AI in drug discovery has been generating headlines for years. Some of them have been justified; many have been premature. In 2026, enough time has passed that we can start to distinguish the areas where AI is genuinely delivering results from those where the technology is still catching up to the claims.
This is a ground-level assessment of where things actually stand, written for scientists who want an honest picture rather than a press release summary.
Where the Evidence Is Strongest: Structure Prediction
The most unambiguous AI success in drug discovery is protein structure prediction. AlphaFold 3, published in Nature in 2024 by Abramson et al., extended AlphaFold’s capabilities to predict the structures of protein-ligand, protein-DNA, and protein-RNA complexes, not just proteins alone. This is directly relevant to drug discovery: identifying where a small molecule binds to its target, and how the complex behaves, is foundational to structure-based drug design.
The practical impact is real. Structural biologists who previously depended on X-ray crystallography or cryo-EM to get a target structure for a new project can now generate high-confidence structural hypotheses in hours rather than months. This is not a replacement for experimental structure determination (especially for key validation points), but it has genuinely accelerated the early stages of structure-based drug design programs at dozens of pharma and biotech companies.
Limitations remain. AlphaFold 3 is less reliable for intrinsically disordered proteins (which includes many therapeutically relevant targets), and predicted structures for complex allosteric mechanisms require experimental validation. But the technology is clearly working and is being used in production drug discovery programs.
Molecular Generation: Promising, but Pipeline Stage Still Matters
Generative AI for small molecule design has attracted enormous investment and substantial claims. The reality in 2026 is more nuanced than either the hype or the skepticism suggests.
Several companies (including Recursion, Insilico Medicine, and Exscientia) have AI-designed molecules in clinical trials. This is a genuine milestone. Whether these molecules will eventually outperform conventionally designed compounds in clinical outcomes is not yet known.
The honest assessment: AI generative models are effective at producing large numbers of molecules that satisfy defined property criteria (molecular weight, predicted binding, aqueous solubility, synthetic accessibility). They are efficient at early-stage hit generation and scaffold exploration. What they do not yet do reliably is predict clinical success, which depends on biological complexity, patient heterogeneity, and ADME/tox properties that remain difficult to model accurately.
The most rigorous published analysis of AI-designed drug candidates remains modest: a 2023 analysis by Jayatunga et al. in Drug Discovery Today examining the clinical-stage AI-assisted pipeline found that the programs are real but that demonstrating a quantitative improvement in attrition rates over conventional approaches will require more time and more candidates reaching late-stage trials.
Target Identification: Mixed Results
AI for target identification (finding which proteins to drug for a given disease) is the area where expectations and reality diverge most. The challenge is that predicting whether a target is causally relevant in human disease, not just statistically associated in a genomic dataset, is genuinely hard. Human genetic validation (Mendelian randomization, rare variant associations) remains more predictive of clinical success than most ML-based target prioritization approaches.
That said, multi-omic integration approaches using large language models and graph neural networks have demonstrated real utility in specific contexts: identifying combination therapy opportunities, finding synthetic lethal interactions, and prioritizing targets within a known pathway. The key is using AI as one input alongside genetic evidence and mechanistic biology, not as a replacement for either.
ADME and Toxicity Prediction: The Underrated Win
One area where AI is delivering consistent, practical value with less fanfare is ADME (absorption, distribution, metabolism, excretion) and toxicity prediction. These are properties that determine whether a promising molecule is actually viable as a drug, and computational prediction of them is now incorporated into most modern medicinal chemistry workflows.
Models like SwissADME and commercial platforms from companies like Schrödinger and OpenEye have improved meaningfully with machine learning. Predicting metabolic liabilities, hERG cardiotoxicity risk, and blood-brain barrier permeability with reasonable accuracy at the virtual screening stage allows chemists to filter out high-risk compounds earlier. This reduces experimental waste even if it does not eliminate the need for wet lab validation.
What This Means for Scientists in Life Science
If you work in drug discovery or biotech, a few things follow from where the technology actually stands:
Structure prediction tools are now a practical part of the workflow for structure-based design programs, and familiarity with AlphaFold outputs is increasingly assumed in computational chemistry and structural biology roles.
Generative chemistry tools are being adopted broadly but their value is in accelerating hit-to-lead work, not in replacing medicinal chemistry expertise. Scientists who understand both the chemistry and the model limitations are more valuable than those who treat outputs as reliable without question.
Target prioritization remains a domain where human judgment, genetic evidence, and mechanistic understanding still outperform pure computational approaches for high-stakes decisions.
The companies most credibly using AI are those integrating it into specific, well-defined steps in their pipeline rather than claiming end-to-end AI drug discovery. The latter claim is not yet backed by evidence.
The Bottom Line
AI is a real and growing contributor to drug discovery in 2026. Protein structure prediction has delivered transformative capability. Molecular generation and ADME prediction are producing genuine workflow improvements. Target identification is the area most in need of more validation data before strong claims are justified.
Scientists working in the field should engage with these tools directly rather than relying entirely on vendor claims or media coverage. The Coursera Drug Discovery and Development course from UCSD provides useful grounding in the broader pipeline context for those newer to the field.
The next two to three years will be important: if the AI-designed molecules currently in clinical trials show improved attrition rates compared to historical benchmarks, the case for AI-driven drug discovery will become substantially stronger. If they fail at similar rates, it will push the field toward more realistic claims about what AI can and cannot do in this context.