This is a narrative review and not a systematic one. Literature from 2023 to October 2026, with earlier foundational papers where a method originates before that date, was identified through PubMed and the arXiv, bioRxiv and ChemRxiv preprint servers, and was supplemented by regulator and standards body websites and developer documentation for licences and release status. Search terms combined artificial intelligence, machine learning or large language model with the pipeline stages, open source, open weight and tool names. Tools were included when they were publicly available, maintained, documented in a peer reviewed paper, regulatory document or repository, and relevant to a task a pharmaceutical scientist performs, and each was classified as open source, open weight, free web service or closed using the criteria in Section Open and closed AI: a working taxonomy. Performance claims that come from developers or from unreviewed preprints are labelled as such. Clinical, regulatory and licensing statements are given as at October 2026 and were checked against trial registries, company disclosures, regulator publications and developer repositories, and a positive pivotal readout is distinguished from regulatory approval. The search was not recorded in a form that would allow it to be called systematic, and no such claim is made. The grades in Figure 2 and the verdicts in Table 1 are qualitative judgments by the authors and not empirical audits.

Discovery is where AI is most visibly transformative and where the open and closed worlds diverge most sharply. Four tasks dominate. First, target identification and validation increasingly use protein structure prediction: AlphaFold and its successors predict three dimensional structures and, in AlphaFold 3, protein interactions with ligands, nucleic acids and ions, a body of work recognized by the 2024 Nobel Prize in Chemistry, shared by Hassabis and Jumper for protein structure prediction and by Baker for computational protein design [21, 46]. The model is partly open [22], and independent open models such as OpenFold [28], Chai-1 [47] and Protenix [32], the last two so far described only in preprints, give the academic laboratory genuinely free options, while Isomorphic Labs commercializes a closed successor engine. A newer open-weight entrant, Boltz-2, jointly predicts complex structure and small molecule binding affinity under a permissive MIT licence and, according to its developers, is the first deep learning model to approach the accuracy of physics based free-energy perturbation calculations on the benchmarks they report, extending the open structure prediction ecosystem from pose to potency [48]. This is a developer reported claim from a preprint that has not yet been peer reviewed or independently replicated at scale. Upstream of structure, the open Open Targets Platform integrates human genetics, functional genomics and clinical evidence to score and rank target disease associations, and PharmGKB curates variant drug associations and clinical dosing guidelines, together giving academic groups a free route to target prioritization and pharmacogenomic risk assessment before a structure is ever modelled [40, 41]. Second, generative de novo design uses deep generative models to propose novel chemotypes; the field reached proof of concept when a model identified potent DDR1 inhibitors in 21 days [49], and open toolkits such as REINVENT and the transformer based MolGPT let academic groups run comparable workflows without licence cost. Third, virtual screening and molecular docking are accelerated by machine learned scoring, with open stacks built on RDKit [18] and DeepChem [19] and free docking engines such as AutoDock Vina, the SwissDock server and the automated CB-Dock2 cavity detection server competing against proprietary physics plus machine learning platforms; the chemical matter for these workflows comes overwhelmingly from the open PubChem, ChEMBL and DrugBank repositories, which catalogue structures, bioactivities, targets and interactions for essentially every known compound, while free reverse screening tools such as SwissTargetPrediction and the Similarity Ensemble Approach predict the likely targets of a structure and support drug repurposing. Diffusion based docking, exemplified by the open DiffDock, reframes pose prediction as a generative rather than a search problem and needs no predefined binding pocket, while the open Protein-Ligand Interaction Profiler (PLIP) automatically extracts and visualizes the hydrogen bonds, hydrophobic contacts and π-stacking that a docked or crystallographic pose implies [36, 37]. An independent evaluation with the PoseBusters test suite found that deep learning docking methods, DiffDock among them, did not yet outperform classical engines such as AutoDock Vina and the commercial CCDC GOLD in physical plausibility or in generalization to novel complexes [50]. Downstream of a proposed structure, the open ASKCOS suite combines retrosynthetic route prediction, reaction outcome prediction and condition recommendation in a single freely available tool, closing the loop from generative design to a plausible synthetic route [38]. Fourth, early property prediction (quantitative structure-activity relationships and ADMET) is now quite robust and largely open: ADMET-AI [27], built on Chemprop and trained on the Therapeutics Data Commons [33], was reported by its developers to rank highest on the Therapeutics Data Commons ADMET leaderboard at the time of its publication, a developer reported and date dependent result, and free web tools such as SwissADME [24] and ProTox [25] cover drug likeness and toxicity. The variable nearest neighbour vNN-ADMET server adds cytotoxicity, CYP450 inhibition, mutagenicity and liver injury endpoints from a distance weighted similarity model [39]. The integrated commercial platforms (Insilico, Recursion, which merged with Exscientia in 2024, and Schrödinger) bundle these steps end to end and have produced the clinical candidates discussed in Section Clinical research [16].

FormulationAI is a web based AI platform that supports in silico pharmaceutical formulation design by predicting key formulation properties from basic drug and excipient information. This enables rapid identification of promising formulations while reducing reliance on trial and error experimentation, shortening development timelines, and lowering costs [45]. Complementing this approach, FormulationDE applies machine learning to predict drug excipient compatibility and identify potential interaction risks based on molecular functional groups. By providing rapid, interpretable predictions, it streamlines pre-formulation studies and reduces the need for labour-intensive compatibility testing [44]. As web services, these tools carry the confidentiality caution of Section Open and closed AI: a working taxonomy.

