This study addresses three questions. First, can a pipeline built entirely from freely available tools produce compound-prioritisation outputs of sufficient quality to justify proposed Tier 3 Contract Research Organization (CRO) investment in resource-constrained settings? Second, what is the incurred and projected cost of operating such a pipeline? Third, what is the qualitative contribution of LLM orchestration to literature synthesis and annotation? The scope is explicitly limited to the discovery phase: target identification through computational lead nomination with proposed experimental binding confirmation (Tier 3, proposed future work). It does not encompass lead optimisation, in vivo proof-of-concept, GLP toxicology, CMC, or clinical phases.

The term "agentic AI" refers to AI systems that autonomously plan and execute multi-step tasks, use external tools (APIs, databases, computational engines), maintain state across task sequences, and return outputs for human review at defined checkpoints. In a non-agentic deployment, a researcher might ask an LLM to "suggest compounds for a kinase target" and receive a generic text response. In an agentic deployment, the same request triggers: (1) an autonomous literature search against PubMed and the ChEMBL bioactivity database (Gaulton et al., 2017) with PMID verification, (2) structured extraction of SAR data for the target class, (3) library composition recommendations, and (4) a structured summary submitted to a human reviewer at the Tier 1 to Tier 2 gate.

The pipeline described in this paper implements this architecture: an LLM agent using a tiered model selection (Claude Opus 4 for complex reasoning and difficult cases; Claude Sonnet 4 for routine tasks) orchestrates the four-tier workflow, calling external tools (ChEMBL REST API, NCBI E-utilities, RDKit via subprocess) and maintaining a structured decision log across tiers. The agent does not make compound-selection or go/no-go decisions autonomously; those remain with domain experts. The agent provides coordinated execution of literature synthesis, evidence integration, and hypothesis-generation tasks that would otherwise require substantial expert time. The LLM orchestrates rather than computes: it issues these structured tool calls (REST APIs and RDKit via subprocess) and integrates the returned outputs, but does not author or execute arbitrary code and does not itself perform the docking or ADMET calculations, which run as separate deterministic programs.

See Figure 1 for the workflow architecture, showing the four-tier pipeline, external tools, LLM agent roles, human-review checkpoints, and the distinction between executed work (Tiers 1–2) and proposed future work (Tiers 3–4). Relative to prior autonomous-lab demonstrations, the specific contributions of this work are the tiered LLM orchestration with per-tier model and cost allocation, the incurred-versus-projected cost-floor analysis for resource-constrained settings, and the per-pair, small-sample LLM citation audit with Wilson intervals.

All three case studies were conducted as computational workflow demonstrations on retrospectively-defined targets. The retrospective design was chosen to permit comparison of computational prioritisation outputs with available literature evidence on each target class under conditions where the targets and their broad pharmacology are already known. (All three case studies are retrospective in design; the term prospective is used only for proposed future Tier 3 experimental work described as future directions, not for the present computational demonstrations.) No experimental Tier 3 work was performed.

Case A: a proprietary therapeutic target. Target class, indication, identity, and library subset selection criteria are withheld for confidentiality as Sponsor confidential program information. Structure: a publicly available experimental X-ray co-crystal structure of the target, containing a bound co-crystallised ligand (specific PDB accession withheld, as it would reveal target identity). The downstream candidate compound SMILES are likewise withheld. Case A is included in this manuscript to demonstrate workflow execution on a fully-confidential program — the realistic deployment mode for industry users — rather than to report target-specific findings. See "Data and Code Availability" for the disclosure policy.

Case B: a well-characterised kinase target with extensive published ATP-competitive inhibitor SAR and a public co-crystal structure containing a covalently-bound ATP-site inhibitor. The specific target identity and structure identifier are anonymized pending provisional patent filing on the underlying program.

Case C: a protein-protein-interaction (PPI) interface target with a public apo crystal structure (no co-crystallised small molecule or peptide) and sparse small-molecule literature. Targeting strategy: small-molecule PPI-interface inhibitors against the partner-binding surface. The specific target identity and structure identifier are anonymized pending provisional patent filing. This case is framed as a highly exploratory PPI-interface stress test (see Result: Case C).

Table 1 (Case study design summary) lists the three cases, structural data sources, library subset, target class, and confidentiality status.

