TLDR: This research paper introduces Biology-Informed Machine Learning (BIML) as an evolution of Physics-Informed Machine Learning (PIML) to address the unique complexities of biological systems. It highlights PIML’s shortcomings in biology due to uncertain knowledge, data heterogeneity, unobserved variables, and network complexity. BIML proposes four pillars—uncertainty quantification, contextualization, constrained latent structure inference, and scalability—and leverages Foundation Models and Large Language Models to integrate diverse biological information, aiming for more robust and interpretable models in life sciences.
A new perspective paper titled “Biology is the Challenge Physics-Informed ML Needs to Evolve” by Julien Martinelli from the ELLIS Institute Finland and Aalto University, proposes a significant evolution in how machine learning can be applied to biological systems. The paper introduces Biology-Informed Machine Learning (BIML) as a principled extension of Physics-Informed Machine Learning (PIML), designed to tackle the unique complexities inherent in biological data and knowledge.
Physics-Informed Machine Learning has achieved remarkable success in fields like fluid dynamics and climate science by embedding known physical laws, often expressed as differential equations, directly into machine learning models. This approach works well when systems are governed by clear, well-understood equations and data is structured and relatively noise-free. However, the paper argues that directly porting PIML to biology, a field rich in dynamic systems, faces substantial hurdles.
Why Traditional PIML Falls Short in Biology
The author identifies four key structural mismatches that limit PIML’s effectiveness in biological modeling:
- Multi-faceted and Uncertain Prior Knowledge: Unlike physics, biological understanding is often qualitative, fragmented, and context-dependent. It rarely translates into precise governing equations, making it difficult to encode as strong theoretical guarantees.
- Data Heterogeneity: Biological datasets are typically collected across diverse individuals, conditions, or perturbations, reflecting inherent variability that needs to be modeled explicitly, not just averaged out as noise.
- Unobserved Species: Many crucial biological components, such as intracellular concentrations or regulatory factors, are unmeasured or unmeasurable. PIML often assumes full observability, which is rarely the case in biology.
- Network Size and Complexity: Biological systems involve large, nonlinear feedback networks with numerous interacting variables. PIML methods, often designed for low-dimensional systems with known structures, struggle with this combinatorial complexity.
Furthermore, the paper highlights a critical absence of benchmarks specifically tailored to the complexities of biological dynamical systems, which currently prevents a systematic assessment of PIML’s true effectiveness in this domain.
Introducing Biology-Informed Machine Learning (BIML)
BIML is presented not as a replacement for PIML, but as an adaptation that retools its methods to operate under softer, probabilistic forms of prior knowledge. It embraces the epistemic and practical constraints of biology, integrating diverse information like partially known interactions and context-dependent structures into ML workflows for scientific discovery.
The framework of BIML is built upon four foundational pillars:
- Uncertainty Quantification: Given the ambiguity of biological knowledge, BIML places a strong emphasis on quantifying uncertainty across model structure, parameters, and data. Probabilistic frameworks, especially Bayesian approaches, are crucial here.
- Contextualization: BIML moves beyond one-size-fits-all models by developing frameworks that can disentangle shared biological mechanisms from variations specific to individuals, tissues, or experimental conditions.
- Constrained Latent Structure Inference: Recognizing that biological systems are intrinsically partially observed, BIML adopts strategies that treat partial observability as the norm. It introduces hypothetical yet interpretable latent variables, constrained by biological prior knowledge, to represent unmeasured components.
- Scalability: To handle the high dimensionality of real-world biological systems, BIML requires computationally efficient inference strategies and model architectures that preserve interpretability and fidelity, even with complex, feedback-rich networks.
The Role of Foundation Models and Large Language Models
A key insight of the paper is the integrative role of Foundation Models (FMs) and Large Language Models (LLMs) in enabling BIML. These advanced AI models can bridge human expertise with computational modeling by:
- Formalizing and weighting uncertain prior knowledge.
- Interpreting and disambiguating heterogeneous contexts.
- Proposing biologically plausible latent components.
- Informing scalable model construction by suggesting modular structures and pruning implausible interactions.
The paper envisions LLMs acting as interactive modeling assistants within “virtual laboratories,” guiding experiment design, querying databases, and validating constraints, thereby fostering a more adaptive and collaborative scientific workflow.
Also Read:
- Guiding Causal Discovery with Known Influences: A New Approach to Understanding Relationships
- Enhancing Personalized Treatment Prediction with Scarce Data through Multi-Source Knowledge Integration
A Call to Action for the ML Community
To facilitate the transition to BIML, the paper offers concrete recommendations. It advocates for rethinking benchmarks to create “epistemic stress-tests” that expose failure modes under realistic biological conditions, rather than just optimizing performance on idealized tasks. It also calls for an application-driven view of ML, where models are judged by their relevance to real-world biological problems. Finally, fostering interdisciplinary collaboration between ML researchers and domain experts (biologists, pharmacologists, clinicians) is crucial to ensure models target meaningful dynamics and generate actionable predictions.
The authors argue that biology’s complexity should not be seen as an obstacle, but as a catalyst for ML innovation. By engaging with these intricacies, BIML aims to push PIML towards greater robustness, expressivity, and scientific maturity, ultimately advancing our understanding of living systems. For more details, you can read the full paper here.
