AI in Biomedicine: From Discovering Drugs to Predicting Biology

Artificial intelligence is transforming biomedicine, but its greatest opportunity lies not in accelerating drug discovery alone — rather in enabling predictive biology: the ability to predict how patients, therapies, and living systems will behave in real-world settings.

While advances in AI have expanded our capacity to identify new targets and therapeutic candidates, the fundamental bottlenecks remain clinical success, patient stratification, and the consistent performance of advanced biologics. Overcoming these challenges requires a new generation of functional biological data that captures cellular behavior over time rather than static molecular snapshots.

Combined with modern machine learning, these datasets can support more accurate diagnostics, improve clinical development, and enable personalized medicine at scale. At TeraCyte, we are building the infrastructure for this transition through high-throughput, image-rich, time-resolved single-cell assays designed to generate the functional data layer needed for predictive biology.

For AI to make a substantial impact on the declining return on biomedical innovation, it must move beyond making drug discovery faster and streamlining existing processes. The real opportunity is to remove the bottlenecks that determine whether medicine actually reaches patients: de-risking clinical trials, matching therapies to the right individuals, and manufacturing advanced biologics at scale. This is the strategic problem TeraCyte is built around.

AI-driven drug discovery can be valuable and commercially attractive, but discovery is not the fundamental constraint in the biomedical ecosystem. We already have more targets, modalities, and candidate therapeutics than we can clinically validate, manufacture, and deliver.

The core bottleneck is prediction: which therapy will work in which patient, when to intervene, how a patient's immune system is responding, and whether a living therapeutic or bioprocess will perform consistently at scale.

I use the term predictive biology deliberately. It is not meant as a new discipline detached from systems biology or dynamical modeling. It is a practical framing for the next wave of biomedical AI: models trained on functional biological behavior and built to predict clinical or industrial outcomes.

To enable that shift, we need two missing elements: scalable biological data and simple inference at the point of care.

The data layer is critical. Much of today's biomedical data is static, fragmented, expensive, or too indirect to support high-confidence decisions. Biology, however, is dynamic and phenotypic in nature. In an immune assay, relevant behavior can include activation, migration, killing, proliferation, cell-cell interaction, and short-term response trajectories. These are observable functions expressed in morphology, motion, interaction patterns, fluorescence dynamics, and survival over time.

This is why live cell imaging and phenotypic data are so important for AI. Genomics and proteomics are powerful, but they often measure potential or state. Imaging measures behavior. It creates dense, high-dimensional data in which every cell can become a training example for representation learning — morphology, movement, timing, intensity changes, and interaction history can all be learned directly by computer-vision models without requiring humans to define every feature in advance.

Vision transformers, self-supervised pretraining, and multimodal architectures are well suited to this problem precisely because they extract structure from large image streams at scale. The goal is not simply to digitize biology. It is to build a scalable functional data layer that lets models learn how living systems behave.

The second requirement is simple inference at the point of care. A model that depends on specialized infrastructure, long turnaround times, or complex expert interpretation will not integrate well into clinical decision pipelines.

The deployment architecture is still being shaped by regulatory, privacy, and clinical workflow requirements — the right balance between local data generation, edge processing, and validated model inference will vary by setting and intended use. What is not variable is the user experience requirement: load the sample, run the assay, receive a decision-support output. The workflow should feel closer to a clinical test than a research experiment, and it should fit within existing laboratory and hospital infrastructure rather than requiring new ones.

This is why diagnostics must be re-imagined. Personalized medicine cannot be unlocked by treatments alone. It requires personalized diagnostics: simple, fast, reliable assessments of a specific patient in their current biological state. Most companion diagnostics still rely on static biomarkers or thresholded readouts that approximate function. Those tests are essential, and some already include functional or semi-functional signals, but the next step is diagnostics that measure function directly: how a patient's cells behave, and whether that behavior predicts response or risk.

AI diagnostics can also solve a major business model problem. Traditional companion diagnostics are often built one drug, one biomarker, and one indication at a time. That creates fragmented markets, slow adoption, and limited commercial upside. An AI-native diagnostic platform changes the logic. The same physical platform can generate standardized data input across many diseases, while different validated models interpret that data for different indications, therapies, or clinical questions. In other words: same assay infrastructure, different AI models for multiple indications.

This creates platform economics for diagnostics. A single scalable system could serve as a companion diagnostic across multiple indications, support trial enrichment for different therapies, monitor response over time, and improve as datasets grow. Each model and intended use would still require appropriate validation, clinical grounding, and regulatory work. But the underlying workflow, data structure, deployment channel, and installed base can be shared. That is a fundamentally different model from traditional diagnostics.

At TeraCyte, we are building this through immunology on a chip: high-throughput, image-rich, time-resolved single-cell assays that combine silicon-chip precision, live-cell imaging, and AI-powered analysis. Instead of reducing biology to a static endpoint, we capture cellular phenotype and behavior over time across very large cell populations, then transform those dynamics into standardized datasets for predictive models.

The training loop must be grounded in controlled assay conditions, reproducible data generation, clinically meaningful labels, and validation against patient outcomes, therapeutic response, potency, or process performance.

The same platform logic applies across the ecosystem's core bottlenecks. In clinical development, it can help de-risk trials by identifying functional responder signatures and enriching patient cohorts earlier. In personalized medicine, it can enable patient-specific immune diagnostics for therapy selection and monitoring. In advanced biologics, it can support potency assays, CAR-T and cell-therapy QC, donor or batch comparability, and bioprocess monitoring — all areas where static measurements often fail to capture functional performance.

Others are building important pieces of this future: better models, better sensors, better assays, and better clinical workflows. But integration alone is not the durable advantage. The compounding advantage comes from the data flywheel: each clinical deployment generates labeled outcomes that improve the model, which improves prediction, which drives broader adoption, which generates more data.

A platform grounded in standardized assay conditions and validated against real patient outcomes builds a dataset that becomes harder to replicate over time — not only because the hardware is proprietary, but because the biological ground truth is accumulated, not downloaded.