single-cell research, [read the full post on lessw-blog](https://www.lesswrong.com/posts/R4xxxAfNpAvpb3LCf/extracting-performant-algorithms-using-mechanistic-5). ### Key Takeaways * Evo 2, a genomic foundation model, implicitly learned the evolutionary tree of life simply by predicting DNA sequences. * The model encoded phylogenetic relationships as a structured, curved manifold within its internal activations. * Mechanistic interpretability allows researchers to extract these geometric representations and compare them to ground-truth biological data. * The author plans to extend this methodology to single-cell foundation models like scGPT to map human cell development pathways. * This approach highlights the potential of AI models to act as engines for novel scientific discovery rather than just predictive tools. [Read the original post at lessw-blog](https://www.lesswrong.com/posts/R4xxxAfNpAvpb3LCf/extracting-performant-algorithms-using-mechanistic-5) --- ## Sources - https://www.lesswrong.com/posts/R4xxxAfNpAvpb3LCf/extracting-performant-algorithms-using-mechanistic-5