AI is also transforming pharmacokinetic (PK) prediction. Integrated artificial intelligence physiologically based pharmacokinetic (AI–PBPK) platforms can predict in vivo drug disposition directly from molecular structures by estimating key physicochemical and ADME properties, including solubility, permeability, protein binding, plasma clearance, and tissue partitioning. These parameters are incorporated into PBPK models to forecast PK profiles, with validation against more than 670 human intravenous and oral datasets demonstrating robust predictive accuracy. By accurately predicting tissue distribution and organ selectivity, AI–PBPK models can accelerate candidate selection and prioritize compounds with favourable PK characteristics, reducing dependence on extensive in vitro and in vivo studies [43]. The formulation and AI–PBPK platforms described in this section are free web services, so proprietary structures should not be uploaded to them.

Beyond computational prediction, AI is beginning to automate experimental chemistry. In 2026, OpenAI and Molecule.one reported the first publicly documented use of a frontier language model as a near-autonomous agent in a real medicinal chemistry workflow. Coupled to an agentic system and a high-throughput robotic laboratory, GPT-5.4 identified TEMPO as a mild oxidant capable of improving the yield of the challenging Chan–Lam coupling reaction for primary sulfonamides. In two microscale screening campaigns comprising 10,080 reactions, the optimized TEMPO condition raised the mean estimated product yield from 16.6% to 25.2% and increased the fraction of reactions exceeding a 30% yield from 15.6% to 37.5%, and bench scale experiments confirmed the benefit in 11 of 14 representative substrate pairs [17]. The screening yields were estimated by liquid chromatography with absorbance correction and not isolated, the mechanism remains unresolved, and the result applies to one reaction class on one platform. Although the work is a preprint authored by the developers that has not been peer reviewed, and required human oversight for experimental design and validation, it demonstrates how AI can extend beyond in silico prediction to assist with autonomous laboratory experimentation while maintaining scientific accountability [17].

A sixth task, less often itemized but central to phytomedicine and natural product research, is the systematic mining of curated databases of bioactive natural products for candidate scaffolds. The open option is COCONUT (COlleCtion of Open NatUral prodUcTs), a free, login free web resource whose 2024 overhaul aggregates more than sixty natural product collections and supports name, structure, substructure and similarity searching, bulk download and a REST API, with precomputed natural product likeness, synthetic accessibility and drug likeness scores to help prioritize compounds [34]. Machine learning has been built directly on it: a support vector machine trained on COCONUT structures assigned biological origin (plant, fungal or bacterial) at 89 to 94 percent accuracy, illustrating how an open database also serves as a substrate for predictive models [51].

The related LOTUS initiative, disseminated through Wikidata, adds more than 750,000 referenced structure organism pairs and is the natural companion for ethnopharmacological and traditional medicine work [35]. The proprietary comparators are the Dictionary of Natural Products and the natural product content of subscription chemistry databases, so the open versus closed contrast of this review extends to natural product space as well. Downstream of scaffold selection, the multi-target mechanisms of natural products and repurposed drugs are increasingly dissected by network pharmacology, for which the open STRING protein interaction database, the Cytoscape network visualization platform and the MetaboAnalyst metabolomics server form a free and widely used toolchain.

In pre-clinical work AI supplements rather than replaces wet lab toxicology and pharmacokinetics. In silico toxicity and ADMET models triage liabilities before synthesis [25, 27, 52]; computer vision reads histopathology and high content imaging, a use that is mature in this stage; and physiologically based pharmacokinetic (PBPK) modelling supports first in human dose projection. PBPK is the clearest example of an open option matching a commercial one: the Open Systems Pharmacology suite (PK-Sim and MoBi) is free and open source [29], standing alongside the proprietary Simcyp and GastroPlus that are standard in industry. Machine learned synthesis planning and formulation optimization round out the stage. Reviews of AI for pharmacokinetics [52] and for pharmacology practice [53] document the breadth of use.

Clinical development uses AI for trial design and simulation, patient recruitment and matching, digital biomarkers, population pharmacokinetic-pharmacodynamic (PopPK-PD) modelling, and prediction of drop out and site risk. The pharmacometric workhorses here are largely closed (NONMEM, Monolix, Phoenix WinNonlin), although open alternatives in R (nlme, the Stan ecosystem) and Julia (the Pumas project) are growing. The headline development of the review period is clinical: Insilico Medicine’s rentosertib (ISM001-055), a TNIK inhibitor in which both the target and the molecule were generated by AI, reported positive phase IIa results in idiopathic pulmonary fibrosis and entered a phase III program in July 2026 (NCT07687459) [12, 16], while Exscientia’s DSP-1181 was the first AI-designed molecule to enter human trials, it was discontinued after phase I, which is a reminder that accelerated discovery does not guarantee clinical fitness, and the physics based computational design platform of Schrödinger, applied at Nimbus Therapeutics, produced zasocitinib (TAK-279), a TYK2 inhibitor that reported positive pivotal phase III results in plaque psoriasis in December 2025 and superiority over deucravacitinib in a head to head phase III study in June 2026 [16, 54, 55]. These results validate the discovery platforms but remain early, and whether any AI-originated drug has completed phase III depends on the definition used. We use AI-originated for a drug whose target and molecule were both identified by AI, as for rentosertib, and computationally designed for a molecule optimized with physics based or machine learning tools against a conventionally selected target, as for zasocitinib. On that definition no AI-originated drug has yet completed phase III or reached the market, whereas a computationally designed drug has reported pivotal phase III results.

