What problem does it solve?
Conventional small molecule drug discovery starts from a target hypothesis and then screens, makes and tests thousands of candidate compounds to find a handful worth advancing, before any clinical testing begins. Insilico Medicine puts traditional early stage discovery at two and a half to four years to a preclinical candidate; Recursion Pharmaceuticals puts the industry average at over 2,500 compounds synthesized and 42 months per program.
An AI native platform tries to spend that cost computationally instead: predicting which targets are worth pursuing and which molecules are likely to bind, be selective and be safe before a chemist makes anything, so the synthesis and testing budget is spent on a much shorter list of higher probability candidates. It does not remove the need for real chemistry, real assays or real clinical trials; it changes how many molecules a team has to make to find one worth testing in a person.
How does it work?
- Model the biology. Multi omics data, scientific literature and patent intelligence feed models that score and rank potential disease targets against criteria such as novelty, druggability and safety.
- Generate candidate molecules. Generative chemistry models propose novel molecular structures against the chosen target, rather than screening only a fixed, existing compound library.
- Predict properties computationally. Models predict binding, selectivity, toxicity and pharmacokinetic properties before any molecule is made, narrowing a large design space to a short list worth synthesizing.
- Synthesize and test the short list. Chemists make and test only the highest scoring molecules, and the assay results feed back into the models for the next design round.
- Advance to preclinical and clinical development. The nominated candidate moves into the same regulated preclinical and clinical pathway as any other drug. In this use case, the AI's role ends at candidate nomination (some platforms, such as Insilico's, also offer separate clinical trial prediction tools, which are a different use case); every later stage here is validated by conventional testing.
- Audience
- Employee facing
- Autonomy
- Copilot
- Adoption
- Early adopters
- Channels
- Internal tools
What is it worth?
Benchmarks are computed from the public deployments below: one data point per organization per KPI, with who made each claim.
| KPI | Median | Reported range | Data points | Claimed by |
|---|---|---|---|---|
| Cycle time reduction | Too few to pool | 60% to 60% | 2 | 2 organization |
Value drivers: Speed and cycle time, Lower cost to serve, Employee productivity.
Indicative value
A biotech starting 10 early discovery programs a year
USD 18 million to USD 72 million
Annual discovery program cost avoided through faster preclinical candidate nomination per year
How this is calculated
Formula: programs * monthsSaved * costPerMonth. The low scenario uses every low input, the high scenario every high input.
| Input | Low | High | Basis |
|---|---|---|---|
| Discovery programs started per year programs, programs per year | 10 | 10 | The reference biotech. |
| Months saved per program versus a conventional discovery timeline monthsSaved, months per program | 12 | 24 | Conservative against the benchmarks on this page (Insilico Medicine reports reaching preclinical candidate nomination in an average of 12 to 18 months against a traditional 2.5 to 4 years; Recursion Pharmaceuticals reports about 17 months against an industry average of 42 months), because a first program on a new platform rarely matches a mature platform's average. |
| Fully loaded cost of a discovery program team per month costPerMonth, USD per program per month | 150,000 | 300,000 | Editorial assumption for a mid sized medicinal chemistry and biology team; replace with your own. |
What it leaves out: Time saved in discovery only. It leaves out the platform's own cost, the preclinical and clinical development that follows candidate nomination (unchanged by this use case), and the fact that a faster nomination is not the same as a successful drug: Recursion Pharmaceuticals' own risk disclosure says the risk of failure in pharmaceutical research and development is high and failure can occur at any stage before or after regulatory approval, AI discovered or not.
Who already uses it?
2 public deployments, strongest evidence first. Grades: A regulator or audit, B the organization itself, C vendor case study, D anonymous or estimate.
Insilico Medicine
Global · Pharma and life sciences · 2026
Insilico Medicine runs an end to end AI platform, Pharma.AI, that combines target discovery (PandaOmics), generative molecule design (Chemistry42) and translational and clinical support tools to move programs from a biological hypothesis to a nominated preclinical candidate. Its most advanced program, rentosertib (formerly ISM001-055 / INS018_055), a TNIK inhibitor for idiopathic pulmonary fibrosis whose target and molecule were both identified and designed with this platform, entered a Phase III clinical trial in July 2026 after a randomized Phase IIa trial published in Nature Medicine showed a dose dependent lung function signal. Beyond rentosertib, the company reports 31 preclinical candidate nominations from its pipeline, 13 of which received IND clearance and 8 of which have reached ongoing Phase I trials.
