TECHNOLOGY EXPLAINER
AI Drug Discovery and Screening
— Even if computation can screen 100 million compounds, synthesis and measurement remain the bottlenecks
Drug discoveryHigh Throughput Screening (HTS)Home Stages of molecular selection for 100 million molecules by calculationContact Us However, if you read the paper, you can choose the candidate by calculationMelting and Metabolism speedCloseSyn and MeasurementIt is decided by the way.
- AI drug discovery and screening (3 lines)
- How to find efficiency
- High Throughput Screening – Count on Experiments
- Perspective of Material Engineer 1: The Library is a 'degradable inventory material'
- Incidence Analysis (1): Docomo - Sprinkle with 100 million slices
- Incident Analysis (2): Structural Prediction — AlphaFold has changed or not changed
- Machine Learning and Generation AI – Learning data is much smaller than the search space
- Automation of Experiments - Robot Scientist
- Examples of advanced clinical stages (one of the first information)
- Material Technician Perspective 2: Candidates of AI can also be obtained with "not melting"
- How regulators handle AI
- Strengths and Limits
- Glossary/Reference Materials/Reference Table
Facts= The publication material describes the contents (with source link)
the analysis of this paper= The value calculated based on the specified premise
Undetermined / Future= Items that are not able to confirm the results of the plan, goal, and、
In addition to this, it is possible to organize structures and read materials and process designs."Description"is distinguished.
About this articleExploration technology (screening, calculation, automation)Home It does not evaluate the effectiveness or safety of individual drugs, and is not medical advice. where clinical trials are touchedStage, Design, and Key Evaluation ItemsThe contents are described in the scope of the first information ( , ClinicalTrials.gov). "Development period has been shortened by AI"Developed companies and authorsContact Us
1. AI drug discovery and screening (3 lines)
- Screening:targeting many compounds (e.g. s and receptors),Steps to pick up effective things (Hit)HOME ExperimentHigh Throughput Screening (HTS)In the calculatorVirtual ScreeningHome
- Where AI Enters:Target proteinThree-dimensional structure prediction(AlphaFold, etc.)Active Prediction(number of points in the study model),rate s(generated AI) andAutomate experiment planning and executionHome
- Unchanged:LastSynthesize, measure and checkrequired. The more you choose,"What to synthesize and measure"The specific gravity of this article is rather large.
The paper of AI drug discovery stands out, “How many hundred millions have you selected?” For example, a study of antimicrobial substances in deep learning,Approx. 1,735 millionIt was evaluated in the and experimented.23 piecesated8 piecesCommentFactsHome ulation is a tool to narrow the candidate,Testing is now underway。
2. How to find efficiency
The research and development of drug discovery has been pointed out that it has not increased results as a result of investment. Scan et al is a paper in 2012, as follows:Facts。
“The number of new drug approvals per billion USD spent on research and development was reduced by half every nine years since 1950, and fell to about one hundred percent after inflation adjustment.”Facts
The paper re-dia ses the cause of the progress that should increase efficiency in terms of science, technology, and management for 60 years.Facts。
There are many successes in clinical trials. Wong et al from January 2000 to October 2015406,038 TopicsClinical trial data (21,143to analyze and 13.8% of the entire development program finally approvedFacts。
How much can I choose "good candidates for muscle" at the stage before entering the clinical practice?the overall efficiency. Expectations for AI and screening technologies are gathered in this article. However,AI AI has improved clinical success rateas seen in Chapter 9, it is still not at the stage that can be seen in primary informationUndetermined / Future。
3. High Throughput Screening – Counts in Experiments
HTSLibraryDMSO, etc. Disp into small wells with robots to measure the reaction with targets at onceHow to use Inglese et al in the Chemical Genomics Center of NIH in the United Statestitative HTS (qHTS)AnnouncedFacts。
“Traditional HTS that test compounds in a single concentration...It is annoyed by frequent false positives and false negatives, and requires a large try.”Facts
qHTSAt least 7 stages of di series4 digit concentration rangeGet the concentration reaction curve of all compounds at once. The concentration of the source plate is most of the compound640 nM〜10 mMCommentFacts。
qHTS5,480 compounds of 60,793 compounds (9.0(atedFactsHome Then, the authors have the same dataOnly 1 point of 11 μMIt is compared to look back when it is judged.
| Judgment (one of 11 μM) | Fake-positive | False Negative |
|---|---|---|
| 3SD threshold from average | 30 compounds (2%) | 845 compounds (40%) |
| threshold of 6SD from average | 5)s (1%) | 1,602 compounds ()) |
AllFacts(Inglese et al 2006 [Reference 3]). False-negative solvents are proportions to compounds that are class 1 and 2 (good curve) with qHTS.
Authors'False-positive can be detected in a trial, but False-negative is not specified in conventional HTS, and the frequency is almost unknown.'NotedFactsHome Also, this precision assay had high accuracy (Z′=0.87)Assays that use cells may reach90% of false positivesFacts。 If the threshold is severe, false positive is reduced, but the threshold increases——What is the design of the measurement? This article explains.
4. Perspective of material ians 1: The compound library is “degradable inventory material”
HTS talks about robots and calculations.Quality control of solution materialsHome You can read three points from primary information.