AI is entering the regulatory stage from two directions. Sponsors use natural language processing and LLMs to draft and quality check dossiers, synthesize evidence and review labelling, and use analytics for manufacturing and for chemistry, manufacturing and controls (CMC) (Table 5). Regulators, in turn, have begun to govern AI used to support decisions. In January 2025 the US Food and Drug Administration issued its first dedicated draft guidance, Considerations for the Use of Artificial Intelligence to Support Regulatory Decision-Making for Drug and Biological Products, proposing a seven step, risk based credibility assessment framework anchored on the model’s context of use [8]. The guidance reflects the agency’s experience with more than 500 drug and biologic submissions containing AI components since 2016 and has itself been the subject of critical appraisal [56]. As far as we could determine, the FDA still listed this document as a non-binding draft guidance in September 2026 [10]. For the pharmaceutical scientist, the practical message is that an AI result intended to support a regulatory claim must carry documented evidence of its credibility for the specific question it addresses. Parallel activity is under way internationally: the European Medicines Agency adopted a reflection paper on AI across the medicinal product lifecycle in September 2024 [9]; the FDA and EMA jointly published ten guiding principles of good AI practice in drug development in January 2026, stressing human centric design, a risk based approach and a clear context of use [10]; ICH adopted its M15 guideline on model informed drug development in January 2026, with a framework of model influence, risk and impact [11]; the MHRA and Health Canada have, with the FDA, identified good machine learning practice principles, and Health Canada issued pre market guidance for machine learning enabled medical devices in April 2026, although these address devices and not medicines directly [57, 58]; and the European Union Artificial Intelligence Act (Regulation (EU) 2024/1689) adds a horizontal legal layer applicable to many AI systems used in drug development, so the United States focus of this section reflects the most developed example rather than the only one.

After approval, AI is used to recover safety and effectiveness signals from data at a scale beyond manual review. Spontaneous reporting databases (the FDA Adverse Event Reporting System, the World Health Organization VigiBase, and the European Medicines Agency EudraVigilance) receive millions of reports annually, and machine learning and natural language processing now augment classical disproportionality analysis for signal detection [59–61]. Open building blocks include the openFDA application programming interfaces, the spaCy and scispaCy text-processing libraries, and MedDRA-aware tooling; closed safety platforms (for example Oracle Argus and ArisGlobal LifeSphere) embed AI for case processing in industry. Generative models are being explored for narrative summarization and causality assessment [62], and narrative reviews emphasize the need for explainability and human oversight in this safety critical setting [63].

The open versus closed contrast extends to biologics, although this review concentrates on small molecules. Protein language models such as the ESM family [64], diffusion based backbone generators such as RFdiffusion [65] and inverse folding tools such as ProteinMPNN [66] are released openly, with licence terms that vary by model, whereas antibody design is served by both academic releases and proprietary platforms whose comparative evidence is mostly developer reported. Predicted structures and sequences still require experimental expression, binding and developability testing, and immunogenicity cannot be predicted reliably from sequence. Maturity and evidence differ substantially between protein design, antibody engineering and manufacturing analytics (Section Manufacturing, chemistry, manufacturing and controls, and process analytical technology).

Manufacturing and CMC are where AI is least visible in the open literature and where closed, validated systems dominate, because the data are proprietary and the systems operate under good manufacturing practice. Machine learning predicts critical quality attributes from composition and process data and underpins soft sensors and multivariate models on process analytical technology (PAT) signals, including in continuous manufacturing, which ICH Q13 addresses [67]. The FDA has sought comment on integrating AI into its manufacturing framework [68], and the joint FDA and EMA principles span manufacturing as well as development [10]. A model that releases product or controls a process is a high risk context of use whose lifecycle, change control and oversight must be qualified within the sponsor quality system.

General statistics, regression, survival analysis and machine learning are well served by free, open and reproducible environments: R (with packages such as nlme, survival and tidymodels), Python (scikit-learn, statsmodels, lifelines), and the point and click JASP and jamovi for teaching. These are open source and citable, which clearly matters for method transparency in a manuscript. Conversational assistants are useful as statistical tutors and code generators but should not be trusted to perform or report analyses without verification, a caution echoed across the AI in pharmacology literature [53, 72].

For non-compartmental analysis, compartmental PK-PD and population modelling, the open options include R (nlme, nlmixr2, mrgsolve), the Julia-based Pumas project, and, for mechanistic PBPK, the open-source PK-Sim and MoBi suite [29]. The commercial standards remain Phoenix WinNonlin, NONMEM, Monolix, Simcyp and GastroPlus [72]. Machine learning is increasingly used to predict PK parameters directly from structure [52]; this is an area where a free model can give a useful first estimate that a validated tool later confirms. Free chatbots have also been evaluated as substitutes for specialized pharmacokinetic software in teaching: prompt-engineered free versions of ChatGPT and Gemini produced accurate, reproducible patient only and Bayesian estimates of volume of distribution and elimination rate constant, making Bayesian analysis more accessible to educators and students, though the results clearly depended on prompt design and required verification [73].

