AI in Drug Discovery: From Protein Structure Prediction to Designed Molecules

Bringing a new medicine to market is slow, expensive and failure-prone. Most candidate molecules fail somewhere between early discovery and clinical approval, often because they do not bind their target well enough, cause unexpected toxicity or simply do not work in humans. A large part of the early cost lies in understanding biological targets and searching an astronomically large chemical space for molecules that might act on them.

Proteins are central to that search. Most drugs act by binding to proteins, and a protein's function depends on its three-dimensional shape. For decades, determining structures experimentally through X-ray crystallography, NMR or cryo-electron microscopy took months to years per protein. That bottleneck is what AI has transformed. My own work sits in health tech rather than wet-lab biology, but this is one of the clearest examples I know of AI producing a genuine scientific breakthrough, and every engineer interested in AI for health should understand it.

AlphaFold 2

DeepMind's AlphaFold 2, described in Nature in 2021, predicts a protein's 3D structure from its amino acid sequence with accuracy that, for many proteins, is competitive with experimental methods. Its key ingredients are:

• Multiple sequence alignments (MSAs). Evolutionarily related sequences reveal which amino acid positions change together, a strong hint that they are close in 3D space.
• The Evoformer. An attention-based architecture that jointly reasons over the MSA and a pairwise representation of residue relationships.
• A structure module. Converts those representations into 3D atomic coordinates, refined iteratively, with a per-residue confidence score (pLDDT) that tells users which regions to trust.

The AlphaFold Protein Structure Database, run with EMBL-EBI, makes predicted structures freely available for a very large number of proteins. The significance of this work was recognised when the 2024 Nobel Prize in Chemistry was awarded to David Baker for computational protein design, and to Demis Hassabis and John Jumper for protein structure prediction.

Language models for proteins

Protein language models treat amino acid sequences like text. Meta's ESM-2 and ESMFold, described in Science, learn structural information from sequence alone through masked language modelling, enabling fast structure prediction without building an MSA.

Beyond single proteins

AlphaFold 3 extends prediction to complexes involving proteins, nucleic acids, small molecules and ions using a diffusion-based architecture, which is directly relevant to how drugs bind their targets.

Predicting structures is only half the story. Generative models now design new proteins and molecules.

• Protein design. RFdiffusion, from David Baker's lab, applies diffusion models, the same family behind image generators, to generate novel protein backbones that bind specified targets. Some designs have been validated experimentally.
• Small-molecule generation. Generative models propose candidate drug-like molecules optimised for predicted binding, solubility and synthesisability. Graph neural networks predict molecular properties directly from chemical structure.
• Virtual screening. ML-accelerated docking and binding affinity prediction let researchers evaluate far more candidates computationally before committing lab resources.
• Target identification. Models over genomic, transcriptomic and literature data help identify which proteins are worth targeting for a disease.

Several AI-discovered or AI-designed drug candidates have entered clinical trials in recent years. It is still early, and the real test, whether AI meaningfully improves success rates in human trials, will take years to answer.

• Static structures. Predicted structures are snapshots. Proteins move, and many are partially disordered; function often depends on dynamics that current models capture poorly.
• Binding is not efficacy. A molecule that binds well in silico may be toxic, poorly absorbed or ineffective in the body. Wet-lab validation and clinical trials remain essential.
• Confidence must be read. Low-confidence regions in AlphaFold predictions should not be treated as reliable structure.
• Data bias. Genomic and clinical datasets heavily under-represent African populations, despite Africa holding the greatest human genetic diversity. Drugs and targets discovered with skewed data may not work equally well for everyone.
• Biosecurity. The same design capabilities that create therapeutics could be misused. Responsible access policies and screening of synthesis requests are part of the field's ethics.

Africa carries a heavy burden of diseases, malaria, tuberculosis, sickle cell disease, neglected tropical diseases, that historically received less pharmaceutical investment. Open tools like AlphaFold DB and open-source models lower the barrier for African researchers to work on these targets without massive infrastructure.

1. Use open resources first. AlphaFold DB, ESM models and open cheminformatics libraries such as RDKit let small teams do meaningful computational work.
2. Collaborate across disciplines. Software engineers, data scientists, biologists and chemists need shared vocabulary. Engineers can contribute robust pipelines, data infrastructure and reproducible workflows.
3. Invest in representative data. Support genomic and clinical data initiatives that include African populations, with strong consent and benefit-sharing frameworks.
4. Build reproducible pipelines. Version data, models and parameters; computational biology results must be reproducible to be trusted.
5. Use cloud compute strategically. Many predictions can run on modest GPU instances; reserve large-scale compute for screening campaigns that justify it.

If you come from backend or data engineering, as I do, the entry point into computational biology is more accessible than it looks. Start by downloading a few predicted structures for a protein linked to a disease you care about from AlphaFold DB, and open them in a free viewer such as Mol* or PyMOL. Learn to read the pLDDT confidence colouring before drawing any conclusions.

Next, run a protein language model such as ESM-2 locally or in a notebook to generate embeddings for a set of sequences, and train a simple classifier on top of them. Explore RDKit to compute molecular descriptors and fingerprints for small molecules. These exercises teach the data formats, FASTA, PDB and mmCIF, SMILES, that underpin the field.

Finally, find collaborators. University departments of biochemistry and pharmacology, including those in Lagos, often have scientific questions and too little engineering capacity. A reliable pipeline, a clean database and a reproducible workflow are contributions scientists genuinely value.

• AlphaFold 2 and protein language models such as ESMFold made accurate structure prediction widely accessible, recognised by the 2024 Nobel Prize in Chemistry.
• AlphaFold 3 and diffusion-based design tools such as RFdiffusion extend AI from predicting structures to designing binders and molecules.
• Predicted binding is not clinical efficacy; experimental validation and trials remain indispensable.
• Under-representation of African populations in biological data is a serious equity and scientific gap.
• Open tools let small, interdisciplinary teams, including in Africa, contribute to drug discovery for neglected diseases.

1. Highly accurate protein structure prediction with AlphaFold — Nature
2. Accurate structure prediction of biomolecular interactions with AlphaFold 3 — Nature
3. Evolutionary-scale prediction of atomic-level protein structure with a language model — Science
4. De novo design of protein structure and function with RFdiffusion — Nature
5. AlphaFold Protein Structure Database — EMBL-EBI and Google DeepMind
6. The Nobel Prize in Chemistry 2024 — NobelPrize.org

Cover image: original illustration created for this article. Cited works remain the property of their respective authors and publishers.