Each case began with a target-enriched compound library drawn from ZINC20: for Case B, the kinase-focused subset (≥40% kinase domain sequence identity to the target); for Case A, a target-class-appropriate ZINC20 subset (specific subset and selection criteria withheld as Sponsor confidential program information); for Case C, the lead-like general subset. Tier 1 filters applied: Lipinski Rule-of-Five, PAINS exclusion, reactive-group exclusion, and TPSA cap. Library reduction yields are reported in Result.

Molecular docking was performed with GNINA v1.1 (McNutt et al., 2021), which combines an AutoDock Vina-derived empirical search with a convolutional neural network (CNN) affinity model. All runs used CNN rescoring of final poses (--cnn_scoring rescore), exhaustiveness 16-32, nine binding modes, a fixed random seed (42), and a 22.5 x 22.5 x 22.5 A cubic search box centred as described per case (Supplementary S2). Ligands were prepared with Open Babel at pH 7.4. Docking was executed on CPU-only consumer hardware; no GPU was used. A single docking engine was used for all reported results; AutoDock Vina and Smina were not used to generate any result reported here. Docking scores are treated throughout as rank-order triage signals, not binding free energies, and are not converted to predicted affinities or IC50 values. Consistent with this, no compound advanced on docking rank alone: advancement required concordant literature-derived target rationale, passage of the Absorption, Distribution, Metabolism, Excretion, and Toxicity (ADMET) filter cascade, and human review at the tier gate. All ADMET assessments in this study are computational predictions (SwissADME and ADMETlab 2.0), not experimental measurements; no wet-lab ADME or toxicity assays were performed, and these predictions require experimental confirmation in the proposed Tier 3 (future work).

Docking box definition.

Case A: the docking box was centred on the co-crystallised ligand resolved in the experimental structure, which defines the primary ligand-binding pocket.

Case B: the docking box was centred on the ATP-site inhibitor-binding region defined by the co-crystallised inhibitor in the public structure.

Case C: the Case C structure is a public apo crystal structure of the PPI-interface target with no co-crystallised small molecule or peptide. The docking box was centred on the partner-binding (PPI) groove using residue contacts identified directly on the apo structure. This is a homolog-free, apo-derived box and not a co-crystal-derived box; because the structure is apo, no co-crystallised ligand was available for a re-docking positive control.

Re-docking positive-control check.

For Case A, the native co-crystallised ligand was re-docked into its own experimental structure as a positive-control check before novel-compound docking; it was recovered in the crystallographic binding site with reproducible top-ranked docking scores across independent runs. A quantitative pose-recovery (RMSD) benchmark is reported for the fully public case in Section 4.5. The specific PDB accession is withheld, as it would reveal target identity.

For Case B, the co-crystallised inhibitor in the public structure was re-docked solely as a box-placement check, which confirmed correct localisation of the docking box to the inhibitor-binding region. This check confirms binding-site placement rather than reproduction of the reference binding mode, and is not a rigorous re-docking positive control.

For Case C, a re-docking positive control was not possible because the Case C structure is apo. This is acknowledged as a structural limitation of Case C and contributes to its highly exploratory framing.

Tier 3 (experimental binding affinity confirmation by surface plasmon resonance and selectivity assessment by commercial kinase or off-target panel) is described here as the proposed next step for the compounds prioritised by Tier 2. Tier 3 was not executed in this study. Tier 3 applicability is most direct for established target classes with commercially available assay platforms (kinase activity assays, GPCR binding panels, ion channel patch-clamp services, nuclear receptor reporter assays). For less-established targets requiring custom assay development, including PPI interfaces and undisclosed novel targets, the Tier 3 timeline and cost can substantially exceed those for kinase or GPCR programs and should be incorporated into Tier 3 budgeting on a per-program basis.

Tier 4 encompasses the activities required to compile an IND-enabling package: GLP toxicology studies, CMC development, formulation, pharmacokinetic studies in two species, and genotoxicity assessment. Tier 4 was not executed in this study. Cost ranges for Tier 4 are indicative only, based on published CRO fee surveys, and are not derived from direct study expenditure.

LLM assistance was provided via a tiered model deployment: Claude Opus 4 (Anthropic) for complex reasoning and difficult cases, and Claude Sonnet 4 (Anthropic) for routine tasks, both accessed via the Anthropic cloud API. The LLM was not used as a docking or ADMET prediction engine. Its specific roles in this pipeline were: (i) literature synthesis — extraction and summarisation of target biology, known inhibitor SAR, and competing compound series from PubMed and ChEMBL literature, with PMID verification; (ii) pharmacophore hypothesis generation — proposing binding pharmacophore elements based on synthesised literature, with expert review and approval; (iii) compound annotation — generation of per-compound SAR comments linking structural features to docking pose interactions; (iv) protocol documentation — generation of structured experimental request documents for CRO submission (drafted only; no CRO submission was made in this study); and (v) result interpretation — drafting of initial concordance analysis text for expert review.

