I enjoyed your thought-provoking post! We think a lot about these topics in our work testing human tumor tissue ex vivo and training models to predict response to multiple therapeutics. You touched on several things we’ve also found important in experimental and model design, particularly the value of perturbations and understanding response trajectories over time.
I agree that general-purpose foundation models have limits, since there is no single, task-independent “best” representation of a biological system. We see this empirically: our models generalize better to unseen patients when we train the representation and treatment-response predictor end-to-end.
One of the biggest challenges I see with clinical datasets is that a patient can only receive one therapy at a time, so responses to alternative treatments (at the same initial state) are never observed... This makes it difficult to disentangle patient biology from treatment effects and learn the patient-by-treatment interactions that ultimately matter for treatment selection. Experimentally perturbing the same patient’s tumor with multiple therapies gives us a very different kind of training signal.
If compensatory capacity is latent rather than directly observable, how would you operationalize its measurement in practice? Would it require repeated perturbations to identify the hidden state, rather than a single response–recovery trajectory?
For my part, I think the biological world model already exists. It is the living organism itself. Its possible “computational power” does not sit in an abstract representation, but in its entire material architecture: in boundaries, flows, feedbacks, and in the capacity to respond to perturbations that have never occurred before. Perhaps the real challenge is not to build a biological world model, but to understand how the one that already exists computes.
I enjoyed your thought-provoking post! We think a lot about these topics in our work testing human tumor tissue ex vivo and training models to predict response to multiple therapeutics. You touched on several things we’ve also found important in experimental and model design, particularly the value of perturbations and understanding response trajectories over time.
I agree that general-purpose foundation models have limits, since there is no single, task-independent “best” representation of a biological system. We see this empirically: our models generalize better to unseen patients when we train the representation and treatment-response predictor end-to-end.
One of the biggest challenges I see with clinical datasets is that a patient can only receive one therapy at a time, so responses to alternative treatments (at the same initial state) are never observed... This makes it difficult to disentangle patient biology from treatment effects and learn the patient-by-treatment interactions that ultimately matter for treatment selection. Experimentally perturbing the same patient’s tumor with multiple therapies gives us a very different kind of training signal.
If compensatory capacity is latent rather than directly observable, how would you operationalize its measurement in practice? Would it require repeated perturbations to identify the hidden state, rather than a single response–recovery trajectory?
For my part, I think the biological world model already exists. It is the living organism itself. Its possible “computational power” does not sit in an abstract representation, but in its entire material architecture: in boundaries, flows, feedbacks, and in the capacity to respond to perturbations that have never occurred before. Perhaps the real challenge is not to build a biological world model, but to understand how the one that already exists computes.