Fitting calibration and standard curves (linear, weighted linear and four-parameter logistic) needs nothing proprietary. R (the drc package), Python (SciPy and lmfit) and open spreadsheet tools fit these models, report weighting and goodness of fit, and can be scripted for reproducibility. An LLM is helpful for generating the fitting script and for explaining why a weighting scheme is appropriate, provided the numerical result is checked against the open library.

Drawing structures, generating and parsing SMILES, canonicalizing and converting between formats, and computing descriptors are core open-source strengths. RDKit [18] handles SMILES parsing, canonicalization, two-dimensional depiction and descriptor calculation programmatically; Open Babel converts between formats; and free editors such as the Ketcher and JSME web sketchers produce SMILES for teaching and manuscripts. LLMs can read and generate SMILES and explain notation, but they make silent errors on stereochemistry and valence, so generated structures should always be validated in RDKit before use.

Prediction of logP, solubility, pKa, permeability and the wider ADMET panel is now a free-tool strength. SwissADME [24] gives drug-likeness and pharmacokinetic summaries from a SMILES string in seconds; ADMET-AI [27] provides fast, accurate, batch predictions with an open Python package; ProTox [25] predicts toxicity endpoints; pkCSM [42], from the University of Cambridge and the University of Melbourne, uses graph based signature machine learning models to predict ADMET properties, including CYP substrates and inhibitors, from a SMILES string or a two dimensional structure; BioTransformer 3.0 [74] combines machine learning with knowledge based rules to predict mammalian, CYP450 and phase II metabolism and to generate the structures of the predicted metabolites from a SMILES string or an SDF file; and the Therapeutics Data Commons [33] supplies the datasets and benchmarks. Commercial predictors (ADMET Predictor, the Schrödinger suite) add breadth and support. In our assessment the free tools are adequate for most teaching, ranking and early stage decisions, although independent head to head comparisons with commercial predictors are limited (Section Accuracy versus accessibility: where the performance gap lies). SwissADME, ProTox, pkCSM, BioTransformer 3.0 and the public ADMET-AI server are web services that receive the submitted structure, so proprietary or unpublished structures should be predicted with a locally installed package such as ADMET-AI. pkCSM is free for academic use, whereas non-academic use requires a server licence.

Research has little effect until it is communicated, and much of a pharmaceutical scientist’s working life is spent turning results into figures, slides, posters, abstracts, manuscripts and regulatory documents. AI now assists at every step of that dissemination: generating and refining figures, images and presentation slides, checking grammar and references, drafting and restructuring text, and helping to assemble and format formal submissions such as investigational new drug applications (Table 5). The gain is in speed and polish rather than in substance, and the responsibility for accuracy stays with the author. LLMs are now embedded in writing, summarization, coding and tutoring. The pragmatic division is that closed assistants (ChatGPT, Claude, Gemini) in our assessment currently lead on reasoning and polish, while open-weight models (Llama, Mistral, Qwen) can be run locally when confidentiality forbids sending text to a vendor. In teaching, these tools generate worked examples, quiz items, case vignettes and code, and they lower the barrier to the computational methods above. Dedicated literature search assistants such as Elicit, Consensus and the Semantic Scholar research tools apply similar language model techniques specifically to systematic search, evidence extraction and citation graph navigation, and are increasingly used alongside general purpose chatbots when scoping a review. Two cautions, consistent with the original editorial and with the pharmacology literature, apply throughout: outputs must be verified because models fabricate plausible references and results, and use must be disclosed in line with journal and institutional policy [1, 53]. A related development is the emergence of AI enabled electronic laboratory notebooks (for example Benchling AI and eLabNext AI), which summarize experiments, extract and structure protocols and build searchable knowledge graphs, extending this assistance from the manuscript to the day to day laboratory record and broadening its industrial relevance.

Biopharmaceutics adds a specific and high value case to the property prediction of Section Physicochemical and ADMET property prediction. The Biopharmaceutics Classification System (BCS) sorts an oral drug into one of four classes by aqueous solubility and intestinal permeability [75], and a BCS assignment can support a biowaiver of in vivo bioequivalence studies under the harmonized ICH M9 guideline for class I and class III drugs [76]. The two discriminating measurements, solubility and permeability, are precisely the properties that machine learning now estimates quickly from structure. Free web tools give a usable first pass: SwissADME reports lipophilicity, aqueous solubility and a gastrointestinal absorption flag from a SMILES string [24], and ADMET-AI returns calibrated aqueous solubility and Caco-2 permeability among its endpoints [27]. A dedicated open platform, FormulationBCS, curates log S, log P, log D and apparent permeability datasets and predicts BCS class directly, showing that in silico provisional classification is now feasible at high throughput for early developability and formulation decisions [71]. The practical division is the same as elsewhere in this review: a free model is well suited to ranking candidates, triaging a library or teaching the BCS, whereas a regulatory biowaiver still rests on the experimental solubility and permeability determinations specified by ICH M9, for which a predicted class is a hypothesis to be confirmed rather than evidence in itself [76]. FormulationAI and FormulationDE, described in Section Discovery and development, extend this approach to formulation design and drug excipient compatibility [44, 45].

