The overlap between artificial intelligence and synthetic biology is entering a new phase, according to Al Jazeera, as computational models move beyond forecasting the structure of proteins and begin proposing new molecular designs that can then be tested in a lab. The Qatar-based broadcaster reported on the shift on Monday, describing a field that spent decades relying on slow, costly experimental work now turning to software that can suggest candidate molecules before any bench work begins.
Al Jazeera frames the change as a transition from prediction to design — the point where generative AI meets synthetic biology, a discipline that applies engineering principles to build biological systems or components for a defined purpose, such as a therapeutic protein, an industrial enzyme or a cell engineered to carry out a specific function.
Why it matters
For years, the bottleneck in computational biology was structural: working out the three-dimensional shape of a protein from its amino acid sequence. Experimental methods could take considerable time and resources, particularly for complex structures, and progress lagged far behind the volume of genomic sequences piling up in databases. Al Jazeera reports that feeding a sequence into a computer model can now yield a predicted 3D shape and hints about how the protein may interact with other molecules.
That matters beyond the laboratory. The same tools that shorten drug discovery timelines could, in the wrong hands, lower barriers to making dangerous biological agents — a tension the reporting highlights through linked coverage warning about the risk of AI accelerating bioweapon development alongside medicine.
From AlphaFold to genomic language models
Al Jazeera traces the current wave to AlphaFold 2, released in 2020, which predicts a protein's three-dimensional structure from its sequence. Google DeepMind, the model's developer, says determining a single protein structure experimentally can demand significant time and resources, especially in complicated cases. The AlphaFold database now holds more than 200 million predicted protein structures, giving researchers a far wider set of known proteins to explore, according to the report.
The next step came in 2024 with the published study on AlphaFold 3, which extended prediction beyond proteins to complexes that include nucleic acids, small molecules and ions. Tests reported in the journal Nature showed the model outperforming several specialized tools at predicting certain molecular interactions, including protein binding to small molecules and nucleic acids and antibody-antigen interactions. Al Jazeera notes that an AlphaFold Server gives researchers access to the model for non-commercial research, while the database offers free access to the predicted structures.
A different approach underpins ESMFold, which draws on language models trained on natural protein sequences. Rather than depending on the multiple evolutionary comparisons once required for accurate prediction, it uses learned patterns to predict structure directly from sequence, speeding up the process and widening exploration of proteins from microorganisms and natural environments. Al Jazeera cites a 2023 study in Science in which researchers used this approach to build an atlas of more than 617 million predicted protein structures from environmental biological data, including over 225 million classified at a high confidence level.
The reporting also covers models aimed at DNA rather than proteins. AlphaGenome examines the relationship between DNA sequence and gene function and regulation — a harder problem than identifying which gene produces a given protein. It analyzes the stretches of the genome that do not translate directly into proteins but govern how genes operate, helping estimate whether a particular mutation or DNA change is linked to disease and predicting multiple regulatory features at once from sequence alone. Evo 2, described as a genomic language model, learns patterns in DNA sequences much as text models learn patterns in human writing; it can read sequences, design new genetic sequences and entire genomes, and estimate the likely functional impact of genetic changes, including variants tied to inherited disease.
Design, drugs and the limits of the evidence
Al Jazeera reports that AI now helps filter promising designs and narrow the list of candidates that need early laboratory testing, shortening the traditional design-build-test cycle. In a 2025 study in Science, researchers showed that a model called ESM 3 could generate a new fluorescent protein whose sequence differs substantially from known ones; the authors estimated the feat was equivalent to simulating more than 500 million years of evolution. The same reporting describes computationally designed binding proteins aimed at antigens linked to brain tumors, incorporated into genetically modified immune cells, with laboratory and animal tests showing improved anti-tumor properties — results still at the preclinical stage.
In drug discovery, the report points to Rentosertib, an experimental candidate discovered with AI assistance that targets a protein involved in pathways of idiopathic pulmonary fibrosis. Early phase 2 trial results were described as encouraging, and the drug has moved into phase 3 testing. Separately, researchers used deep learning to design small proteins that bind neurotoxins found in snake venom; lab and animal tests showed the proteins could neutralize the toxins and protect mice from lethal doses.
Al Jazeera is careful about the caveats: most medical applications remain in research or early clinical trials and cannot yet be treated as proven replacements for conventional treatments. A successful computational design does not guarantee a drug will work in humans, and the reporting notes that clinical testing remains the decisive step.
What to watch: whether the next round of AI-designed candidates clears late-stage trials, and how regulators and biosecurity bodies respond as design tools become more widely accessible.