(1) Reduce during storage.KozikowskiApprox. 7,200 compoundsHome20 mM DMSO solutionstored at room temperature for 1 year and tracked by mass analysis. The probability of the compound is92% in 3 months, 83% in 6 months, 52% in 1 yearCommentFacts。 The library is not an asset of “finish once” but an inventory with expiration and storage conditions(This article explains).
(2) In the aqueous solution, it becomes a "particle".Feng and Shoichet "At micromolar concentrations, many low molecular weight colloidal cohesives self-assemble, and non-specific enzymes and other proteins."FactsHome The frequency30 μM maximum 19% of cinal s and 1-2% of 5 μM.Considering that many HTS campaigns aim to have a hit rate of less than 1%, it is also a significant proportion.FactsHome How to distinguishAdd the non-ionic surfactant (Triton X-100 at 0.01% v/v) to disappear.Home Particle50〜1,000 nmYou may be able to directly observe dynamic light scattering (DLS)Facts。
(3) The structure itself interferes with the measurement system.Baell And Holloway In Many Biochemistry HTSCompounds that appear as repeated hits (PAINS)reporting and identifying partial structure TheyIt is more likely to be treated as a promising starting point in literature, but notand encourage attentionFacts。
"The molecules that were dissolved were actually a dispersion of nanoparticles."——This is one of the most familiar 、omena for colloid and interface engineers. The concept of critical micell concentration, dispersion by surfactant, and particle size measurement in DLSQuality control tools for drug discovery screening(This article explains). As seen in Chapter 7 and later,AI learning data is created from such experimental resultsTherefore, false positive will be the mistake of learning as it is (in this article).
Lipinski et al.Lead compounds obtained by HTS have a higher molecular weight and Log P than those of the previous era of HTS, which tends to be less turbidity.Contact UsFactsHome This paper is famous for its easy-to-understand oral absorption.5 Laws(Hydrogen-binding base 5 or more, receptor 10 or more, molecular weight 500 or more, calculation Log P 5 or more)Facts。 It is easy to deviate to hydrophobic and difficult to dissolve molecules that 'pick up strong bindings by hitting a lot'.That is This article explains. Absorption and dissolution are handled in detail in this series.
5. Incidence Analysis (1): Docomo - Sprinkle with 100 million slices
docomoput candidate s in the targeted protein tri-structure spill (combined part), How much does it work?Simulation. Experimental HTS,molecules that have not yet been synthesizedYou can also evaluate.
How many molecules can be candidates? RuddigtenC・N・O・S・Halogenenumerate organic s, 1,664 billionCreated a database GDB-17. This paper is the scope of this size.Includes many pharmaceuticals and typical lead compoundsFacts。
In 2019, Ryu et al reported the results that actually examined a part of this “expansion space”Facts。
| Paper Description | |
|---|---|
| 130 established reactionsand Enamine70,000 units (building blocks)“make-on-demand” compound made from.Less than 3% of products sold from other suppliers | |
| Size | C β-lactamaseOver 9,900 millionDopamine D4For receptorsMore than 800 million unitsDo |
| Volume | D4about 70 trillion complexes43,563 core time (approx. 1.2 days with 1,500 cores) |
| Syn and Testing | Select 51 to C44 (86'sHome5 in is ts.(Hit Rate 11。 D4Home549 piecesSyn and testing |
| Result | D4Home81 new skeletondiscover and30 Sub Micro MoleHome The higher the number of hits, the higher the number of hits.22〜26%If the number of points becomes worse, it goes down to monotonous.0 |
AllFacts(Lyu et al 2019 [Reference 9]). The 1,664 billion of GDB-17 depends on Ruddigkeit et al 2012 [Reference 8].
Lyu"What kind of molecules can be found?"defined asFactsHome Of the 51 selected44 pieces(86(Facts。
If this is a material development,Search from the "ideal composition found in calculation" but from the "composition made by hand raw materials and processes"Design This article explains. Search spacedefined by the processThat’s why the calculation result is immediately real. Materials Informatics“I can’t make predictions.”It is a design judgement of the suggestion to engineers who have experience in the wall.
OthersHit rate has been down single with the number of docking pointsThe result is also importantFacts。 The number of points in the calculation acts as a tool to obtain “the winning probability”, but it does not guarantee the existence of individual sThis is an example of 549 tests.
6. Incident Analysis (2): Structural Prediction - What AlphaFold Changes and Not Changes
Docomo3D structure of target proteinSo far, it was only decided by experiments such as X-ray crystal analysis. Jumper et al is a paper of 2021.Facts。
“A huge experimental effort has determined the structure of approximately 100,000 unique proteins, but this is just a few of the known billions of protein sequences.”Facts
AlphaFold is an international Blind Evaluation CASP14 for structural prediction, centralized view of main chain accuracy0.96 Å(Evaluated by 95% of Cα αs, root deviation's and residue's), the following method is2.8 ÅCommentFacts。
2024 Nobel Chemical Award to David Baker"Protein design by calculation"And the remaining half is demis Hasabis and John JumperProtein structure predictionAwarded inFacts。
Presentation by AlphaFold2“Almost all 200 million proteins identified by researchers”structure can be predicted,190+ countriesIt says that it is usedFacts(It is written by the same presentation).