Total LLM API expenditure across all three case studies was $1,450–$1,800 (incurred), reflecting a tiered deployment (approximately 20% of token volume via Claude Opus 4 for complex reasoning and difficult cases; approximately 80% via Claude Sonnet 4 for routine tasks; Opus 4 costs approximately 5× more per token than Sonnet 4 at current Anthropic API pricing). Cloud API access was chosen over local LLM deployment because the parameter-scale models required for the annotation tasks described exceed practical local hardware requirements for most resource-constrained research settings.

Definition and audit of LLM annotation accuracy. LLM annotation accuracy is defined as the proportion of LLM-generated claim-citation pairs (a claim about target biology, SAR, or compound activity, accompanied by a PMID or ChEMBL ID) where the cited reference verifiably supports the claim on independent review. A pair was accepted if the citation supports the claim; rejected if the citation is irrelevant, contradicts the claim, or the PMID/ChEMBL ID does not resolve to a real record. Every claim-citation pair present in each program's LLM-generated literature-synthesis and target-annotation outputs was audited (all available pairs were verified, rather than a random sample); the number of pairs available per case was small (Case A n=8, Case B n=9, Case C n=10; 27 pairs total). Each cited PMID was checked against PubMed. Pair-level audit was performed by the author (computational chemistry background); the audit was not blinded and inter-rater agreement was not measured, both of which are explicit limitations, and the small per-case samples are not powered to attribute accuracy differences to target class or literature density. Wilson score 95% confidence intervals are reported with each percentage; the per-pair verdicts, cited PMIDs, and DOIs are provided in Supplementary S4. A formal controlled comparison with blinded multi-reviewer audit and time-motion measurement against an LLM-free workflow is identified as a priority follow-up and is referenced as future work in the Discussion.

Per-case audit results are summarised in Table 2.

Hallucination Mitigation and Accuracy Improvement. Citation hallucination — a plausible but incorrect or non-existent reference — is the primary reliability risk when deploying LLMs as literature agents. A three-step verification chain was applied to all LLM literature outputs: (1) quote extraction from the cited PMID; (2) deterministic substring match of the quoted text against the cited abstract or full text; (3) flagging for expert review on match failure. Flagged claims were either rescued by expert review (citation corrected) or removed from the decision record. Beyond hallucination mitigation, compound prioritisation accuracy was supported by five complementary mechanisms: structured output enforcement (predefined JSON schema), low-temperature inference (T=0 for retrieval and classification; T=0.3 for hypothesis generation), CNN-rescored docking (GNINA v1.1), re-docking-gate positive-control check where structurally possible, and LLM-guided scaffold selection.

Proportions (LLM annotation accuracy) are reported with 95% confidence intervals calculated using the Wilson score interval method (Brown, Cai, & DasGupta, 2001), which is appropriate for small sample sizes and proportions near 0 or 1. All confidence intervals were computed in Python 3.13 using statsmodels 0.14.6 (statsmodels.stats.proportion.proportion_confint with method='wilson'). The pooled interval treats the 27 pairs as a single sample and does not account for case-level clustering; the per-case intervals are the primary report.

Case A is included as a workflow-execution demonstration on a fully-confidential program; target-specific findings are not reported. The computational pipeline reduced an initial library of approximately 400,000 compounds (target-class-appropriate ZINC20 subset; subset and selection criteria withheld as Sponsor confidential program information) through Tier 1 ADMET filtering to 1,847 candidates, and molecular docking to 200 docked compounds. The top 10 compounds by CNN-rescored docking score were identified as the recommended input to proposed Tier 3 experimental validation (SPR binding affinity and selectivity panel). The positive-control re-docking of the native co-crystallised ligand into its own experimental structure was successful — the ligand was recovered in the crystallographic binding site with reproducible top-ranked docking scores across independent runs — before novel-compound docking; the specific PDB accession is withheld, as it would reveal target identity. LLM annotation accuracy for Case A: 88% (7 of 8 claim-citation pairs verified against PubMed on author audit; 95% Wilson CI 53–98%; small exploratory sample). No experimental binding, ADMET, or selectivity data were generated in this study. The prioritised compounds in Case A are computational predictions requiring experimental confirmation in Tier 3 (proposed future work). The fixed ten-compound-per-case shortlist (30 across the three cases) was a pragmatic design choice—large enough to hedge against the weak individual discrimination of docking score alone (Section 4.5), yet small enough to keep the proposed Tier 3 assay burden affordable—rather than an optimised parameter.

