Molecular structure and interaction
Predict shapes, binding, conformational change, and interactions among proteins, nucleic acids, and small molecules.
Layer 7 — Scientific AI · Chapter 3
Biology is a layered information system extending from molecules and cells to organs, organisms, populations, and ecosystems. Scientific AI can model each level, but the greatest difficulty—and opportunity—lies in connecting them.
A molecular change may alter a protein, affect a cellular pathway, produce a tissue-level response, and influence a patient differently depending on age, environment, genetics, and other treatments. Useful biomedical AI must therefore combine data across scales without pretending that one level fully determines another.
Predict shapes, binding, conformational change, and interactions among proteins, nucleic acids, and small molecules.
Relate sequence variation and regulatory context to expression, function, and disease risk.
Model how cells respond to perturbations, environments, drugs, and developmental signals.
Extract biological patterns from microscopy, radiology, and tissue images.
Estimate disease progression, treatment response, adverse events, and patient subgroups.
Study epidemiology, evolution, public health, and interactions between genes and environment.
AI can improve each stage, but no model eliminates the need for biological evidence. A candidate may look excellent computationally and still fail because the disease mechanism is wrong, the molecule behaves differently in living systems, the therapy cannot reach its target, or the clinical effect is too small.
Biomedical datasets often describe what happened to patients rather than what would have happened under a different treatment. This makes causal reasoning essential. A model may learn that a treatment is associated with worse outcomes because the sickest patients receive it—not because the treatment causes harm.
Randomized experiments, natural experiments, mechanistic knowledge, and careful causal design remain necessary. AI can help identify hypotheses and subgroups, but retrospective prediction should not be confused with proof of therapeutic benefit.
Many diseases are labels for multiple underlying mechanisms. Patients with similar symptoms may respond differently because their biology differs. Scientific AI can search for mechanistic subtypes and treatment-response patterns, supporting a move from broad categories toward more precise intervention strategies.
This creates a tension: finer personalization may improve treatment selection, but smaller subgroups provide less data. Models need uncertainty estimates and must avoid presenting weak subgroup evidence as certainty.
AI can assist with eligibility criteria, site selection, recruitment, protocol feasibility, endpoint modeling, safety monitoring, and identification of patients likely to benefit. It can also help design adaptive experiments that update allocation as evidence accumulates.
However, optimization against a surrogate endpoint can produce a trial that succeeds numerically without delivering meaningful patient benefit. Scientific and regulatory judgment must determine whether an endpoint reflects survival, function, symptoms, quality of life, or another outcome that matters.
| Opportunity | Potential value | Central risk |
|---|---|---|
| Target and mechanism discovery | Improve selection of biological hypotheses before expensive development. | Association may not translate into an effective intervention. |
| Molecular and protein design | Explore candidates beyond conventional libraries. | Generated candidates may be unstable, unsafe, or difficult to manufacture. |
| Experimental optimization | Reduce laboratory iterations and improve information gained per test. | Automation may optimize the wrong measurement. |
| Patient stratification | Match therapies to biological subgroups. | Small or biased subgroups may not generalize. |
| Trial operations and simulation | Improve feasibility, recruitment, monitoring, and decision-making. | Models may not capture real-world adherence or clinical complexity. |
| Biomarker platforms | Create measurable indicators of disease or treatment response. | A predictive marker may lack causal or clinical significance. |
The most important biomedical model is not necessarily the one that predicts best on an existing dataset. It is the one that helps choose an intervention that works in living patients.