2024 AlphaFold 3 not only proteinComplexes including nucleic acid, low s, ions and modifier residuesYou can now predict the structure at the same time.Much higher accuracy than cutting-edge docking tools for protein and low interactionindicatesFactsHome On the other hand, the same paper is clearly written.
“The key limitations of protein structure prediction models are to predict the structure of biological、s in the solution, not dynamic swelling, as seen in PDB.”Home AlphaFold 3Facts。
For three-dimensional chemistry, you can provide the correct rigid reference structure as input.Output may not protect theeven if you assign penalties to rankings4.4% offline rate during pregnancyIt says that it was observedFactsHome Because it is a method of generation type,may occurFacts。
Meaning for drug discoverystructure predictionThe wall of the entrance that "I can not start calculation because there is no structure"HomeMovements and behaviors in water that affect whether the drug is effectiveis still outside the prediction (in this article).
7. Machine Learning and Generation AI – Learning data is much smaller than the search space
(1) Line up to "Ef ive Order" in the study model
kesto suppress the growth of coliCreated a deep neural network to predict new antibacterial substancesFactsHome Flow is 3 stages.1 Learn with experimental data → Try to predict 2 huge libraries → Check the top 3。
| Stage | Paper Description |
|---|---|
| Learning Data | FDA approved drugs and natural products2,335 species(After duplication). Hit the one that suppresses the growth of more than 80%,120 species(5.14%)Hits |
| Predictions | Drug Repur Hub6,111 speciesWuXi Antituber osis Library9,997 speciesZINC15107,349,233 species |
| Time | Approx. 1,735 million predictions4 daysdone |
| Experiment | Predictions from ZINC1523 types of antibacterial activityHome A known antibiotic was structurally apart |
| Examples of results | Drug Repur HubHaricinshowed effect on mouse infection model |
AllFacts(Stokes et al 2020). Halicin is a compound in the study phase and this article does not evaluate its ethical meaning.
(2) Create 'molecules themselves' with generated AI
the structures that meet the conditionsis generated AI. Zhavoronkov et al in 2019, deep generation model GENTRL」d the ease of」, novelty, and biological activityHome kinase 1 in tor21 daysI found it.4 compounds are active in biochemical assays.and2 cells confirm with cell assayand one compound indicates a good drug content in the mouseFacts。
21 daysPeriod of design stages reported by authorsThis article does not indicate the duration of the evaluation and development.
8. Automation of Experiments: “Robot Scientists”
It is an experiment to check even if the candidate is narrowed down by calculation. HomePlan, execute, and analyze experiments into one loopThere is an attempt. Williams et al 2015, robot scientist「Eve」ReportedFacts。
“Robot scientists are laboratory automation systems that use artificial intelligence (AI) methods to discover scientific knowledge through experimental cycles.”Home EveScreening of libraries, confirming hits, and creating leadsrate and automatecycle of quantitative structural activity correlation (QSAR) learning and testingTurnFacts。
EquipmentOver 10,000 compounds per dayIt is possible to change the structure so that it can be used for reading, light absorption, and cell form in mid-high throughput. About Us14,400 compoundsthe library of multiple parasite enzymes (DHFR)Anticancer compound TNP-470 is a strong in tor of DHFR of the three-day thermal malaria wormI'm reporting that I foundFacts。
authors use economic models,The method of choosing compounds with AI is more economical than standard total screeningindicatesFacts(It is evaluation by the model of the author).
"Excellent economic" here isInstead of measuring all compounds, select the next compound while learningThis article explains how to reduce the number of measurements. Chapter 3 qHTS"Carefully calculate all "Eve is the evolution of the direction"Sensify order"The evolution of the direction. both are not confronted,Quality of measurement determines the quality of the entire loopIt is common in terms.
9. An example of progressing to the clinical stage (one of the first information)
The announcement that the drug found in AI has progressed to clinical trials has been increasing.Reviewed papers and those that have confirmed the stage with the ClinicalTrials.gov
| Search | Involvement of AI | Clinical stage (primary information) |
|---|---|---|
| rentosertib (old name 001-055/INS018 055) Target: TNIK | Select TNIK as the target for stabilization with the AI target search base (PandaOmics) and design in tors with the AI design base (Chemistry42). 42 is30 generation modelsto be used in | Phase 1Health78 peoplerandomized double- d placebo controlled trial (NCT05154240, completed) Phase 2: Patients with special pulmonary conditions71 people21 facilities in China, 12 weeks, placebo control (NCT05938920, complete) |
AllFacts(Ren et al 2024 [Reference Material 16], Xu et al 2025 [Reference Material 17], ClinicalTrials.gov [Reference Material 18-19]). At the time of investigation of this article,The primary information indicating that this compound has been approved by the regulatory authorities was not confirmedUndetermined / Future。
Design:Adult of special pulmonary fibrosis, rentosertib 30 mg once a day (18 people), 30 mg twice a day (18 people), 60 mg once a day (18 people), randomly assigned to four groups of placebo (17 people), 12 weeks.Facts。
Key Evaluation Items:Patients with at least one adverse event after clinical trial administration (evaluation of safety and tolerance),Like each groupand reportedFactsHome Changes in lung function (Effort pulmonary active volume)Sub-order Evaluation ItemsReported asFacts。
Author's own retention:Shorter test periods, in all groups16 out of 71 (22.5Larger and Longer Phase 2 or 3 test requiredFacts。
This article does not evaluate the effectiveness or safety of this test.
