Artificial intelligence can shorten parts of drug discovery. That does not establish that an AI-assisted medicine will reach patients sooner, cost less to develop or have a better chance of approval. Each claim needs different evidence. A model that ranks compounds accurately has answered a narrower question than a clinical trial that measures whether treatment helps patients. For investors, the task is to identify which uncertainty a platform has reduced, then decide what that improvement is worth. Within the broader assessment of biotech science and drug technologies, AI deserves scrutiny at the level of individual research decisions. The useful question is whether it improves the selection and development of medicines against a credible alternative, with the costs and unsuccessful projects included.
What is AI doing in drug discovery?
AI in drug discovery covers several activities, rather than one technology or business model. Machine learning models learn patterns from training data and use them to predict properties, rank possibilities or propose new designs. Generative models can suggest molecular structures; other models, prioritise existing compounds or disease targets. Assess the output before judging the platform. A target ranking, a predicted binding structure and a molecule with measured activity represent different levels of progress. So do a laboratory hit and a development candidate, which must meet a broader set of requirements. Treat the label “AI-discovered” as the beginning of the investigation. It does not reveal how much work the model performed, which decisions scientists overruled or whether a conventional workflow could have reached the same result.
| Research activity | Evidence to request | Question that remains |
|---|---|---|
| Target selection | Experiments testing whether changing the target alters disease biology | Will changing it help patients without unacceptable harm? |
| Structure prediction | Comparison with experimental structures and relevant binding measurements | Does the predicted arrangement support useful drug activity? |
| Compound ranking | Prospective testing of nominated compounds against a matched baseline | Does ranking improve the quality of compounds entering development? |
| Molecular design | Synthesised molecules with measured activity and suitable properties | Can the designs become practical development candidates? |
| Property prediction | Predictions checked against new exposure, metabolism or toxicity measurements | Does using the model improve project decisions? |
AlphaFold 3 illustrates both technical progress and the boundary around it. The model predicts joint structures involving proteins, nucleic acids, small molecules and other components. Its 2024 research paper reported improved performance across several structural prediction benchmarks, while documenting errors involving molecular geometry, overlapping atoms and some target classes. It also distinguished predicted static structures from the behaviour of molecules in solution. These are meaningful capabilities, with measurable limitations. They do not establish that a proposed drug has suitable exposure, tolerability or clinical benefit. A strong structural prediction can justify an experiment; treating it as a substitute for the rest of development overstates what was tested. The evidence is set out in the original AlphaFold 3 research paper.
Separate biological reasoning from molecular engineering. A platform might design an effective inhibitor of a protein yet fail to establish that inhibiting that protein improves the disease. Conversely, a compelling target can remain difficult to address with a practical medicine. Ask which problem the company claims to solve, and which evidence connects that problem to its candidate. The assessment of drug targets and mechanisms of action remains relevant whether scientists selected the target through genetic evidence, experimental screening or an AI model. Better predictions cannot rescue a development programme built around the wrong biological assumption.
Define the clock before judging faster development
Any claim of faster drug discovery needs a start date, an end date and a comparable alternative. “From target to candidate” can exclude the years spent building a dataset, developing assays or studying a disease. “From design to trial” may begin after substantial biological work has already occurred. Neither measurement is automatically misleading, but both become misleading when compared with a broader industry timeline. Request a dated project history that includes assay preparation, synthesis, testing and candidate selection. Then separate discovery time from the work covered by Phase 1, Phase 2 and Phase 3 trials. Time saved before human testing should not be presented as an equivalent reduction in the duration of clinical development.
Consider a hypothetical programme with two years of discovery followed by eight years of subsequent development. Cutting discovery to one year halves that stage, but reduces the full timeline from ten years to nine, assuming everything else remains unchanged. The benefit is real. Its size depends on the denominator. The same discipline applies to claims that a platform screens millions of compounds in hours. Ask how long the complete workflow takes, including obtaining or making compounds and confirming their properties. A computation that finishes overnight creates little calendar advantage if the next experiment waits months for capacity. Measure progress along the sequence of tasks that determines the project’s completion date, rather than the speed of its most photogenic step.
