The Single-Model Illusion in AI-Driven Drug Discovery: Introducing a Systems-Level Multi-Model Framework for Translational Discovery

Abstract

Recent advances in AI-driven drug discovery have led to widespread narratives suggesting that a single model or platform can generate viable therapeutic candidates and, when combined with automated laboratory systems, rapidly progress to clinical development. These narratives often imply that AI-driven design coupled with robotic execution can substantially compress the path to Phase I trials and accelerate the treatment of complex diseases within a few years. However, practical implementation reveals a significant gap between model-level performance and end-to-end drug development success.

This work presents a systems-level analysis of current AI-driven pipelines and introduces a multi-model, feasibility-aware, closed-loop framework designed to address key structural limitations.

We identify three primary constraints:

(1) fragmentation across specialized tools requiring multiple independent models prior to synthesis,

(2) misalignment between benchmark performance and biological outcomes, and

(3) limited compatibility between novel compound generation and standardized experimental workflows.

We further demonstrate that downstream automation platforms, including robotic laboratory systems, accelerate specific classes of high-throughput experimental workflows but do not eliminate upstream fragmentation or the need for bespoke synthesis strategies. As molecular novelty increases, compatibility with standardized automation decreases, reinforcing the need for adaptive, human-guided experimental design.

We argue that drug discovery is fundamentally a decision-making system under uncertainty, rather than a single-model prediction problem. The proposed framework integrates multi-model orchestration, early-stage feasibility filtering, tiered candidate prioritization, and closed-loop experimental feedback to improve translation from computational predictions to real-world outcomes.

As of March 2026, even leading integrated platforms such as Isomorphic Labs have not released public wet-lab validation data despite strong computational benchmarks, underscoring the persistent translation gap. Bridging this gap requires not only improved models, but coordinated systems that integrate prediction, constraint, and experimental validation into a unified workflow.

This distinction between perceived and actual workflows is illustrated in Figure 1, and the underlying fragmentation across current platforms is detailed in Figure 2.

1. Introduction

Artificial intelligence has rapidly advanced capabilities in molecular design, protein structure prediction, and binding affinity estimation, with systems such as AlphaFold and emerging integrated platforms from Isomorphic Labs demonstrating strong benchmark performance.

Public and industry narratives often imply a simplified workflow:

AI model → candidate molecule → rapid preclinical → clinical entry

This perception is illustrated in Figure 1 (top panel), which reflects a commonly assumed linear pathway in which a single AI model, combined with automated laboratory execution, directly produces clinical-stage candidates.

However, real-world drug discovery is a multi-stage process requiring integration across heterogeneous tools, experimental validation, and iterative refinement. As shown in Figure 1 (bottom panel), actual workflows involve multiple computational steps, candidate prioritization, and extensive experimental validation prior to clinical translation.

Figure 1

Figure 1. Perceived vs. Actual AI-Driven Drug Discovery Workflows.Top panel: Commonly portrayed workflow in public and industry narratives, in which a single AI-driven platform combined with automated laboratory systems is assumed to directly produce clinical-stage candidates. Bottom panel: Actual workflow, requiring a multi-model computational pipeline (often 6–10 distinct tools), followed by candidate prioritization and extensive experimental validation, including in vitro and in vivo studies, prior to clinical translation. This contrast highlights the gap between narrative compression and system reality.

During the development of internal multi-model systems (2023), the author initially assumed that AI pipelines would produce directly actionable candidates for synthesis. In practice, it became clear that multiple independent tools—often 6–10 distinct systems—are required prior to synthesis, each addressing a different component of the problem.

This observation highlights a critical disconnect between perceived and actual system architecture, motivating the need for a more integrated framework.

2. Limitations of Current AI-Driven Drug Discovery Pipelines

2.1 Fragmentation of Functional Capabilities

Current AI systems are typically specialized:

• structure prediction

• virtual screening

• binding affinity estimation

• generative design

• ADMET prediction

• synthetic accessibility

No single system reliably spans all components. Even advanced platforms primarily compress subsets of these steps rather than replacing the full pipeline.

This fragmentation across specialized tools and vendors is illustrated in Figure 2, which shows the typical distribution of distinct platforms required across the pre-synthesis pipeline.

Figure 2. Fragmented AI Tool Stack and Execution Gap in Drug Discovery. Illustration of the multi-step computational pipeline required prior to synthesis, with representative companies and tools at each stage. Current workflows typically require 6–10 independent platforms spanning target identification, structure prediction, screening, affinity estimation, generative design, ADMET prediction, and prioritization. Downstream execution through automated laboratory platforms or contract research organizations (CROs) occurs after candidate selection and does not eliminate upstream fragmentation. Increasing molecular novelty further limits compatibility with standardized high-throughput workflows, necessitating bespoke synthesis strategies.