Ren et al. is considered to have completed "approximately 18 months" from the detection of targets to the no、 of pre、ical candidatesFactsHome This isReport by researchers of companies developedIt is not verified in line with the conventional development of the same condition that is compared (in this article).
Jayatunga et al. analyzes the clinical pipeline of biotech companies that focus on AI.The success rate of the first phase of the AI found in AI is 80-90%, and the second phase is about 40% (the sample number is limited).ReportFactsHome The authors themselves specify the small size of the specimen for the second phase,It is not a stage where AI can conclude that clinical success rate has increasedUndetermined / Future。
10. Perspective of Material Technician 2: Candidates of AI can also be obtained with "not melting"
Chapter 9: The paper of rentosertib has a section that is easy to overlook and is important. The first lead compounds that were designed and synthesized in the generated AINanomorphic bond affinityHome
In vitro absorption, distribution, metabolic and excretion (ADME) profiling of primary lead compounds, high clearance in human and mouse liver microsomes, and inhibition of cytochrome P450 (IC).50 (less than 10 μM), the velocity is less than 2 μMFacts
HomePrioritize ADME improvements during lead optimizationand the result was a candidate compound.Facts。
Melt level less than 2 μMThe number is in the same digit as in Chapter 4 "5 μM even 1 to 2% of s become cohesive"FactsHome In other wordsEven if AI finds a “well-connected” form, how much the、 melts into water and how fast it disappears with、remains as another problem, I've found itMeasurement of hepatic microsomes and dissolution levelsComment
Materials with material development"The function of the purpose (combination) is calculated, but the workability and stability (melt and metabolic) is measured"It is a composition. Materials InformaticsIt is difficult to simultaneously optimize secondary physical properties rather than predicting primary functionsThere are many scenes that you feel. It is possible to read from the primary information that there is the same wall in the drug discovery. ADME's test method is handled in detail in this series "pharmacokinetics and previous clinical trials".
11. How regulators are doing AI
On January 6, 2025, the FDA demonstrates the idea of using AI in the development of pharmaceuticals and biologicsDraft guidanceAnnouncement This is"The first guidance of the authorities on the use of AI in the development of pharmaceuticals and biologics"FDAMore than 500 applications including AI elements since 2016I have reviewedFacts。
Target:When creating information and data that support regulatory decisions on safety, effectiveness, and quality with AI modelsFacts。
Not applicable:“This does not handle the use of AI models in (1) drug discovery, or (2) work efficiency that does not affect the safety of patients, quality of drugs, and reliability of non-clinical and clinical trial results.”Facts
Frame:Risk7 stepsReliability Assessment (one definition of one question → Definition of the context of two use (COU) → Evaluation of 3 model risk → Planning of 4 reliability → solution → results of 6 results → Judgment of validity against 7 COU). Model riskModel risk of "the influence of the model" and "the weight of the return when the judgment is incorrect"Contact UsFacts。
Example:Injectables for multiple dose vials,Using AI image analysis for 100% automatic appearance inspection of filling volumeExampleFacts。
FDA and FDA jointly on January 14, 2026"Guidance Principles for Good AI Practice in Drug Development"as10 PrinciplesAnnouncement The target is the generation and analysis of evidence at each stage of non-clinical, clinical, post-marketing, and manufacturing.Data sources, processing, and analysis decisions in a trackable manner(Principle 6)Corres ing to data drift with regular monitoring and re-evaluation(Principle 9)FactsHome AIRe animal testing dependencies by improving human toxicity and effectiveness predictionIt also shows expectationsFacts。
danceAI in the drug discovery stage is not coveredData obtained by non-clinical, clinical, and manufacturing subsequentlyComment This article explains. Candidates selected by AIThe last is evaluated in the same testHome So the search phase AI is not regulation“Is it possible to get a candidate through a later test?”is evaluated.
As an example of manufacturingAutomatic inspection of filling volumeLikeIn the situation where AI enters the quality judgment itself, it is necessary to verify the context and risks of useFactsHome For pharmaceutical professionals who are in charge of testing equipment and measuring equipment for pharmaceutical manufacturing,De what AI is used in the judgment firstThe idea is how to write design specifications as it is (in this article). The GMP framework is handled by this series "GMP".
12. Strengths and Limits
| Technology | Strengths confirmed by primary information | Limits found in primary information |
|---|---|---|
| qHTS | Concentration reaction curves of all compounds are added, avoiding single-concentration HTSTSs | False positives due to degradation of the library, coagulation, and compounds |
| Docomo | Able to evaluate un s at a scale of 100 million. The higher the number of points, the higher the hit rate | The number of points stays immediately after the probability, and the number of points or less is the hit rate 0 |
| Structure Prediction | ulation can be started even if there is no experimental structure. Predicting complexes | Phantom structure, malfunction of three-dimensional chemistry, hallucination of disorder area |
| Learning and Generation Models | You can create a space of tens of thousands of learning data in a short time | Learning data is small and depends on the quality of the experiment. ADME and melting rate are required separately |
| Decide the order to measure and reduce the number of | Flexibility of device side is required for changing the type of readout and target |
Each item is based on the description of the primary information [Reference Material 3 - 17] quoted in Chapter 3 - 10Facts。This article is divided into five technologies.Comment
(1) I don’t know whether the success rate of clinical is higher with AI
As seen in Chapter 9, the clinical stages that can be confirmed by reader-reviewed papers and test registration are still limited,I couldn't confirm the example of approval at the time of the survey.Home AI AI improves overall development success rate and durationMatters to be determined by future dataHomeUndetermined / Future。
(2) The number of "sh effect" of development cost and period is not stated
The company’s presentations include shortening periods and costs.I could not check primary information with comparison control of the same conditionSorry, this entry is only available in Japanese. Ren et al's "approximately 18 months" is quoted as his own report.