Two model-based methods extend the previous subsections from prediction toward evidence, and AI assists in both. The first is virtual bioequivalence (VBE). Where Section Biopharmaceutics: solubility, permeability and BCS classification and formulation development used machine learning to assign a BCS class, VBE goes a step further and simulates the bioequivalence study itself: a physiologically based biopharmaceutics model (PBBM), built on the same physiology as PBPK, generates a virtual population, propagates formulation and physiological variability, and predicts whether a test and a reference product would fall within bioequivalence limits [77]. Regulators including the FDA, EMA, MHRA, Health Canada, ANVISA and PMDA now actively discuss VBE for formulation bridging, dissolution safe space justification and selected biowaivers [77], and the open PK-Sim and MoBi suite [29] offers a free entry point to the underlying PBPK, although validated commercial platforms such as Simcyp and GastroPlus remain standard for regulatory submissions. As with the BCS biowaiver of Section Biopharmaceutics: solubility, permeability and BCS classification and formulation development, revisited in Section Software of unknown provenance (SOUP) and the governance of AI in biopharmaceutics, a VBE result intended to support a regulatory claim is a high-stakes context of use whose model must be qualified and documented [8, 77].

The second method moves to the clinical trial. An AI digital twin is a model, trained on historical patient data, that forecasts a participant’s likely trajectory under control or standard of care from their baseline characteristics alone [78]. Used as a prognostic covariate rather than as a synthetic control, the digital twin preserves randomization while increasing statistical power, which can in turn permit a smaller control arm. This approach, prognostic covariate adjustment, received a qualification opinion from the European Medicines Agency in 2022 and underlies the commercial platform already listed for clinical development in Table 4 (Unlearn) [79]. Digital twins make the same point as the rest of this review: the value lies not in replacing the experiment but in using a well governed model, with a documented context of use, to make the experiment more efficient while a human remains accountable for the inference [8].

Predicting human pharmacokinetics and a first in human dose from preclinical data is one of the few areas of drug development where the free ecosystem is unusually deep, and it spans both mechanistic and machine learning approaches. Classical interspecies allometric scaling, the log-linear regression of clearance and volume of distribution on body weight, needs nothing proprietary and runs in R or Python. Mechanistic physiologically based extrapolation is served by the open-source Open Systems Pharmacology suite, whose PK-Sim and MoBi components carry species specific physiological databases, support cross species scaling and first in human projection, and are used across academia, industry and regulatory science [29]; a freely reusable library of qualified PBPK models shortens model development for new compounds. AI–PBPK platforms, described in Section Discovery and development, predict profiles directly from structure [43]. A purpose-built free tool, DallphinAtoM (Drugs with Allometry and Physiology Inside, Animal to huMan), integrates published PBPK and allometric methods in a transparent graphical interface and predicts human PK parameters and concentration-time profiles from physicochemical properties, in vitro ADME data and animal PK data, with the explicit aim of letting the user move between allometric scaling and PBPK as the available data allow [70]. The machine learning option is PKSmart, which predicts human clearance, volume of distribution and half-life directly from molecular structure together with predicted animal PK, with accuracy that its developers report as comparable to industry standard approaches, in a peer reviewed but developer reported evaluation while remaining open-source and freely web-hosted [69]; structure based ADMET predictors such as ADMET-AI return related estimates [27]. The proprietary comparators are the validated suites of Simcyp, GastroPlus and Phoenix WinNonlin, whose advantage lies in regulatory acceptance and support rather than in availability. The division is the familiar one: these free tools are well suited to ranking candidates, projecting an initial human dose and teaching the methods, whereas a dose that enters a regulatory submission still rests on a qualified model and experimental confirmation.

If a reader adopts only a handful of free tools this year, the starter kit in Table 6 spans the recurring tasks of Sections Statistics and data analysis, Pharmacokinetic and pharmacodynamic modelling, Calibration and standard curves, Structure drawing, SMILES and chemical notation, Physicochemical and ADMET property prediction, Scientific writing, literature work and teaching, Biopharmaceutics: solubility, permeability and BCS classification and formulation development, In silico bioequivalence and digital twins, Allometric scaling and human pharmacokinetic prediction, and every item can be cited by version in a manuscript. Statistics and scripting rest on R, Python and the point and click JASP and jamovi; pharmacokinetics on the R packages nlme, nlmixr2 and mrgsolve and on the open source PK-Sim and MoBi suite [29]; chemistry on RDKit [18] with the Ketcher and JSME sketchers; and reproducible pipelines on Jupyter and KNIME. For physicochemical, ADMET and biopharmaceutics screening, SwissADME [24], ADMET-AI [27], ProTox [25] and FormulationBCS [71] return estimates in seconds, but they are web services unless installed locally, and proprietary structures should not be submitted to them. For protein structure, the public AlphaFold server and the open reproductions OpenFold, Chai-1 and Protenix [22, 28, 32, 47] avoid licence costs, although the server is restricted to non-commercial use and receives the submitted sequence. For confidential writing and coding, a locally run open weight model keeps text on local hardware. Recording the exact version used is the reproducibility discipline most worth teaching a trainee. The full list of tools with access addresses is provided as Supplementary Table S1.

