Layer 7 — Scientific AI · Chapter 3

Biology, Medicine, and Drug Discovery

Book chapterLayer 7

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.

Major biological modeling tasks

Molecular structure and interaction

Predict shapes, binding, conformational change, and interactions among proteins, nucleic acids, and small molecules.

Genomics and gene regulation

Relate sequence variation and regulatory context to expression, function, and disease risk.

Cellular state

Model how cells respond to perturbations, environments, drugs, and developmental signals.

Pathology and imaging

Extract biological patterns from microscopy, radiology, and tissue images.

Clinical trajectories

Estimate disease progression, treatment response, adverse events, and patient subgroups.

Population biology

Study epidemiology, evolution, public health, and interactions between genes and environment.

The drug-discovery pipeline

  1. Understand disease biology: identify mechanisms, pathways, and targets that may influence a condition.
  2. Find or design candidates: screen existing molecules or generate new compounds, proteins, antibodies, or other interventions.
  3. Optimize: balance potency, selectivity, stability, delivery, manufacturability, and intellectual-property constraints.
  4. Evaluate safety: estimate toxicity, off-target effects, metabolism, interactions, and risk across biological systems.
  5. Validate experimentally: test in biochemical assays, cells, models, and eventually human studies.
  6. Design clinical development: select patients, endpoints, doses, combinations, and trial structures.

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.

From correlation to intervention

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.

Patient heterogeneity

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.

Clinical research and trial design

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.

Biomedical data constraints

Business and research opportunities

OpportunityPotential valueCentral risk
Target and mechanism discoveryImprove selection of biological hypotheses before expensive development.Association may not translate into an effective intervention.
Molecular and protein designExplore candidates beyond conventional libraries.Generated candidates may be unstable, unsafe, or difficult to manufacture.
Experimental optimizationReduce laboratory iterations and improve information gained per test.Automation may optimize the wrong measurement.
Patient stratificationMatch therapies to biological subgroups.Small or biased subgroups may not generalize.
Trial operations and simulationImprove feasibility, recruitment, monitoring, and decision-making.Models may not capture real-world adherence or clinical complexity.
Biomarker platformsCreate 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.