- qHTS measured 60,000 pieces in 30 hours.40% of good active compounds were spilled in one-point judgmentFacts
- Up to 19% of clinical trials are colloidal cohesive in 30 μM.The quality of the hit depends on the solutionFacts
- Docomo evaluated 33,800 million units in about 1.1 core seconds per。.It is about 250,000 yen.the analysis of this paper
- AlphaFold has lowered the walls that are not structured, but the limitations of movement, three-dimensional chemistry, and hallucinity are given by the paper itself.Facts
- Candidates of AI can also be found in less than 2 μM and high clearance.Lastly, I decided to measureFacts
- The FDA draft guidance does not cover the stage of drug discovery AI and is intended for use in regulatory decisions.Facts
13. Glossary
- Screening
- The process of targeting a large number of compounds and picking up active (Hit).
- HTS
- High throughput screening. Experimental method to measure a large number of wells in a robot.
- qHTS
- 。ntitative HTS. A method to measure all compounds in a concentration series and obtain a concentration reaction curve.
- Library
- A large number of compounds are arranged in a plate with a solution for screening.
- False-positive/false-negative
- In fact, there is an error/activity that is said to be hit without activity.
- Colloidal cohesive
- particles that can be self-s。。 in water. Non-specific inhibition of protein.
- PAINS
- A compound with an easy-to-interference partial structure that appears as a repeat hit in many measuring systems.
- 5 Laws
- The number of weights, Log P, and hydrogen bonds indicate an estimate of molecules prone to oral absorption.
- Structure of the drug and its pharmacological activity are described in Chapter 10.
- Analysis done in a calculator. Use with in vitro and in vivo.
- Docomo
- ulation of candidate s on the target binding site, and pointing to the alignment.
- make-on-demand
- A group of compounds enumerated from reactions and composition units as a premise to synthesize after receiving orders.
- AlphaFold
- AI model that predicts the three-dimensional structure of protein from amino acid sequences.
- AIrate AI
- Create a new structural framework that meets the requirements.
- QSAR
- Quantitative structural activity relationship. The relationship between the structure and activity of the is expressed by formulas and models.
- ADME
- Absorption, distribution, and excretion. The four processes representing the movement of drugs in the body.
- Use context (COU)
- Defining which role and scope of AI models are used for.
14. Reference Materials
- Scannell, J.W., Blanckley, A., Boldon, H., Warrington, B.「Diagnosing the decline in pharmaceutical R&D efficiency」Nature Reviews Drug Discovery 11(3), 191–200 (2012). doi:10.1038/nrd3681 — pubmed.ncbi.nlm.nih.gov
- Wong, C.H., Siah, K.W., Lo, A.W.「Estimation of clinical trial success rates and related parameters」Biostatistics 20(2), 273–286 (2019). doi:10.1093/biostatistics/kxx069 — pmc.ncbi.nlm.nih.gov
- Inglese「Quantitative high-throughput screening: A titration-based approach that efficiently identifies biological activities in large chemical libraries」PNAS 103(31), 11473–11478 (2006). doi:10.1073/pnas.0604348103 — pmc.ncbi.nlm.nih.gov
- Kozikowski, B.A.「The effect of room-temperature storage on the stability of compounds in DMSO」Journal of Biomolecular Screening 8(2), 205–209 (2003). doi:10.1177/1087057103252617 — pubmed.ncbi.nlm.nih.gov
- Feng, B.Y., Shoichet, B.K.「A detergent-based assay for the detection of promiscuous inhibitors」Nature Protocols 1(2), 550–553 (2006). doi:10.1038/nprot.2006.77 — pmc.ncbi.nlm.nih.gov
- Baell, J.B., Holloway, G.A.「New substructure filters for removal of pan assay interference compounds (PAINS) from screening libraries and for their exclusion in bioassays」Journal of Medicinal Chemistry 53(7), 2719–2740 (2010). doi:10.1021/jm901137j — pubmed.ncbi.nlm.nih.gov
- Lipinski, C.A., Lombardo, F., Dominy, B.W., Feeney, P.J.「Experimental and computational approaches to estimate solubility and permeability in drug discovery and development settings」Advanced Drug Delivery Reviews 46(1-3), 3–26 (2001). doi:10.1016/S0169-409X(00)00129-0 — pubmed.ncbi.nlm.nih.gov
- Ruddigkeit, L., van Deursen, R., Blum, L.C., Reymond, J.-L.「Enumeration of 166 billion organic small molecules in the chemical universe database GDB-17」Journal of Chemical Information and Modeling 52(11), 2864–2875 (2012). doi:10.1021/ci300415d — pubmed.ncbi.nlm.nih.gov
- Lyu「Ultra-large library docking for discovering new chemotypes」Nature 566, 224–229 (2019). doi:10.1038/s41586-019-0917-9 — pmc.ncbi.nlm.nih.gov
- Jumper「Highly accurate protein structure prediction with AlphaFold」Nature 596, 583–589 (2021). doi:10.1038/s41586-021-03819-2 — pmc.ncbi.nlm.nih.gov
- Abramson「Accurate structure prediction of biomolecular interactions with AlphaFold 3」Nature 630, 493–500 (2024). doi:10.1038/s41586-024-07487-w — pmc.ncbi.nlm.nih.gov