- Cycle time reduction: about 60%, as stated in Insilico's June 2025 Nature Medicine publication release (22 nominated candidates, 2021 to 2024); the July 2026 release quoted above restates the same 12 to 18 month range without an updated candidate count
"While traditional early-stage drug discovery typically takes 2.5 to 4 years, Insilico has consistently reached preclinical candidate (PCC) nomination in an average of just 12 to 18 months, with only 60 to 200 molecules synthesized and tested per program."
Claimed by: organization
Recursion Pharmaceuticals
United States · Pharma and life sciences · 2026
Recursion Pharmaceuticals, which completed its business combination with Exscientia in November 2024, runs an AI native operating system that combines phenomic screening with automated, precision small molecule chemistry to take programs from an initial hit to a development candidate. As of its February 2026 results, the company reports the platform has delivered more than 10 development candidates to date, including REC-617, a CDK7 inhibitor identified as lead candidate in under 11 months with 136 novel compounds synthesized, and REC-7735, a PI3Kα H1047R inhibitor precision designed with 242 compounds synthesized from first novel hit to REC-7735 in 10 months, now in IND enabling studies. Recursion's partnership with Sanofi has the potential for up to 15 AI designed small molecule programs, of which 5 or more span immunology and oncology, and had reached five progress based milestones as of the same results.
- Cycle time reduction: about 60%, reported February 2026; average for advanced candidates delivered by the platform, per program
"Advanced candidates have been delivered by synthesizing ~330 compounds per program in ~17 months, compared to industry averages of over 2,500 compounds and 42 months, respectively."
Claimed by: organization
How do you implement it?
A model agnostic playbook: what to prepare, the order to build in, and what goes wrong.
Data you need
- Curated multi omics data (genomics, proteomics, disease models) for target scoring
- A large internal dataset of assay results (binding, ADMET, toxicity) to train and validate predictive models against
- Structural biology data for the targets in scope, where available
Systems to integrate
- Electronic lab notebook and assay data management systems
- Compound registration and inventory systems
- Computational chemistry and structural biology tools
Complexity: High
The AI shortens discovery, not validation. A nominated candidate still needs full preclinical safety and efficacy testing, an investigational new drug filing and clinical trials under existing pharmaceutical regulation. Getting real predictive power out of the models needs a large, curated internal dataset of assay results to train and validate against, not just public data.
- 1
Pick a target class the model can actually score
Start where public and internal data on the biology and known chemical matter is strongest; a completely novel target with almost no data starves the model of signal.
- 2
Set a synthesize and test budget per round
Cap how many AI proposed molecules a chemistry team commits to making per cycle, and measure hit rate against that budget, rather than running an open ended search.
- 3
Keep a qualified chemist in every loop
A medicinal chemist reviews every proposed structure for synthesizability and known liabilities before it is made, and can veto a candidate the model scored well.
- 4
Track the whole funnel, not just discovery speed
Preclinical candidate nomination is not success. Follow programs into IND filing and Phase 1 to see whether faster discovery produces better drugs, not just more of them, faster.
- 5
Validate predictions against your own assay data before trusting them
A platform's published benchmark numbers come from its own historical programs. Run a blinded internal validation before betting a program's timeline on the model.
Guardrails
- A qualified medicinal chemist reviews and approves every AI proposed molecule before synthesis
- Predictive model performance is validated against blinded internal assay data, not only the platform's own published benchmarks
KPIs to instrument
- Molecules synthesized and tested per program, against the target budget
- Time from project initiation to preclinical candidate nomination
- Attrition rate of AI nominated candidates through IND filing and Phase 1, compared with the company's historical baseline
Human in the loop
Chemists and biologists review every AI proposed target and molecule; the AI narrows the design space and predicts properties, it does not decide what gets synthesized or what advances. A program only moves to candidate nomination after conventional wet lab confirmation of the predicted properties.