To see how these free tools combine in practice, consider a graduate student building a first physiologically based pharmacokinetic model. The student might use a conversational assistant to draft and annotate initial R code, SwissADME or ADMET-AI to obtain first-pass physicochemical and permeability estimates from the candidate structure, PK-Sim and MoBi to run the mechanistic simulations, and an established pharmacometric platform to confirm the completed workflow before any result is reported, checking each tool against the next as the analysis proceeds. The pipeline is assembled entirely from freely available components, yet every step remains the responsibility of the modeller rather than of any single program.

The reproducibility and opacity concerns above have a concrete regulatory analogue that pharmaceutical scientists will increasingly meet by name. In medical device software, any component that is already developed and generally available, and that was not produced under a controlled development lifecycle, is termed software of unknown provenance (SOUP) under the international standard IEC 62304 [30]. SOUP is not prohibited; instead the manufacturer must inventory it, specify what it is required to do, and manage the residual risk through testing and monitoring proportionate to the harm a failure could cause [30]. The same posture has been proposed for third-party machine-learning models, which are SOUP with the added difficulties that their training data are usually undisclosed and their behaviour can drift after deployment [82]. For computerized systems in regulated manufacturing and laboratory practice, the parallel framework is the risk-based, supplier-leveraging approach of GAMP 5 [31], now being extended to adaptive AI. Adoption therefore depends not only on technical merit, but also on governance, training, procurement and local IT constraints.

This matters directly for the biopharmaceutics tasks of Section Biopharmaceutics: solubility, permeability and BCS classification and formulation development. An open predictor used to rank candidates or teach the BCS is low stakes and needs little beyond sensible checking. The moment a model output is offered in support of a regulatory claim, for instance a predicted BCS class advanced toward an ICH M9 biowaiver, that output becomes a high-stakes use of software whose provenance, version, training data and validation must be documented, exactly as the FDA credibility framework requires for the model’s context of use [8, 71]. The open versus closed distinction is then not merely a question of cost. An open tool pinned to a known version, with inspectable code, is easier to qualify and to defend as fit for purpose than a closed service that may change silently, which is the practical reason transparency and reproducibility, rather than licence price, should drive tool choice wherever a result must withstand audit [8, 30, 82]. This does not make open tools automatically qualified: a commercial platform brings a vendor qualification package whereas an open tool must be qualified by the user, and both routes are workable. Neither an open nor a commercial label establishes qualification by itself, which depends on the defined context of use, change control, verification, validation and governance. R based analysis packages have passed FDA review in the R Consortium submissions pilots [83], the open PK-Sim platform has a published automated qualification framework with a geometric mean fold error of 1.4 for CYP3A4 mediated interaction predictions [84], and nlmixr2 has been described and benchmarked in peer reviewed work [85, 86]. Table 3 therefore asks for a qualified platform, commercial or open.

In this review credibility means trust, established through evidence proportionate to risk, in the performance of an AI model for a specific context of use, following the FDA draft guidance [8]. It is not a single statistic. Credibility is not an intrinsic property of a model and is not established by a high coefficient of determination or by agreement with a training set. It should be distinguished from three related concepts: verification, which establishes that an implementation operates according to its specification; validation, which tests whether predictions of the quantities of interest are accurate and reliable enough for the intended use; and reproducibility, which asks whether results can be regenerated under documented conditions. Credibility integrates these and other lines of evidence to decide whether reliance on the model is justified for a specific decision. We divide it into five measurable dimensions (Table 7): predictive performance and robustness, reproducibility, data quality and provenance, regulatory and governance alignment, and human oversight and accountability. Alignment with regulatory standards means that the evidence is organized as the relevant framework expects, for example the FDA seven step process built on model influence and decision consequence [8] or the model risk and model impact assessment of ICH M15 [11]. The rigour required scales with the stakes, so a model that ranks compounds for teaching needs far less evidence than one that supports a dose or a biowaiver. Thresholds are specific to the context of use and should be prespecified in the credibility assessment plan, and Table 7 therefore gives examples of metrics and conventions, such as the two fold error criterion for pharmacokinetic predictions, and not universal limits. Depending on the use, the evidence may need to cover calibration, discrimination or prediction error, uncertainty, applicability domain coverage, external validation, data provenance and representativeness, version control, known limitations, and the consequences of false positive and false negative results. Not every metric suits every model, and the criteria must be matched to the scientific task.

Human oversight is a condition for using AI in pharmaceutical science, because the consequences of error fall on patients. The FDA and EMA principles place human centric design first [10], the EMA reflection paper stresses a risk based and human centred approach [9], and Article 14 of the EU Artificial Intelligence Act requires effective human oversight of high risk systems [90]. Three arrangements should be kept distinct: human in the loop, where a person approves each output before it takes effect; human on the loop, where a person supervises an autonomous process and can intervene, as in a self driving laboratory with defined stopping rules; and human in command, where a governing body decides whether, where and how a system is deployed and retired. The arrangement should follow model influence and decision consequence, so that dose selection, batch release and regulatory claims call for review of each output while low risk screening can run under supervision. Oversight in this sense covers more than checking a final output: it spans the selection of data, the formulation of the scientific question, the evaluation of model suitability, the interpretation of predictions, verification against experimental or authoritative evidence, and authorization of consequential decisions. Literature organization or preliminary screening can accept light touch, periodic review, whereas predictions that inform toxicological risk, clinical dosing, manufacturing release or regulatory claims need stricter verification and documented accountability.

