AI and a Future Without Disease

In September, the first patients began receiving rentosertib in a Phase III clinical trial. It is a drug candidate for idiopathic pulmonary fibrosis whose discovery and design were supported by generative AI. In this disease, lung tissue progressively scars, making breathing increasingly difficult. AI helped identify a protein involved in the disease process and design a molecule intended to inhibit its activity. The trial is expected to enrol 320 people and assess the treatment’s effectiveness and safety over a year. (Insilico Medicine: first patients treated and the Phase III trial plan).

What exactly did AI do here? It combined the analysis of biological data with knowledge from scientific publications and helped select TNIK, an enzyme involved in signalling within cells, as a promising treatment target. A generative model then used its three-dimensional structure to propose molecules that could inhibit its activity. Chemists synthesised selected compounds, tested them and refined their properties. It took approximately 18 months to go from identifying the target to selecting a candidate for further preclinical development. (Nature Biotechnology: selecting TNIK, designing the molecule and conducting preclinical studies).

An earlier Phase II trial involving 71 people showed a signal of improved measured lung capacity in one of the treatment groups. However, the trial lasted only 12 weeks and focused primarily on safety. Rentosertib is not yet an approved medicine. The larger study will investigate whether the benefit persists and what risks the treatment carries. To me, reaching this stage is news worth paying attention to. (Nature Medicine: Phase II results, lung function measurements and adverse events).

September also brought the release of AlphaGenome Atlas: a database of predicted effects for approximately nine billion individual changes in human DNA. When investigating the cause of a rare disease, simply reading a patient’s DNA is not enough. Researchers still need to determine which differences matter. The Atlas helps identify changes worth investigating. These are predictions for possible variants, not nine billion experimentally confirmed results. (Google DeepMind: the scope of the Atlas and how its predictions support genetic research).

The model was trained on experimental results measuring gene activity in different cells and tissues. Among other things, it predicts how much RNA a particular DNA sequence will produce and how the cell will join RNA segments before making a protein. By comparing predictions for the original and altered sequences, it estimates the possible effects of changing a single “letter”. The Atlas contains these calculations in advance, so other teams do not have to repeat them. (How AlphaGenome works: training data and comparisons between original and altered DNA sequences).

There is already a concrete example. The tool helped identify an overlooked variant affecting DNM1, a gene associated with a severe neurological condition involving epilepsy. The model predicted abnormal RNA splicing and the production of an abnormally extended protein, and experiments confirmed this mechanism. The Atlas remains a research tool, not a diagnostic system. But a better answer to “what is going wrong here?” helps researchers determine where to look for a treatment. (The DNM1 example: identifying an overlooked variant and experimentally testing its effects).

In another experiment, approximately 950 Claude agents spent 21 hours searching genetic data. These were programs using an AI model and analytical tools, working in parallel. They gathered data on more than 200,000 enzymes that copy RNA into DNA, selected 3,500 potentially interesting biological systems and narrowed the list to 20 candidates for closer examination.

One agent noticed repeated DNA sequences next to the gene for an already known enzyme. The novelty was the entire system: the enzyme, an additional protein and the characteristic repeats. Scientists began investigating it in the laboratory. We do not yet know what it does or whether it will have a medical application. What interests me is how it was found: analysing many possibilities in parallel and selecting those worth testing experimentally. (Anthropic: the agents’ research process, the enzyme system they identified and the questions that remain unanswered).

A similar approach is being used to search for new antibiotics. In a study published in 2025, generative AI proposed more than 36 million chemical structures. After analysis and testing of selected compounds, two showed efficacy in mouse models of infection: one against drug-resistant bacteria that cause gonorrhoea, the other against MRSA, a drug-resistant form of Staphylococcus aureus. This is still preclinical work, but it addresses a specific need: treating infections against which existing antibiotics are losing effectiveness. (MIT: designing new antibacterial compounds and testing them in the laboratory and in mice).

Progress also extends to checking results. In 11 days, AI expressed an existing proof of Fermat’s Last Theorem in approximately 13 million lines of Lean code, allowing a computer to verify each step of the reasoning. The theorem had already been proved; the achievement concerns its verification and builds on many years of work by mathematicians. Alongside the ability to generate answers, we are developing ways to establish whether they are correct. (Anthropic: formalising the proof in Lean, computer verification and the mathematical work it builds on).

In biology, that requires experiments and clinical trials. This is why I would like to see more investment in laboratories and in making well-documented data available. If AI generates more hypotheses, we need to become better at selecting those worth testing and publishing the results of unsuccessful experiments too.

In late September, I had the pleasure of attending Moonshots LIVE in Los Angeles, where people talk only about such a bold vision. An important part of the event is showing these possibilities: how technology can help solve specific, often enormous problems. How to identify the cause of a disease sooner. How to test more ideas for treatments. How to make results available to other teams. These are ambitious topics, far more interesting than speculation about an inevitable catastrophe.

AI is not limited to major scientific discoveries, though. Similarly advanced technologies can be used in businesses, including here in Poland. Models connected to company data and analytical tools can support performance analysis and production optimisation: finding relationships between process parameters and product quality, helping investigate the causes of waste, or comparing different ways of using machinery, materials and energy. These are far more interesting tasks than producing another automated response to a routine support ticket.

I see the greatest potential in assigning AI increasingly sophisticated analyses—work that previously required substantial time, or that a company never attempted at all. But this requires expertise. You need to understand the production process, choose appropriate analytical methods, connect data from different systems and check whether a proposed change actually improves the outcome. Access to a good model alone will not do that. It takes this preparation to use AI for better decisions and business development.

If AI in your company is still mainly handling routine support tickets, it is worth considering whether you could give it something more ambitious to work on. And if you would like to know how to approach that, come and talk to us at TailoredByte.

When I look at headlines and discussions on social media, I feel these possibilities still receive less attention than their significance deserves. Far more often, I come across another prediction of mass unemployment or loss of control. This is not about uncritical enthusiasm. It is about giving a promising drug candidate, a better understanding of disease or a new way to verify knowledge just as much attention. And asking more often which of these possibilities we can put to use ourselves.