2.2 Local Optimization vs System-Level Success

Most tools optimize for:

• benchmark accuracy

• correlation metrics

• computational efficiency

rather than:

• experimental success

• biological activity

• clinical viability

This leads to locally optimized components without global optimization.

2.3 Risk Transfer Across Pipeline Stages

In modular pipelines:

Model A → Model B → Model C → Experimental stage

Each stage:

• optimizes its own objective

• transfers uncertainty downstream

Result: No single component is responsible for end-to-end success

2.4 Compounding Error

Errors accumulate across stages:

structure error → affinity error → ADMET error → biological failure

Even small inaccuracies at each step can result in failure at synthesis or assay.

2.5 Absence of Early Feasibility Constraints

Many pipelines prioritize:

• binding affinity

• novelty

while neglecting:

• achievable exposure

• pharmacokinetic feasibility

• realistic dosing constraints

This results in candidates that are computationally promising but clinically unrealistic.

2.6 Execution Layer Limitations and Automation Constraints

Recent advances in laboratory automation, including platforms such as Ginkgo Bioworks, have led to the perception that experimental validation can be seamlessly integrated into AI-driven pipelines. Systems such as Ginkgo’s Reconfigurable Automation Carts (RACs) and cloud laboratory infrastructure demonstrate the potential of closed-loop, high-throughput experimentation. For example, recent work integrating AI-driven experiment design with automated execution has shown the ability to evaluate tens of thousands of reaction conditions in parallel and achieve measurable optimization outcomes.

However, these systems primarily accelerate high-throughput, standardized biological workflows, including plate-based assays, repetitive liquid handling, and certain classes of ADME profiling. Their performance is strongest in settings where experimental conditions can be held constant across large numbers of parallelized reactions.

In contrast, they are less suited for bespoke small-molecule synthesis, particularly multi-step organic chemistry involving variable reaction conditions, anhydrous environments, complex workups, or structurally novel intermediates. As molecular novelty increases, synthesis pathways become less predictable and less compatible with standardized robotic workflows, often requiring significant protocol adaptation or fallback to traditional contract research organizations (CROs) with human medicinal chemists.

This leads to a fundamental constraint:

Increasing molecular novelty reduces compatibility with high-throughput automation systems and necessitates adaptive, human-guided chemical strategies.

Thus, while automation platforms can substantially accelerate certain execution steps, they do not eliminate the need for:

• human-guided medicinal chemistry

• adaptive synthesis design

• iterative experimental refinement

These limitations reinforce that experimental execution remains a critical and non-trivial component of the drug discovery pipeline, particularly for first-in-class compounds and novel chemical space.

As shown in Figure 2, the execution layer remains downstream of candidate selection and does not resolve upstream fragmentation, reinforcing the separation between computational prediction and experimental realization.

2.7 Runway and Business Model Pressures (2026 Landscape) Many AI-native discovery companies face mounting pressure as 2026–2027 approaches. With customers unwilling or unable to maintain subscriptions to 6–10 overlapping tools, and big pharma preferring targeted partnerships over broad platform licensing, several platforms risk exhausting runway unless acquired or pivoted toward asset development. This economic reality further highlights the limitations of tool-centric versus therapeutics-centric business models.

3. Reframing Drug Discovery as a Decision System

Drug discovery should be understood as: a sequential decision-making process under uncertainty rather than: a prediction problem solved by a single model.

Key requirements include:

• integration of heterogeneous models

• uncertainty estimation

• prioritization under constraints

• iterative experimental feedback

This shift from prediction to decision-making forms the basis for the proposed framework, illustrated in Figure 3, which contrasts fragmented pipelines with an integrated, closed-loop system.

Figure 3. From Fragmented Pipelines to Orchestrated Closed-Loop Drug Discovery Systems.

Left: Current state, in which independent models and tools operate with limited integration, leading to fragmented decision-making and high uncertainty. Right: Proposed system-level framework integrating multi-model orchestration, feasibility and constraint-aware filtering (e.g., exposure, toxicity, synthesizability), and tiered candidate prioritization. Experimental execution is incorporated within a closed-loop feedback system, enabling iterative refinement and improved translation from computational prediction to real-world outcomes.

4. Proposed Framework: Multi-Model, Feasibility-Aware, Closed-Loop System

The proposed system-level architecture is shown in Figure 3, which demonstrates how multi-model orchestration, constraint-aware filtering, and experimental feedback can be integrated into a unified decision-making framework.