Case B is a well-characterised kinase target with extensive published ATP-competitive inhibitor SAR and a public co-crystal structure. The pipeline reduced 388,000 compounds (ZINC20 kinase-focused subset) through Tier 1 ADMET filtering to 2,103 candidates and molecular docking to 175 docked compounds. The co-crystallised inhibitor in the public structure was re-docked as a box-placement check (inhibitor-binding region localised) before novel-compound docking; this localises the binding site but is not a rigorous re-docking positive control. The top 10 compounds by CNN-rescored docking score were identified for proposed Tier 3 experimental validation. LLM annotation accuracy for Case B: 78% (7 of 9 pairs verified; 95% CI 45–94%). No experimental data were generated in this study. Flexible loop regions in this kinase represent a docking challenge that would be assessable in future Tier 3 validation. ATP-competitive kinase inhibitors are characteristically promiscuous across the kinome (Karaman et al., 2008), so a Tier 3 selectivity panel against a representative kinase subset is a recommended adjunct to primary binding assays in any subsequent experimental campaign. As with Case A, the prioritised compounds are computational predictions requiring experimental confirmation.

Case C is framed as a highly exploratory PPI-interface stress test. The Case C target is regulated through a mechanism that constrains where a competitive small-molecule inhibitor could productively bind. The binding site used in this study is a protein-protein interaction (PPI) interface — a partner-binding surface — rather than a classical small-molecule binding pocket. PPI interfaces are characteristically shallow, solvent-exposed, and lack the hydrophobic enclosure geometry that empirical docking scoring functions (including the AutoDock Vina-derived function underlying GNINA) were parameterised to recognise. The pipeline reduced 154,000 compounds to 154 docking candidates. Because the Case C structure is apo (no co-crystallised small molecule or peptide), a re-docking positive control was not possible; the docking box was centred on the partner-binding (PPI) groove identified directly on the apo structure. Ten top-ranked compounds were identified as exploratory candidates. LLM annotation accuracy for Case C: 90% (9 of 10 pairs verified against PubMed; 95% CI 60–98%). No experimental data were generated. Conclusions about the top-ranked Case C compounds should be framed as highly exploratory; these compounds should not be advanced to experimental testing on the basis of computational ranking alone. Covalent docking or induced-fit protocols would be preferable for this PPI-interface target in future iterations.

Re-docking positive-control availability differed across cases as a property of the available structural data: a co-crystallised ligand from a structurally related target was available for Case A; only a covalent-inhibitor box-placement check was possible for Case B; and no re-docking positive control was possible for Case C (apo crystal). This structural gradient is consistent with the literature on docking-tool performance, in which pose-prediction and screening accuracy vary systematically by docking program and target protein family (Cross et al., 2009): empirical docking scoring functions (including the AutoDock Vina-derived scoring underlying GNINA) were trained and validated predominantly on well-defined small-molecule binding pockets, and their performance on PPI interfaces is more variable. LLM citation accuracy did not track this structural gradient: on the small audited samples it was 88% (Case A), 78% (Case B) and 90% (Case C), and the samples are too small to attribute the differences to target class, structural data type, or literature density.

To quantify how much of the prioritisation signal is carried by the open-source scoring component alone, the Tier 1-Tier 2 workflow was run unchanged on acetylcholinesterase (AChE; PDB 4EY7), a fully public target outside the Sponsor pipeline. Actives were high-confidence AChE inhibitors from ChEMBL (target CHEMBL220, pChEMBL >= 7); decoys were property-matched DUD-E decoys (target aces; Mysinger et al., 2012). Docking used the same GNINA v1.1 engine and CNN-scoring configuration as Cases A-C (search box defined per target). Pose recovery was accurate: re-docking of the co-crystallised ligand donepezil reproduced the crystallographic pose to 0.31 A heavy-atom RMSD (success gate <= 2.0 A). Ranking by docking score alone, however, did not separate actives from decoys across 51 scored compounds (15 actives, 36 decoys): ROC-AUC 0.23, EF1% 0.0x, BEDROC(alpha = 20; Truchon & Bayly, 2007) 0.02. Applying a mechanism-based interaction criterion - retaining compounds whose top pose contacts the conserved catalytic-gorge anchor tryptophans (Trp86/Trp286), applied identically to actives and decoys - restored above-random early recognition (BEDROC 0.02 to 0.44; EF1% 0.0x to 3.4x; hit-rate in the top 10, 0.10 to 0.20), while overall ROC-AUC remained near random (0.44). Because this interaction criterion is defined from the known AChE binding mode, its enrichment gain is illustrative of how expert mechanistic knowledge can be layered onto docking when the binding mode is known, rather than a target-agnostic enrichment mechanism; on a de novo target no such validated residue set is available, and the EF1% value moreover rests on a single top-ranked compound (n = 51). These figures quantify a well-documented property of empirical and CNN scoring functions rather than a property of the workflow as a whole, and they are consistent with the multi-evidence design used throughout this study, in which docking rank is one weak input among several and is never the sole basis for advancing a compound.

