The allegory of strange loops: A perspective on AI guided cancer research
By Fayyaz Minhas (With AI editorial support)
I often begin my talks about my research in understanding cancer pathobiology with Plato's allegory of the cave. Prisoners sit chained to a wall, watching shadows cast by unseen objects behind them. They debate the nature of reality by analysing those shadows, unaware of the processes that generate them. In many ways, we as scientists find ourselves in a similar position. We observe measurements, images, molecular readouts, and clinical outcomes. These are shadows cast by deeper biological processes. Our task is not merely to describe the shadows, but to infer the mechanisms that generate them, so that we can intervene meaningfully in the real world.
Modern cancer research is often framed as a choice between levels of description. At one extreme lie molecular profiles such as mutations, gene expression, and epigenetic marks, treated as privileged causes of disease. At the other lie tissue images such as whole slide pathology and spatial maps, treated as rich, high dimensional summaries of downstream phenotypic effects. Increasingly, the proposed solution is multimodality: integrate everything using artificial intelligence, and the truth will emerge.
But multimodality is a tool, not a philosophy. One can be multimodal and still profoundly reductionist.
The deeper problem is not missing data, but how we, particularly as computer scientists, think about causation, explanation, and understanding in biology. Cancer and many other complex pathologies are not poorly understood because we lack modalities. They are poorly understood because we continue to force self organising, non linear systems into explanatory frames designed for linear causation.
Genes explain evolution, not construction
The phrase “it's all in the genes” is deeply ingrained in biological thinking. Richard Dawkins' The Selfish Gene is often taken as support for this idea, but this is a misunderstanding. Genetic determinism is the belief that knowing the genes is sufficient to predict biological outcomes, that DNA functions as a blueprint whose execution yields inevitable form and behaviour. Dawkins does not make this claim. His argument is precise and limited. Genes are replicators, units of evolutionary persistence. They explain why certain traits spread through populations over evolutionary time, much like memes spread culturally. They do not explain how organisms are constructed, how tissues organise, or how disease unfolds dynamically within an individual.
Philip Ball's How Life Works makes this distinction explicit. Genes are not programs or blueprints. They do not encode form directly. Biological order arises from interactions among genes, physical constraints, cellular dynamics, and environmental context. Development, robustness, and pathology are processes rather than executions of code. Form is constructed through interaction, not read out from DNA.
Yet in cancer biology, gene centred thinking quietly hardened into genetic determinism. One gene, one solution narratives flourished, sometimes implicitly and sometimes explicitly. These narratives achieved spectacular success in particular contexts, such as subsets of acute myeloid leukaemia, where single mutations can explain disease mechanisms and enable targeted therapy.
These successes were real, but they were also misleading. They encouraged the belief that cancer in general is a molecular wiring problem waiting to be debugged, perhaps one day to be edited away. Most diseases are not like this. They are not deterministic, linear, or reducible to isolated molecular causes.
Life, order, and the strange loop
My thinking here owes much to Gödel, Escher, Bach and I Am a Strange Loop by Douglas Hofstadter. I did not fully understand these books when I first read them, and I may never fully understand them. This is not a weakness of the books. As Nassim Nicholas Taleb has observed, if a book can be fully understood by a summary, it likely did not need to be written. Books that continue to yield new interpretations upon rereading are often those grappling with genuinely deep ideas.
Hofstadter introduced a way of thinking that has proven indispensable. Complex systems generate meaning and behaviour through circular causation. In a strange loop, higher level patterns emerge from lower level interactions and then act back upon them. The system becomes self referential. Escher's drawings, where hands draw one another or staircases loop back onto themselves, are visual demonstrations of this principle.
This is Hofstadter's account of consciousness, but it is also a powerful account of biology. Cancer, viewed through this lens, is not a simple accumulation of mutations. It is an emergent reorganisation of a tissue level system. It adapts, persists, and exhibits apparent goal directedness. When we say that a tumour invades or that cancer metastasises, we are not invoking intention, but we are not speaking nonsense either. We are describing emergent teleology, behaviour that appears purposeful because it arises from feedback driven organisation.
Order against entropy and the origin of emergence
Where does this apparent purpose come from in life and disease. William James Sidis, in The Animate and the Inanimate, approached this question from a thermodynamic perspective. Sidis suggested that life locally reverses the second law of thermodynamics. This is not correct as literal physics. Living systems do not violate thermodynamic laws. However, Sidis captured something enduring and important. Living systems are non equilibrium systems that maintain and propagate structure against dissipation through continuous flows of energy and information.
This sustained organisation creates feedback loops. These loops stabilise patterns, amplify certain behaviours, and suppress others. Over time, such loops give rise to structures that persist, adapt, and interact. Even pathological life does this. Cancer is not random noise. It is structured, patterned, and persistent, almost like one strange loop interacting destructively with another.
Teleology in biology does not arise from intention but from stability. Patterns that persist appear purposeful because they have survived the dynamics that would otherwise dissolve them.