- The Royal Swedish Academy of Sciences"The Nobel Prize in 2024 — Press release" October 9, 2024 — nobelprize.org
- keskes, J.M.「A deep learning approach to antibiotic discovery」Cell 180(4), 688–702 (2020). doi:10.1016/j.cell.2020.01.021 — pmc.ncbi.nlm.nih.gov
- Zhavoronkov「Deep learning enables rapid identification of potent DDR1 kinase inhibitors」Nature Biotechnology 37(9), 1038–1040 (2019). doi:10.1038/s41587-019-0224-x — pubmed.ncbi.nlm.nih.gov
- Williams「Cheaper faster drug development validated by the repositioning of drugs against neglected tropical diseases」Journal of the Royal Society Interface 12(104), 20141289 (2015). doi:10.1098/rsif.2014.1289 — pmc.ncbi.nlm.nih.gov
- Ren「A small-molecule TNIK inhibitor targets fibrosis in preclinical and clinical models」Nature Biotechnology 43(1), 63–75 (2025, March 2024). doi:10.1038/s41587-024-02143-0 — pmc.ncbi.nlm.nih.gov
- Xu, Z. etc.「A generative AI-discovered TNIK inhibitor for idiopathic pulmonary fibrosis: a randomized phase 2a trial」Nature Medicine 31(8), 2602–2610 (2025). doi:10.1038/s41591-025-03743-2 — pmc.ncbi.nlm.nih.gov
- ClinicalTrials.gov「A Phase 1, Evaluate the Safety, Tolerability, and Pharmacokinetics of INS018_055 in Healthy Subjects」NCT05154240 — clinicaltrials.gov
- ClinicalTrials.gov「Study Evaluating INS018_055 Administered Orally to Subjects With Idiopathic Pulmonary Fibrosis (IPF)」NCT05938920 — clinicaltrials.gov
- Jayatunga, M.K.P., Ayers, M., Bruens, L., Jayanth, D., Meier, C.「How successful are AI-discovered drugs in clinical trials? A first analysis and emerging lessons」Drug Discovery Today 29(6), 104009 (2024). doi:10.1016/j.drudis.2024.104009 — pubmed.ncbi.nlm.nih.gov
- FDAFDA Proposes Framework to Advance Credibility of AI Models Used for Drug and Biological Product Development fda.gov
- FDA"Considerations for the Use of Artificial Intelligence to Support Regulatory Decision-Making for Drug and Biological Products — Draftancedance" January 2025 (PDF) — fda.gov
- EMA・FDA“Guiding principles of good AI practice in drug development” January 2026 (PDF) — ema.europa.eu
- European Medicines Agency“FDA and FDA set common for AI in medicine development” January 14, 2026 — ema.europa.eu
15. Response Table of Claim and Sources (Audit)
| Content | Home | |
|---|---|---|
| The R&D cost was reduced by half every nine years since 1950, and inflation was reduced to about one-eighth of 80. Re-diaing the cause in comparison with the advancement of science, technology, and management for 60 years | Scan et al (2012)[Reference 1]https://pubmed.ncbi.nlm.nih.gov/22378269/ | Facts |
| From January 2000 to October 2015, we analyzed 406,038 clinical trial data (over 21,143 compounds) and estimated that 13.8% of the development process will finally be approved | Wong et al (2019)[Reference 2]https://pmc.ncbi.nlm.nih.gov/articles/PMC6409418/ | Facts |
| The conventional HTS of a single concentration is suffering from frequent false positives and false negatives, and requires a large try. The concentration range of about 4 digits in at least 7 stages of di series is 640。M-10mM in most of the source concentration. 1,536 hole plate, 23。L pin transfer. Measurement of 60,793 compounds, 368, 565,248 wells in a。 30 hours. 5,480 compounds (9.0。 were classified as active. False-positive 30(2%), False-negative 845(40%), False-positive 5(1%), False-negative 1,602(。) for 6SD threshold. False negatives are rarely known by HTS. When Z′ was 0.87, a large cell-based assay reached0.8 of false positives | Inglese et al (2006)[Reference 3]https://pmc.ncbi.nlm.nih.gov/articles/PMC1518803/ | Facts |
| 7,200 compounds were stored at room temperature for 1 year as a DMSO solution of 20 mM, and the detection probability was 92% and 6 months at 83% and 52% in 1 year. | Kozikowski et al (2003)[Reference 4]https://pubmed.ncbi.nlm.nih.gov/12844442/ | Facts |
| Many low-affinity at the micromolar concentration self-aggregate and make colloidal cohesive in solution. It is a great proportion to HTS that aims to be less than 1% of hit rate by about 1 to 2% of 20% of 20-cMnM and about 1 to 2% of μM. 0.01% vol/vol Triton X-100 to observe 50-1,000 nm particles with DLS | Feng and Shoichet(2006)[Reference 5]https://pmc.ncbi.nlm.nih.gov/articles/PMC1544377/ | Facts |
| Reported partial structures that identify compounds (PAINS) that are repeatedly hit by many biochemical HTS, and said that they are more likely to be treated as promising starting points | Baell and Holloway (2010)[Reference Material 6]https://pubmed.ncbi.nlm.nih.gov/20131845/ | Facts |
| "5 Laws" (Hydrogen-binding base 5 or more, 10 or more receptors, 500 or more weight, and CLogP 5 or more) HTS leads tend to have lower molecular weight and higher turbidity than previous HTS leads | Lipinski et al (2001)[Reference Material 7]https://pubmed.ncbi.nlm.nih.gov/11259830/ | Facts |
| GDB-17 enumerated up to 17 C, N, O, S, and halogen)s of 1,664 billion (166.4 billion), which is a range of sizes that contain many pharmaceuticals and typical leads | Ruddigkeit et al (2012)[Reference 8]https://pubmed.ncbi.nlm.nih.gov/23088335/ | Facts |