Oversight is only as good as its design. Reviewers need the expertise, time and authority to disagree, otherwise automation bias turns review into a formality, and routine reliance on AI may deskill the workforce so that oversight becomes nominal [88]. Human involvement alone does not guarantee safety, because reviewers also need access to the evidence behind a prediction and clear responsibility for the decision, and plausible but wrong outputs are hard to detect without domain knowledge. Practical safeguards are independent verification against an authoritative source, audited review, documented override and decision logs consistent with data integrity expectations, and reviewer training, as proposed for human in the loop use in ongoing process verification under EU GMP Annex 22 and ICH Q9 [87]. The EMA accepted in 2025 clinical trial evidence generated by an AI tool supervised by a human pathologist, which shows that an explicit oversight arrangement can make AI output acceptable to a regulator [91]. A recent perspective applies the human in the loop principle to drug design, clinical trials, synthesis and manufacturing, analysis and quality control, and formulation, and assigns moderate to high oversight needs to classical machine learning and high oversight needs to deep learning and language models [89]. AI should augment scientific judgment without obscuring the responsibility of the investigators, developers, manufacturers and qualified professionals who rely on its outputs.

Whether each stage should use its own AI tool is a related question. A modular design gives each tool a narrow, validatable context of use, makes errors easier to localize, and creates a checkpoint at each hand off where a human can review and sign off, which strengthens human authority. It also permits a stage specific choice between open and closed tools (Table 4). Separate specialized tools could support target identification, molecular design, retrosynthetic planning, analytical characterization, formulation development, preclinical prediction, pharmacokinetic modelling, clinical trial analysis and manufacturing quality assessment, each with a specified intended use, defined inputs and outputs, and independently assessable performance criteria. The costs are interface and format mismatches, errors that propagate silently when modules are validated only in isolation, and the loss of the cross stage learning that integrated platforms exploit. We therefore recommend a modular architecture with explicit human review at each hand off and end to end validation of the pipeline for the decision it supports. Interfaces should carry explicit acceptance criteria and version control so that an uncertain or invalid output is not propagated automatically, and responsibility for the final decision should remain with qualified scientific and regulatory personnel. Modularity is not synonymous with safety, because errors can accumulate across components and integration introduces risks of its own.

The case for open tools in this review rests on cost, transparency and reproducibility and is not a claim of equal performance. Commercial platforms bring integrated and optimized models, proprietary training data that open tools cannot match, support and indemnification, and in several domains they may remain more accurate (for example molecular simulation, free energy estimation, physiologically based pharmacokinetic modelling and regulated pharmacometric analysis, where tightly integrated workflows, specialized physics based calculations and curated reference datasets are consequential). Where comparative evidence exists it is uneven (Table 8). An independent docking benchmark found that deep learning methods did not outperform classical engines, open (AutoDock Vina) or commercial (CCDC GOLD), in physical plausibility or on novel complexes [50]. The open weight Boltz-2 is reported by its developers to approach free energy perturbation at a small fraction of the cost, but this rests on a preprint and developer chosen benchmarks [48]. A simvastatin case study found comparable predictive performance in the open PK-Sim and the commercial Simcyp, although model development strategies differed [92]. Docking pose validity, ADMET prediction and PBPK simulation test different things and are not interchangeable forms of validation. Leaderboard performance of ADMET-AI on public data [27] does not show that a model generalizes to the chemical series of a company, where proprietary models trained on internal data may do better. Neither licensing model guarantees scientific validity: an open model may be poorly calibrated or used outside its applicability domain, whereas a proprietary model may be difficult to scrutinize independently or to reproduce after an undocumented update. Meaningful comparison therefore requires task specific benchmarking against appropriate reference data, prespecified performance criteria, assessment of uncertainty and applicability domain, and attention to the operational and regulatory requirements of the intended use.

Systematic head to head comparisons of free and commercial tools on identical independent datasets for ADMET, BCS classification, allometric scaling and virtual bioequivalence are largely absent. Our view that free tools are adequate for teaching, ranking and early stage work is therefore a judgment from the available evidence and practice and not a measured equivalence, and a dated, versioned, independent benchmark including both kinds of predictor would be the most useful way to settle the question. The practical objective is not to maximize openness or proprietary capability in isolation, but to select the tool that offers the most defensible combination of performance, transparency, reproducibility, integration and accountability for the decision at hand.