Rank the evidence behind an AI platform
Retrospective benchmarks: useful, but easier to overinterpret
A retrospective benchmark tests performance on data that already exist. Its value depends partly on how training and test data are separated. If closely related compounds appear in both groups, a model may face an easier task than predicting unfamiliar chemistry. Scaffold splits separate molecules by structural framework; time splits train on older observations and test on newer ones. The MoleculeNet benchmark implemented these approaches and highlighted why random splitting can be unsuitable for chemical data. Neither approach should be treated as an automatic guarantee of a realistic test. Request the split method, data provenance and comparison models, then check whether the evaluation resembles the company’s intended use. A result on familiar compounds does not answer every question about performance on a new target. The MoleculeNet benchmark paper provides a practical foundation for examining these distinctions.
Prospective experiments: predictions made before results are known
Prospective testing provides a stronger test of a discovery claim when the model nominates compounds before their experimental outcomes are available. For an investment assessment, request a documented selection process and the full denominator: how many designs were proposed, how many could be made, how many were tested and how many met the stated criteria. Ask whether those criteria changed after the results arrived. A useful comparison gives the AI workflow and an alternative workflow comparable resources and the same experimental standards. Even a well-run prospective study remains evidence about the tested task. The distinction between a promising laboratory finding and a medicine is covered in the assessment of what preclinical research can establish.
CACHE Challenge #1 offers a concrete example of prospective computational evaluation. Twenty-three participants nominated compounds predicted to bind the WDR domain of LRRK2, a protein associated with Parkinson’s disease. Compounds underwent experimental testing, followed by a further selection and testing round. The organisers reported multiple confirmed chemical series, but characterised all the hits as weak. They also used additional assays and checks for aggregation or poor solubility to challenge apparent binding results. That combination matters: the programme produced experimental starting points while making their limitations visible. It did not turn a successful binding prediction into evidence of a Parkinson’s treatment. The CACHE Challenge #1 experimental results show the distinction between finding a hit and establishing a useful medicine.
Rentosertib: clinical evidence, with clear boundaries
Rentosertib, developed by Insilico Medicine, provides a documented clinical example. Its development used AI for target identification and molecular design, producing a small molecule inhibitor of TNIK for idiopathic pulmonary fibrosis. The 2025 publication reported preclinical candidate nomination within 18 months of starting target discovery. Its randomised Phase 2a trial enrolled 71 patients across placebo and three treatment groups for 12 weeks. The primary endpoint concerned treatment-emergent adverse events, rather than efficacy. At the highest dose, 60 mg once daily, mean forced vital capacity increased by 98.4 ml, compared with a decrease of 20.3 ml for placebo. Liver toxicity and diarrhoea were among the events leading to discontinuation. These findings support further investigation; a small, short study does not establish durable benefit or broad safety. It also does not show how quickly an equivalent conventional discovery programme would have progressed. Those boundaries are visible in the published rentosertib Phase 2a trial.
Keep two assessments separate when a candidate reaches patients. One concerns the medicine: whether the design, endpoints and results support further development. The other concerns the discovery platform: whether its contribution can be measured and repeated. A favourable trial result can strengthen the case for the candidate without proving that every programme using the platform deserves a higher success probability. A disappointing result, meanwhile, requires examination of where the original prediction stopped and the failed hypothesis began. Use the same discipline applied when reading clinical trial results, rather than allowing the discovery method to change the standard of evidence.
Why clinical success rates need careful interpretation
Published success rates can look more decisive than the underlying dataset permits. A 2024 analysis of pipelines at AI-native biotech companies reported Phase 1 success rates of 80–90% and a Phase 2 rate of about 40%, with a small Phase 2 sample. Those were historical findings from that analysis, not a present-day probability for any AI-designed asset. The distinction between phases also matters: the authors interpreted the early results as encouraging for identifying molecules with drug-like properties, while the Phase 2 outcome was comparable with historical industry averages. The 2024 analysis of AI-discovered drugs in clinical trials is evidence to examine, not a universal success-rate assumption to copy into a valuation.
Before accepting a comparison, inspect the cohort. Does “AI-discovered” include repurposed medicines, familiar targets and wholly new molecular designs? Were discontinued programmes captured consistently? How were trials still running at the analysis date treated? Compare disease areas and development objectives, too. A group containing different kinds of programmes cannot isolate the effect of the discovery method without further analysis. For valuation work, assign probabilities to the asset’s evidence and remaining risks, rather than attaching a platform-wide uplift to every candidate. A disciplined approach to estimating clinical and regulatory success probabilities should make the reason for any adjustment explicit. Otherwise, a promising observation becomes an assumption repeated across the entire pipeline.