| In the make-on-demand library of 130 reactions and 70,000 units of Enamine, the market is less than 3%. More than 9,900 million units were docked to the C and more than 135 million units were docked to the D4 receptor. D4 evaluated about 70 trillion complexes and required 43,563 core time (about 1.2 days with 1,500 cores). 44 of 51 in51C (86% succeeded in the test of 5 in ts (549 in D4) discovered 81 new skeletons and 30 were sub-micromoles. Hit rate is higher than 22 to 26%, the number of points is worse, and the number of points is lower, and the number of points is less than 0. | Lyu et al (2019)[Reference 9]https://pmc.ncbi.nlm.nih.gov/articles/PMC6383769/ | Facts |
| The structure of the unique protein determined by the experiment is roughly 100,000, and it is only part of the known array. The center value of the main chain accuracy 0.96 Å in CASP14, the next point was 2.8 Å | Jumper et al(2021)[Reference Materials 10]https://pmc.ncbi.nlm.nih.gov/articles/PMC8371605/ | Facts |
| AlphaFold 3 predicts complexes, including protein, nucleic acid, low, ion, and modifier residues, and coordinates directly with module modules, and shows much higher accuracy than cutting-edge docking tools with protein–low interactions. The limits that do not reproduce dynamic behavior in the solution without predicting only static structure, violation rate 4.4%, hallucinations in areas without order | Abramson et al (2024)[Reference Materials 11]https://pmc.ncbi.nlm.nih.gov/articles/PMC11168924/ | Facts |
| The 2024 Nobel Chemical Award was awarded half of the protein design by calculation to the baker, and the remaining half was awarded with the "protein structure prediction" to the husabi and jumper, and the announcement date October 9, 2024. Almost all of the structures of approximately 2 billion proteins can be predicted in AlphaFold2, and more than 2 million people in 190 countries | Nobel Foundation Press Release[Reference Material 12]https://www.nobelprize.org/prizes/chemistry/2024/press-release/ | Facts |
| 2,335 learning data (120 species and 5.14% hit), predictive targets are Drug Repur Hub 6,111 species, WuXi 9,997 species, ZINC15 to 107,349,233 species, and prediction in four days, 23 species were tested from ZINC15 prediction and 8 species were off structurally with known antibiotics with antibacterial activity, Halicin showed effect on mouse infection model | kes et (2020)[Reference 13]https://pmc.ncbi.nlm.nih.gov/articles/PMC8349178/ | Facts |
| GENTRL optimizes the susceptibility, novelty, and bioactivity of GEN1 in tors in 21 days, the GEN1 in tors were activated in biochemical assays, 2 compounds confirmed in cell assays, 1 compound showed good drug content in mice | Zhavoronkov et al (2019)[Reference 14]https://pubmed.ncbi.nlm.nih.gov/31477924/ | Facts |
| Defining robot scientists, Eve integrated library screening, Hit confirmation and lead creation, turning QSAR learning and testing cycles, over 10,000 compounds per day, reading, library of 14,400 compounds, and TNP-470 found to be a strong inhibitor of the three-day thermal malaria primary DHFR, showing that AI selection in economic models is more economical than standard screening | Williams et al (2015)[Reference Materials 15]https://pmc.ncbi.nlm.nih.gov/articles/PMC4345494/ | Facts |
| hemOmics targeted TNIK and designed in。tors with Chemistry42 (30 generation models). The primary lead showed the binding affinity of the nano-morweight, but the high clearance of human-mouse liver microsome, CYP inhibition (IC50 less than 10 μM), and the speed theory dissolution less than 2 μM, prioritized ADME improvement by lead optimization. The first phase (NCT05154240) is a randomized double-。d placebo controlled trial of 78 healthy adults. Explanation that completed in approximately 18 months from target discovery to pre ical candidate no | Ren et al(2024)[Reference 16]https://pmc.ncbi.nlm.nih.gov/articles/PMC11738990/ | Facts |
| The 2a phase test is a multicenter (China 21 facilities) double-blind randomized trial that was assigned to a special pulmonary fibroblast adult 71 30 mg once a day (18 people), 30 mg twice a day (18 people), 60 mg once a day (18 people), and placebo (18 people) for 12 weeks. | Xu et al (2025)[Reference Materials 17]https://pmc.ncbi.nlm.nih.gov/articles/PMC12353801/ | Facts |
| NCT05154240 is the first phase test for health adults of INS018 055, and the status is complete. | ClinicalTrials.gov[Reference Materials 18]https://clinicaltrials.gov/study/NCT05154240 | Facts |
| NCT05938920 is the second-phase trial for special pulmonary fibroblast, with 71 registered people (prog), and the situation is complete. | ClinicalTrials.gov[Reference Materials 19]https://clinicaltrials.gov/study/NCT05938920 | Facts |
| Reported the 1st phase success rate of AI to 80-90%, 2nd phase to 40% (sampled) | Jayatunga et al (2024) abstract[Reference Materials 20]https://pubmed.ncbi.nlm.nih.gov/38692505/ | Facts |