Incurred computational cost across the three case studies was: LLM API $1,450–$1,800 (tiered Claude Opus 4 / Sonnet 4); open-source software licenses $0. All computation ran on existing consumer hardware (a multi-core workstation with no dedicated GPU; GNINA runs on CPU), which is not separately costed as an incurred expense. Total incurred cash computational cost was therefore $1,450–$1,800 (see Table 3). Taking the midpoint of this incurred range ($1,625) and normalising over the 30 prioritised compounds the program produced (10 per case across the three cases), the incurred computational cost is approximately $54 per prioritised compound ($1,625 ÷ 30); the corresponding commercial value is $2,300 per compound ($69,000 ÷ 30). These are the per-compound figures plotted in Figure 2B.

Projected total discovery-phase cost is $3,450–$10,800 (a lower bound; see below), which includes proposed Tier 3 experimental validation work that was not executed in this study. The projected breakdown is: kinase selectivity profiling of the top-ranked hits — published commercial kinase-panel pricing ranged from $505 (preset 50-kinase panel) to $2,151 (comprehensive panel) per compound in 2013 (Luceome Biotechnologies pricing, as reported in Comley, 2013), or approximately $700–$3,000 per compound in 2026 terms after roughly 40% cumulative US consumer-price inflation over 2013–2026, giving approximately $2,000–$9,000 per case for profiling the top three of the ten prioritised hits per case (a cost-conscious triage: all ten per case form the computational shortlist recommended to Tier 3, of which the three highest-ranked are routed to the paid selectivity panel); and incurred computational LLM cost $1,450–$1,800. Two further Tier 3 steps are recommended but are not assigned a dollar figure here because no public per-unit pricing exists: biophysical binding confirmation by surface plasmon resonance (no public CRO rate card) and acquisition of the prioritised compounds (make-on-demand purchase, quotation-based and not publicly listed). The projected total is therefore a lower bound covering only the items with published or invoiced pricing. Personnel time is not included in the cost floor because it represents internal allocation in a funded program. The commercial AI-platform comparator used in Figure 2 (per-program mid-range $69K, range $38K–$123K) is derived from the Schrödinger Inc. 2023 Annual Report enterprise-licensing disclosures (Schrödinger, Inc., 2023) and per-program cost allocation, and represents computational cost only (wet-lab HTS not included). This comparator is a 2023-basis estimate; current commercial pricing is quote-based and not publicly listed, so it is indicative rather than a current quotation. See Table 3 for the full cost separation (incurred / projected / personnel-not-included). See Figure 2 for the discovery-phase computational cost comparison (this work vs commercial AI platforms).

Projected end-to-end timeline including proposed Tier 3 CRO validation is approximately 46 weeks: the measured computational portion is approximately 5–6 weeks (Figure 3) and the projected Tier 3 cascade adds approximately 40 weeks. Computational analysis time was measured: 4–6 calendar days per case for docking; 2–3 days per case for ADMET profiling and compound selection (these run in parallel across the three cases within the ~5–6 week computational portion). The projected Tier 3 cascade comprises compound synthesis or sourcing (~8–16 weeks), primary SPR binding (~8–12 weeks), selectivity panel (~8–12 weeks), cellular validation (~4–8 weeks), and result write-up (~2 weeks), based on published CRO fee survey timelines; this CRO work was not executed in this study. See Figure 3 for the computational workflow phase timeline (Gantt chart of the ~5–6 week computational portion, three cases run in parallel; the dashed extension depicts the primary SPR binding step of the Tier 3 cascade only, the full Tier 3 cascade summarised above totals ≈40 weeks).