Four further considerations bear on tool choice. Security: open models and pipeline packages are software components that can contain vulnerabilities or malicious code and whose dependencies may become unmaintained, and agentic systems that call external tools widen the attack surface [93], so pinned dependencies, verified checksums, isolated execution and the SOUP style inventory above apply to open and closed tools alike. Intellectual property: licences determine whether outputs, fine tuned weights and derived models can be used commercially, and the ownership and patentability of AI generated molecules and text depend on the jurisdiction and on the human contribution, so the terms of each tool should be read before it is built into a pipeline. A tool’s licence is distinct from the ownership or protectability of what it generates. Sustainability: training and large scale inference consume substantial energy and carry a carbon cost, self hosting transfers that cost to the institution, and academic web services depend on grant funded servers that can be withdrawn, as the footnote to Supplementary Table S1 notes. Access to compute is itself unequal. Equity: free tools lower the cost barrier, but only where infrastructure, training and representative local data are also available (including language coverage, affordability and sustainable access in low and middle income settings), so open weights and open data are necessary but not sufficient for equitable access.

Forecasting in a fast-moving field is hazardous, but the trajectory of 2024–2026 supports a few disciplined predictions. First, the rentosertib result will be tested by outcome: within this horizon at least one AI-originated drug should report a phase III readout, and the field’s credibility will hinge on whether efficacy survives a larger and more heterogeneous population rather than on the discovery speed that first drew attention [12, 16]. Second, the open and closed worlds will continue to converge, with capable open-weight models closing the reasoning gap and structure prediction becoming routine infrastructure through open reproductions [22, 28, 32, 47]; the durable advantage of closed platforms will narrow to integration, support and regulatory familiarity rather than raw capability. Third, regulation will move from guidance to expectation, as the FDA credibility framework and its analogues become a standard part of submissions and SOUP-style inventory and qualification of AI components migrate from medical device practice into routine biopharmaceutics and manufacturing workflows [8, 30, 31, 82]. Fourth, biopharmaceutics will see in silico solubility, permeability and BCS classification used earlier and more confidently for triage and developability, while the biowaiver decision itself remains anchored to experimental confirmation [71, 76]. Fifth, the binding constraint will shift from model capability to data and verification: curated, well governed datasets such as the Therapeutics Data Commons, and the discipline of checking and documenting outputs, will matter more than access to any particular model [33]. A further structural shift, still nascent in pharmaceutical work, is the move from standalone language models toward agentic systems that plan and execute multi-step tasks by invoking external tools and data, a transition that emerging open standards such as the Model Context Protocol are beginning to formalize [93]. The logical extension of agentic systems is largely automated scientific discovery, an early and still unreliable version of which has already been demonstrated end to end and published [94]. The one prediction offered with confidence is the one the original editorial made, namely that the scientists who prosper will be those who adapt, choosing tools by evidence and keeping a human accountable for the result [1]. Beyond this decade the trajectory can only be sketched, and Table 9 offers such a sketch out to 2050, as a disciplined extrapolation rather than a forecast; across every horizon the pattern is the same, namely that capability migrates steadily to the machine while judgment, accountability and the verification of results remain with the pharmaceutical scientist.

With respect to AI self-awareness we see no basis for predicting a date, because self-awareness and consciousness have no agreed operational definition or validated test. The most systematic current approach derives indicators from neuroscientific theories to inform credences, without concluding that present systems are conscious and without excluding that future ones could become plausible candidates [95]. Views differ on whether such features could ever suffice: computational functionalists hold that the right functional organization could be enough, whereas biological naturalists argue that consciousness depends on being a living organism, which makes artificial consciousness unlikely along current trajectories [95, 96]. Present pharmaceutical models are not candidates on either view. Structure predictors and generative design networks are function approximators that run in an input, computation and output cycle with no continuous internal state, and a language model that writes that it is analyzing a molecule reproduces the statistics of language and does not report an inner state. The transformation of drug discovery rests on capability, not awareness. For pharmaceutical use the relevant properties are more mundane and measurable, namely reliability, calibration, controllability and goal directed autonomy, and these are already changing the risk profile whether or not any system is aware. Systems that already plan multistep workflows, invoke external tools, generate hypotheses and coordinate computational and laboratory tasks show expanding operational capability, not subjective experience, and what they require is governance: explicit operating boundaries, traceable actions, independent verification, human authorization before consequential steps, and mechanisms for interruption or rollback. Even if a system were judged aware, accountability would not change, because sponsors, marketing authorization holders, investigators and qualified persons remain legally responsible and regulators assess a model for its context of use and not its inner life. Increasing autonomy, not awareness, is what should shape the oversight arrangements of Section Human oversight and the modular design of AI workflows, and we would revisit the question if a validated and accepted test of machine consciousness emerged.

The recurring lesson of this review, that judgment rather than tool use is the enduring skill, carries a clear educational implication. Just as contemporary pharmaceutical curricula incorporated statistics, pharmacokinetics and evidence-based medicine, future curricula will require formal training in AI literacy, model validation, reproducibility, algorithmic governance and regulatory evaluation. Equipping pharmacy undergraduate and graduate students to interrogate an AI tool, to recognize its failure modes and to document its credibility for the decision at hand is now as fundamental as teaching them to design an experiment or interpret a pharmacokinetic concentration vs. time plot, and it is the surest way to ensure that the next-generation of pharmacists and pharmaceutical scientists directs these tools rather than defers to them [1, 97]. Table 10 sets out a concise competency map for this purpose, pairing each skill with the reason it matters. Competence in designing and exercising human oversight, including awareness of automation bias, belongs on that list (Section Human oversight and the modular design of AI workflows).