| The FDA announced draft guidance on January 6, 2025 that it is the first guidance on the use of AI in pharmaceutical and biologics development, and has reviewed over 500 applications including elements of AI since 2016. | FDA Press Release[Reference Materials 21]https://www.fda.gov/news-events/press-announcements/fda-proposes-framework-advance-credibility-ai-models-used-drug-and-biological-product-submissions | Facts |
| Not covered by drug development (when making data that supports regulatory decisions on safety, efficacy, and quality with AI) at 7 levels of reliability. The model risk is a combination of the influence and judgment of the model. Examples of clinical patient ification and production of 100% automatic inspection of multi-dose vials with AI image analysis. Must be drafted in January 2025 | FDA draft guidance (PDF)[Reference]https://www.fda.gov/media/184830/download | Facts |
| FDA and FDA’s 10 mentoring principles are subject to non-clinical, clinical, commercial, post-marketing, and manufacturing phases of evidence, in principle 6 (trackable documentation of data sources, processing and analysis decisions), in principle 9 (regular monitoring and re-evaluation, data drifting), and expectations for reducing animal testing dependencies by improving AI’s toxicity and effectiveness prediction | FDA and FDA ding principles (PDF)[Reference 23]https://www.ema.europa.eu/en/documents/other/guiding-principles-good-ai-practice-drug-development_en.pdf | Facts |
| FDA and FDA jointly announced 10 principles on January 14, 2026 | News[References 24]https://www.ema.europa.eu/en/news/ema-fda-set-common-principles-ai-medicine-development-0 | Facts |
| 565,248 ÷ 30 hours = about 18,800 WELL / h (5 WELL / sec), 565,248 ÷ 60,793 = about 9.3 WELL / compound. 43,563 core time ÷ 3,800,000 = approx. 1.1 core seconds /。, 549 ÷ 3,800,000 = approx. 150,000. 107,349,233 ÷ 2,335 = about 4.6 million times, 8 ÷ 23 = about 35%. Fig. 2 and Fig. 4 pole length (e.g., the number of regular pairs) | 。ulation of this article. It is a simple de ation of the published value of the paper, and it does not indicate the medium rate of the general speed and model of the actual efficacy and calculation of the device | the analysis of this paper |
| AI AI has improved clinical success rate and development time. whether rentosertib is approved | At the time of the investigation of this article (September 2026), primary information that indicates primary information and approval based on the comparison of the same condition was not confirmed. | Undetermined / Future |
| AI’s value of the time and cost of AI drug discovery companies, and the overall development of individual companies | I can not check the primary information with comparison control. Clinical stage examples are limited to reader-reviewed papers and clinicalTrials.gov | |
| Read and organize the library as "inventory material with expiration", "colloid dispersion". Read the relationship between threshold and spill. Read make-on-demand as a search space defined by the process side. Set the number of dots as the probability index. Infer the meaning of structure prediction. Position Eve as the direction of wise measurement. Concurrent optimization of dissolution and ADME. Reads that the regulation sees the scene where AI output is used. Chapter 12: Technologies | and commentary of this article based on the publication content. It is not the opinion of authors and regulatory authorities of each paper | |
| Evaluation of the efficacy and effect of each candidate compound, clinical results and safety | This article is an explanation of search technology and does not evaluate the treatment effect or safety. Phase 2 trials include design, key evaluation items, and author retention only | |
| Fig. 1 to Fig. 4 and Fig. 6 are drawings for explanation, not actual equipment, structure, data. Hero image and figure 5 are AI generation images | Notes by this article |
Last Updated: September 26, 2026/Source has been limited to primary information (review papers, Novel Foundation announcements, ClinicalTrials.gov, FDA/。 published documents). They are distinguished from the facts that have been sourced as "instructions" because they include the design reading of screening and calculations. We do not describe the development period and cost reduction effect by AI, improvement of clinical success rate, and approval of candidate compounds because we can not confirm primary information or approval of comparison-controlled primary information. FDA guidance is as of January 2025. This article is an explanation of the search technology.It does not evaluate the effectiveness and safety of the treatment, but it is not medical advice. All diagrams are illustrations for explanation. Figure 1, Figure 2, Figure 3, and Figure 5 are vector drawings, Figure 4 is an image, and Figure 6 is an AI-generated image.No actual equipment, data, or data.