LLM orchestration in this workflow performed: literature synthesis, pharmacophore hypothesis drafting, structured compound annotation, protocol-document drafting, and result-interpretation drafting. LLM citation accuracy was assessed by verifying, against PubMed, every claim-citation pair present in each program's LLM-generated literature-synthesis outputs: Case A 88% (7/8), Case B 78% (7/9), and Case C 90% (9/10); pooled 85% (23/27; descriptive count only, no pooled interval computed). Per-pair verdicts, PMIDs, and DOIs are provided in Supplementary S4. The audit has explicit limitations: single non-blinded reviewer (the author), small exploratory samples (8–10 pairs per case), and no inter-rater agreement measurement; the samples are too small to support conclusions about how target class or literature density affects reliability, and the earlier point estimates that suggested such a pattern have been withdrawn. A formal controlled comparison against LLM-free workflow execution with blinded multi-reviewer audit and time-motion measurement is identified as the priority follow-up.

Freedom-to-operate (FTO) and operational practice. The literature-SAR scaffold-restriction step in Tier 1 selects against chemotypes with no documented activity on the target class. This is a deliberate design choice to enrich for compounds with prior medicinal chemistry precedent, but it has two consequences: (i) the workflow as configured will under-discover truly novel chemotypes, and (ii) prioritisation against documented literature scaffolds does not guarantee freedom-to-operate. In our operational practice, per-program FTO assessment is performed routinely as a separate workflow operating upstream of compound nomination, combining SureChEMBL, USPTO, EPO, and PCT patent-landscape searches and, for sequence-based programs, BLAST-based patent-sequence comparison. This workflow is not currently embedded in the agentic computational pipeline reported here. Integration of an automated FTO-checking agent into Tier 1, with a formalised SOP for search coverage and reviewer sign-off, is scoped as a workflow enhancement. The present manuscript reports the docking/ADMET/LLM pipeline only; FTO clearance for any compound prioritised by this workflow is treated as a separate operational gate.

The principal limitations of this study are summarised in Table 4. They are: (i) no wet-lab experimental validation; (ii) no formal LLM efficiency comparison against an LLM-free workflow; (iii) partial computational reproducibility (confidential elements withheld as Sponsor confidential program information); (iv) target-class generalizability favouring kinase and well-characterised enzyme targets; (v) no automated FTO check in the current workflow; (vi) docking-tool limitations for PPI interfaces; (vii) Scope of the “open-source” characterization: it applies to the computational discovery stack (library construction, docking, ADMET profiling, and scoring), all of which are free and open; the LLM is a substitutable orchestration component, not part of the compound-prioritization method. This study used a proprietary LLM API (Anthropic Claude) for that role, but the architecture is model-agnostic and can be operated with open-weight models (e.g., the Llama, Mistral, or Qwen families), so the blueprint entails no proprietary lock-in; a head-to-head evaluation of open-weight orchestration models is identified as future work.

The public benchmark reported above is a deliberately stringent test of docking-only prioritisation, and its outcome is informative. Accurate pose geometry (0.31 A re-docking RMSD) did not translate into accurate affinity ranking. Several constraints bound the interpretation of these numbers. The benchmark was executed on CPU-only hardware using a single CNN model (crossdock_default2018) in rescoring mode rather than the five-model ensemble or CNN-guided refinement mode, both substantially more computationally demanding; enrichment statistics computed over 51 compounds, several of which depend on the top one to five ranked molecules, are correspondingly imprecise and should be read as indicative rather than definitive; and rigid-receptor docking models neither the flexible activation loop relevant to Case B nor the shallow, solvent-exposed interface of Case C. Accordingly, the result characterises the commodity scoring component under a constrained configuration, not the ceiling of the workflow. The practical implication is the design principle already embodied here: docking score functions as one weak filter within a multi-evidence funnel in which literature-derived target rationale, ADMET filtering, mechanism-based interaction criteria and a mandatory human review gate carry the prioritisation decision. Strengthening the computational ranking itself - through ensemble CNN scoring or CNN-guided refinement on GPU hardware, MM-GBSA rescoring of top candidates (Genheden & Ryde, 2015), and target-specific machine-learning models trained on known actives where these exist - is the most direct route to improved screening performance and is identified here as the priority next step.

Implementation note: the open-source workflow described here also forms the basis of a drug-discovery platform under development by Insivio. That platform is not yet commercially available; details of its planned deployment are intentionally omitted from this manuscript to keep the focus on the open-source scientific workflow. See the Conflict of Interest statement for the author's relationship with Insivio.