From Bench to Bedside: How AI Is Reshaping Medicine
AI is already being used across life sciences for specific, well-defined jobs, from designing candidate molecules and analysing trial data to interpreting pathology slides and identifying patients who may need further investigation. Following the pipeline from discovery towards clinical care, this feature looks at what organisations listed on Benchscope are building, where the technology may be useful and what still has to be proved.
For most people, AI at work still means fairly ordinary things: drafting emails, summarising documents, searching for information or automating repetitive tasks. In life sciences, some of the most important systems look quite different, built for a single scientific task and trained on specialised data.
AlphaFold is arguably the best-known example. Developed by London-based Google DeepMind, it predicts the three-dimensional structure of proteins from their amino acid sequences, a problem that had challenged researchers for decades. In the years since, AI has spread into other specialised areas of life sciences, from designing candidate medicines to analysing clinical and biological data.
In this feature, we follow that shift along the pipeline, from discovery through to clinical care, using examples from UK companies listed on Benchscope.
Discovery: narrowing the search
At the start of the pipeline, AI is being used to reduce the number of possibilities researchers need to investigate. Drug discovery still depends on laboratory work, but computational models can help decide where that work might be most useful.
London-based Isomorphic Labs is developing an AI-first drug design system intended to predict molecular interactions and support the design of potential medicines. Its work spans several therapeutic formats including small molecules, antibodies and peptides.
Etcembly takes a more specialised approach. The Oxfordshire company uses generative AI to design T cell receptors and antibody therapeutics, applying language-model techniques to biological sequence data. Its lead programme targets PRAME, a protein associated with several cancers.
The attraction is straightforward. If a model can push researchers towards better candidates earlier, fewer resources may be spent testing poor ones. That does not make the rest of drug development optional. Chemistry, biology, toxicology, manufacturing and clinical trials still determine whether a candidate can become a medicine.
Preclinical research: testing ideas against biology
A promising prediction is only useful if it survives contact with real biology. Further along the pipeline, AI is being combined with laboratory models, imaging and multi-omic data to help researchers decide which ideas are worth taking forward.
Oxford-based Scripta Therapeutics uses AI alongside imaging and patient-derived models in a lab-in-the-loop workflow for neurodegenerative drug discovery. Rather than treating the model as the final answer, experimental results can be used to refine the next round of predictions.
Sonrai Analytics in Belfast works across a broader set of research data. Its platform brings together multi-omic, imaging and clinical information for work including target identification, biomarker discovery and patient stratification. Its customers span preclinical research through clinical trials.
This stage is a useful reminder that AI in life sciences often works best as part of a feedback loop. A model helps select what to test, the experiment produces new data and that data can improve the next decision.
Clinical development: learning what works in people
Once a treatment reaches clinical development, the questions change. Researchers need to understand which patients should be studied, how disease is progressing and whether a treatment is producing a meaningful response.
Queen Square Analytics, a UCL spin-out, applies machine learning to MRI and clinical trial data in neurological disease. Its work includes patient stratification, treatment response prediction and imaging analysis for trials in multiple sclerosis and other neurodegenerative conditions.
Here the AI is not designing the medicine. It is helping researchers make sense of the evidence around it. That may improve how trials are designed or how changes in disease are measured, but the usual standards of clinical evidence still apply.
Diagnosis and decision support: closer to clinical care
The closer AI moves towards routine care, the more important validation becomes. Systems that influence a diagnostic pathway need to work reliably across the people and settings where they are actually used.
Panakeia is developing software that analyses routine H&E-stained tissue images and predicts molecular biomarkers from them. Its breast cancer product reports ER, PR and HER2 status from standard pathology slides and is available as an IVD medical device in Great Britain and the EU.
Mendelian approaches a different problem. Its MendelScan software analyses electronic health record data to identify patients who may warrant further investigation for a rare disease. The system is designed to flag and prioritise cases for clinical review rather than make a diagnosis itself.
Vivid Dx is working on speed. The Oxford spin-out is developing a platform that combines Raman spectroscopy, microfluidics and deep learning to identify bacteria directly from blood samples. The company is targeting pathogen identification in around 30 minutes and phenotypic antibiotic susceptibility results in about three hours.
These examples are quite different, but the role of AI is similar. It sits between a large or complex source of information and a decision that still needs scientific or clinical judgement.
Towards the patient
Some applications move closer still to the treatment itself. Cambridge-based BIOS Health develops neural interfaces and software that analyse signals from the peripheral nervous system. Its platform uses machine learning to model neural and physiological responses with the aim of helping personalise neuromodulation therapies.
This is a more experimental area than analysing an established pathology image. Neural signals are complex, treatment settings are highly individual and the route from a promising model to broad clinical use is demanding.
It also shows how wide the phrase 'AI in healthcare' has become. The same broad family of methods can be used to rank molecules at the start of drug discovery, analyse trial data years later or interpret signals coming directly from the body.
The movement is going both ways
Life sciences companies are pulling AI deeper into research and clinical workflows. At the same time, frontier AI companies are building products specifically for scientists.
Anthropic launched Claude Science in June 2026 as an AI workbench for scientific research, with tools intended to sit alongside databases, coding environments and other research software. In September, it introduced a Life Sciences Verification Program that gives verified teams access to models with safeguards adapted for legitimate work in areas including drug discovery, research biology, clinical development and manufacturing.
That does not mean Anthropic is becoming a biotech company. The contrast is more useful than that. Life science organisations are adopting AI to deal with biological complexity, while AI companies are adapting their products to the practical needs of scientists. The two industries are beginning to meet in the middle.
What comes next?
The next stage for medical AI is likely to be less about finding another eye-catching demonstration and more about proving which tools deserve a place in routine research and healthcare.
Many systems already show promise in controlled studies or specialist settings. The harder test is whether they remain useful across different laboratories, hospitals and patient groups. That means validation, regulation and integration matter just as much as model performance.
If the strongest tools are adopted, patients may rarely notice the AI itself. It could sit inside pathology, imaging, drug discovery or clinical research as one part of a much larger process.
The useful question is therefore not whether AI will take over healthcare. It is which applications can show a measurable benefit, work reliably in the real world and become ordinary parts of how medicines are researched, tested and delivered.
Explore more AI & Machine Learning companies working across drug discovery, diagnostics, clinical research and healthcare on Benchscope. If your company is listed and would like to update its profile, or if you think we’ve missed a relevant organisation, submit an update here.