Assess the data advantage, not the model’s vocabulary
When assessing a claimed data advantage, ask what the company can do with its information that an alternative team cannot readily reproduce. Dataset size alone is a weak answer. Request evidence about experimental consistency, coverage of relevant chemistry and the inclusion of unsuccessful experiments. Ask whether new laboratory results return to the modelling workflow, and whether that feedback has improved later decisions on unseen projects. Also examine transfer: a model trained around one assay or target family should be evaluated separately before its performance is assumed elsewhere. This is where the distinction between a biotech platform and a single-asset business becomes useful. A platform claim calls for repeatability across projects, rather than repeated presentations about its best candidate.
Check control over the data as well as access to them. Request clear disclosure of whether partner-generated results can be reused, whether access survives the end of a collaboration and whether restrictions prevent the company from combining datasets. These are diligence questions, not assumptions about any particular agreement. Consider the economics of the feedback process, too. If every extension into a new disease area requires a fresh laboratory operation, the business may need more capital and staff than its software presentation suggests. The relevant test is whether each additional programme benefits from previous work, and whether that benefit is visible in measured output.
Translate research speed into business value
Research acceleration becomes financially useful through identifiable effects: lower spending to reach the same milestone, earlier access to a decision, or better selection of programmes worth funding. Keep these effects separate. In a hypothetical example, completing a discovery project for $6 million rather than $10 million saves $4 million before accounting for the platform’s allocated costs. It does not establish lower trial costs, better clinical outcomes or greater commercial demand. Include computing, data acquisition, laboratory work and failed projects when comparing the two workflows. Then examine whether faster candidate generation increases spending elsewhere. A company that starts more studies may consume cash faster despite improving discovery efficiency, making cash burn and runway relevant to the assessment.
The business model determines who captures any improvement. For a licensing business, inspect the research payments and asset rights alongside conditional milestones. For a company developing its own medicines, examine how much development spending remains before a value-changing result. A large announced partnership value should not be treated as cash already earned, or as experimental proof that the platform works. Request the disclosed payment structure and the activities required to trigger future receipts. The framework for valuing biotech licensing deals helps separate funded work from contingent rewards. Keep any platform valuation distinct from the pipeline valuation, so the same expected benefit is not counted twice under different labels.
AI discovery claims and regulatory evidence are different issues
The FDA’s January 2025 draft guidance proposed a risk-based framework for assessing AI model credibility when model-generated information supports regulatory decisions about drug safety, effectiveness or quality. Its scope expressly excluded AI used for drug discovery, along with certain operational uses that do not affect patient safety, product quality or study reliability. It was issued as draft, nonbinding guidance. That distinction prevents a common category error: a framework for AI-generated regulatory evidence is not blanket validation of an AI discovery platform. When a company cites regulatory discussions, ask which model use and development question those discussions concern. Do not treat them as proof of faster approval or acceptance of every platform output. The boundaries appear in the FDA draft guidance on AI supporting drug regulatory decisions.
Questions to put to management before accepting an acceleration claim
Request a project-level evidence package rather than another description of the model architecture. The package should connect a dated prediction to an experiment, a decision and a measurable consequence. It should also make unsuccessful work visible. A claim of fewer synthesis cycles is easier to assess when the company discloses the candidate criteria, the cycles completed and the comparison workflow. A claim of better target selection needs different evidence. Decide which claim is being made before asking whether the presentation supports it.
- What started and stopped the clock? Identify work excluded from the advertised timeline.
- What was the alternative? Request a relevant comparison with comparable resources and objectives.
- What is the denominator? Include failed designs, abandoned projects and compounds that could not be made.
- Was the test prospective? Establish when predictions and success criteria were fixed.
- Has the result repeated? Look for performance on new projects, not repeated analyses of one success.
- Who receives the economic benefit? Check asset ownership, spending obligations and conditional payments.
Record unanswered questions as gaps in the investment case. Do not fill them with industry averages or assumed platform advantages. Cross-check presentations against papers, trial records and financial disclosures using a consistent approach to biotech research sources. Where disclosure stops at a benchmark, describe the evidence as benchmark performance. Where it reaches prospective laboratory validation, state the tested task and its limits. Before the next announcement, write down what result would justify changing the assessment. That gives the next model update, candidate nomination or clinical readout a defined job to do.