Cancer emerges at the tissue level
Complex systems often exhibit properties that are more than the sum of their parts. Simple local rules can give rise to complex global behaviour, a principle explored extensively by Stephen Wolfram in A New Kind of Science. Cellular automata governed by trivial rules can suddenly cross thresholds into rich, unpredictable dynamics.
Biology appears to operate in exactly this way. Below certain organisational thresholds, behaviour remains simple and local. Above them, new levels of causation emerge.
This view is formalised in the emergent theory of cancer articulated by Sigston and Williams, who reject the sufficiency of the somatic mutation theory. They argue that cancer is an emergent system arising at the level of the functional tissue unit, not at the level of genes or individual cells.
In this framework, tissue architecture is not a downstream consequence of molecular change but a primary explanatory level. Causation is multi scale and bidirectional. Genes influence tissue organisation, but tissue organisation constrains gene expression, cellular behaviour, and evolutionary trajectories.
This relationship can be formalised statistically and information theoretically. Imaging patterns reduce uncertainty about mutations or cellular states and vice versa, revealing measurable mutual information between morphology and molecular profiles. This statistical dependence enables prediction. Certain architectural patterns correlate with aggressive disease, particular mutations, or coordinated expression programs.
But statistical dependence is not causation, nor does it explain the latent mechanisms that generate both. Diagnosing cancer from genes alone is conceptually incomplete because it ignores the constraints imposed by tissue organisation. Explaining cancer purely from morphology risks mistaking stabilised patterns for primary causes. Both errors arise from collapsing a circular system into a one way explanation.
Images are not explanations either
When I moved from bioinformatics into computational pathology, I initially thought of whole slide images as another downstream layer of molecular events. Over time, this assumption eroded. Computational pathology is too often reduced to machine learning on slides rather than understood as a way to study disease processes themselves using computation, particularly the latent causes that give rise to observed patterns.
There has been substantial work on predicting gene expression or mutations from morphology. These efforts are useful, but the predictions are never perfect. The standard explanations invoke noise, limited datasets, or weak labels. These explanations are convenient but insufficient.
There is growing evidence that mechanical forces and geometry influence gene regulation. Studies have shown that nuclear deformation can alter chromatin organisation and transcriptional programs in colorectal cancer and other systems. Shape affects transcription. If mechanical perturbation of nuclei can regulate genes, then tissue architecture cannot be treated as a passive readout. Form acts back on function, reminiscent of Escher's hands drawing one another.
This is the strange loop again. Genes shape tissue. Tissue reshapes gene expression. The system stabilises around emergent phenotypes that we observe as tumour evolution. The difficulty of predicting molecular states from images is not merely noise or dataset limitation. It reflects the fact that causal prediction is fundamentally harder than anticausal prediction, especially in systems with circular causation, as emphasised by Bernhard Schölkopf and others.
Multimodality is not the point
At this stage, it is tempting to declare multimodal integration as the solution. A more holistic picture is necessary, but multimodality can easily become a fetish. One can integrate images, transcriptomics, and proteomics while still treating one level as ground truth and the others as noisy reflections. This reproduces reductionism under the guise of completeness.
The goal is not to stack modalities. The goal is to change the explanatory frame.
Causality, not comfort
This is where Judea Pearl's The Book of Why becomes essential. Pearl's central argument is that prediction is not explanation. Correlation cannot answer what if questions. Understanding requires causal models that encode interventions, counterfactuals, and feedback.
AI in biology is often asked to be explanatory in a psychological sense, to reassure us, to pacify us, to tell stories that feel intuitive or controllable. But explanatory AI can become an anaesthetic. It can obscure rather than illuminate when it reinforces simplistic narratives.
The task is not to explain cancer in comforting terms, but to reason about it correctly, because lives depend on it.
Rethinking experimentation and evidence
This shift has implications for how we design experiments and therapies. Much of modern biology still revolves around isolating individual cells, increasingly through organoids, perturbing them with drugs, and extrapolating upward. This approach has value, but it remains limited.
As Professor David Epstein once observed, death is a non subjective assessment. Outcomes matter. Retrospective patient data, longitudinal disease pathways, tissue organisation, treatment histories, and clinical endpoints encode causal information that no isolated assay can capture alone.
Understanding cancer and designing therapies should not proceed from isolated mechanisms in isolation. It should integrate retrospective patient trajectories, tissue level organisation, clinical trial outcomes, and molecular interventions, not as parallel silos but as components of a single causal system.
Toward a new view
Images alone are not enough. Mutations alone are not enough. Multimodality alone is not enough.
Cancer is neither a genetic bug nor a visual pattern. It is a failure mode of a complex, self organising system. The actionable path forward is clear. We must adopt causal, multi scale models as the foundation of AI in cancer biology. We must treat genes, cells, tissues, patients, and treatment pathways as interacting levels rather than hierarchical outputs. We must use AI not merely to predict or explain, but to test interventions and counterfactuals across scales. We must anchor discovery in real patient trajectories rather than isolated experimental abstractions.
AI should not be a pattern recogniser or a storyteller. It should be a tool for navigating strange loops.