Common failure modes
- Optimizing for a metric that is not a drug
- A model can hit its binding or property targets and still nominate a molecule that fails for reasons the model was not trained to predict, such as manufacturability or an off target effect found only in later testing. Track attrition through clinical development, not only discovery speed.
- Overfitting to public and vendor benchmark data
- A platform's published cycle time and molecule count figures come from its own historical programs and target classes. Validate on your own target and internal data before assuming the same numbers apply.
What are the risks and rules?
EU AI Act
Minimal risk
Target scoring and molecule generation are not a safety component of an Annex I product and are not one of the Annex III high risk areas (Article 6), so they do not become high risk on that route. Insofar as the platform and its training are themselves scientific research and development, activity that stays there falls outside the Regulation entirely under the Article 2(6) research exclusion. The resulting drug candidate is separately regulated as a medicine, not as an AI system, through the normal pharmaceutical approval pathway; conventional preclinical and clinical testing validates the AI's outputs before anything reaches a patient.
Rules that apply
Guidance
- Article 2: Scope (European Union, Europe). Article 2(6) says the Regulation "does not apply to AI systems or AI models, including their output, specifically developed and put into service for the sole purpose of scientific research and development"; Article 2(8) extends this to "any research, testing or development activity regarding AI systems or AI models prior to their being placed on the market or put into service." Article 6 itself, by contrast, says nothing about research: it sets the Annex I safety component route and the Annex III route for high risk classification.
- Considerations for the Use of Artificial Intelligence To Support Regulatory Decision-Making for Drug and Biological Products (US Food and Drug Administration, North America). This January 2025 draft guidance covers AI models used to produce information that supports a regulatory decision on a drug, such as evidence in a submission. It explicitly excludes this use case's discovery stage: "the use of AI for the purposes of drug discovery is not in the scope of this guidance," so a platform that stops at candidate nomination sits outside it; the guidance's credibility assessment framework becomes relevant only once AI generated outputs are used to support a later regulatory decision.
Controls to put in place
- Documented model validation against internal assay data before a program relies on a prediction
- Named chemist and biologist sign off on every candidate before it advances to synthesis and formal preclinical testing
- Data provenance and access controls over the proprietary assay data used to train and validate models
Frequently asked questions
- Has an AI discovered drug reached patients yet?
- Insilico Medicine's rentosertib, a candidate for idiopathic pulmonary fibrosis whose target and molecule were both identified with its AI platform, became the company's first asset to enter a Phase III clinical trial, in July 2026, after a randomized Phase IIa trial published in Nature Medicine showed a positive lung function signal. According to Insilico, rentosertib remains investigational and has not been approved by any regulatory authority.
- How much faster is AI native drug discovery?
- Insilico Medicine reports reaching preclinical candidate nomination for 22 programs between 2021 and 2024 in an average of 12 to 18 months, against a traditional 2.5 to 4 years, synthesizing only about 60 to 200 molecules per project. Recursion Pharmaceuticals reports advanced candidates have been delivered by synthesizing about 330 compounds per program in about 17 months, against industry averages of over 2,500 compounds and 42 months. Both figures describe the companies' own historical programs, not a guarantee for any specific target.
- Does this replace medicinal chemists?
- No. The platform proposes structures and predicts properties, but the molecules still have to be made and tested in real assays, and the programs on this page still rely on chemistry and biology teams to validate what the models propose. We recommend a qualified chemist review every AI proposed structure before synthesis (see the implementation guardrails).
- Is this the same as AI used in clinical trials or regulatory writing?
- No. This use case covers the discovery stage, from target identification to a nominated preclinical candidate. Matching patients to trials and drafting clinical study reports are separate, later stage use cases with their own evidence.
How to cite this page
Blits.ai AI Use Case Library, "AI native platform for drug target discovery and molecule design", last verified 28 September 2026, https://www.blits.ai/ai-use-cases/ai-drug-discovery-platform. Licensed under CC BY 4.0. Method: how we verify use cases.
Changelog
- 28 September 2026: First published