4.1 Multi-Model Orchestration

Multiple models operate in parallel:

• structural prediction

• generative design

• affinity estimation

• physics-based methods

Outputs are compared to assess:

• agreement

• disagreement

• uncertainty

4.2 Consensus and Disagreement Scoring

Candidates are evaluated based on:

• cross-model agreement

• ranking stability

• diversity

This reduces reliance on any single model.

4.3 Early Feasibility Filtering (Exposure-Aware Screening)

Binding affinity is used as a proxy for required exposure:

This enables estimation of whether required concentrations are:

• physiologically achievable

• toxic

• unrealistic

Candidates requiring implausible exposure are deprioritized early.

4.4 Tiered Candidate Prioritization

Rather than binary selection:

• Tier 1: synthesis-ready candidates

• Tier 2: retained for iteration

• Tier 3: archived for future exploration

This preserves optionality while improving efficiency.

4.5 Hybrid Execution Model

The framework separates:

In-house (high-value):

• model orchestration

• decision logic

• IP generation

Outsourced (standardized):

• synthesis

• assays

• routine ADME

4.6 Closed-Loop Experimental Feedback

Experimental results are reintegrated:

prediction → synthesis → assay → feedback → updated prioritization

This enables:

• model recalibration

• improved decision-making

• reduction of systematic bias

5. Implications for TechBio Business Models

5.1 Tool vs Therapeutics Misalignment

Many AI companies optimize for:

• model performance

• customer adoption

rather than:

• drug development success

5.2 The Single-Model Illusion

Public narratives often imply: one model → one drug

In reality: multi-model systems and experimental validation are required

5.3 Hybrid Model as a Scalable Approach

A hybrid structure:

• preserves capital efficiency

• maintains IP control

• enables access to standardized infrastructure

6. Discussion

The expectation that AI systems can directly generate synthesis-ready drug candidates reflects a broader narrative compression within the field, as illustrated in Figure 1. In practice, and as detailed in Figure 2, multiple computational layers and independent tools are required prior to synthesis, and experimental validation remains essential.

In practice:

• multiple computational layers are required

• experimental validation remains essential

• automation does not eliminate synthesis constraints

Even when combining AI-driven design tools with automated laboratory systems, the pipeline remains fragmented. The primary bottleneck is not candidate generation, but integration of

predictions, constraints, and experimental feedback into a coherent decision system.

6.1 Analogous Therapeutic Development in Environmental and Biological Systems

The multi-model, constraint-aware framework described here has been applied by the author in the development of therapeutic strategies for microplastic and nanoplastic mitigation. In this context, compounds were identified and evaluated through iterative experimental workflows.

Rather than relying on a single predictive model or screening metric, candidate systems were prioritized using multiple criteria, including aggregation behavior, physicochemical interactions, and experimentally observed outcomes. This approach enabled the identification of functional therapeutic mechanisms that would not have been reliably predicted by any single model or metric alone.

Combinations of compounds were designed, prioritized, and experimentally evaluated in wet-lab systems, including ex vivo large-volume assays designed to approximate physiologic conditions.

Importantly, these systems required integration of heterogeneous signals and rapid experimental feedback, reinforcing that actionable therapeutic development emerges from coordinated decision-making across models and experiments. These findings support the broader conclusion that effective drug discovery frameworks must incorporate multi-model orchestration, feasibility constraints, and closed-loop validation, particularly when targeting complex or poorly characterized systems.

7. Conclusion

AI has significantly improved early-stage molecular design and search space exploration.

However, current approaches remain limited by fragmentation, local optimization, and incomplete integration with experimental systems.

We demonstrate that:

• multiple specialized tools are required prior to synthesis

• no single AI system spans the full pipeline

• automation platforms do not eliminate synthesis constraints

• increasing molecular novelty reduces compatibility with standardized workflows

We propose a shift from model-centric approaches to system-level orchestration, integrating multi-model predictions, feasibility constraints, and experimental feedback.

Ultimately, successful drug discovery depends not on individual models, but on the ability to construct robust, closed-loop systems capable of making decisions under biological uncertainty.

The transition from fragmented pipelines to integrated, closed-loop systems (Figure 3) represents a necessary evolution for accelerating the translation of AI-driven discovery into real-world therapeutics.

  This paper is also available at: https://doi.org/10.5281/zenodo.19240171

Keywords

AI drug discovery, multi-model systems, feasibility filtering, pharmacokinetics, closed-loop learning, techbio, drug development, automation, medicinal chemistry