The Dose Makes the Poison: Exposure-Aware Evaluation of Generative Drug Design

arXiv:2611.13398 · q-bio.QM, cs.AI, cs.LG · Submitted 2026-11-10 · Read on arXiv

Listen

Radio episode about this paper

Transcript

Introduction to the show: ident: Genomics Radio. Generated commentary on the latest computational biology and genomics papers.

Ines: We're looking at the paper: "The Dose Makes the Poison".

Marcus: Here's the gist of the paper. Generative models are increasingly used to propose molecules for drug discovery They can produce far more candidates than can be experimentally tested, making computational selection essential This work investigates this gap by using physiologically based pharmacokinetic (PBPK) simulations to re-evaluate molecules selected from three PARP1-targeted generative pipelines Of 60 shortlisted molecules, seven fail to achieve sufficient target exposure, while 28% of the remaining candidates show predicted safety risks at their required dose Moreover, 85% of candidates with an achievable dose are outperformed by molecules overlooked during static selection These findings suggest that accounting for dose and exposure can substantially change which generated molecules are considered promising for drug development.

Ines: Exposure-Aware Evaluation Methodology We evaluate generated molecules based on the dose needed to achieve sufficient target exposure and the associated safety margins For each candidate, we combine predicted molecular properties with physiologically based pharmacokinetic (PBPK) simulations to estimate drug exposure following oral administration We then determine the dose required for target engagement and assess safety at that dose, accounting for uncertainty in both the molecular predictions and simulated exposure The minimum effective dose (MED) is determined by searching for the minimum oral dose at which the simulated free (unbound) drug concentration remains above this threshold during the dosing interval This MED depends on both potency and pharmacokinetics, as even a highly potent molecule may require a large dose if it is poorly absorbed or rapidly eliminated Safety at the required dose is assessed using the exposure-adjusted safety margin Mj = IC50j / Cj (d∗) This allows us to compare candidates based on their predicted behavior at a relevant dose.

Marcus: Pipeline Diversity and Static Selection We investigate three PARP1-targeted generative pipelines that differ in how they generate and optimize molecules These pipelines include MOOD, which is a docking-guided diffusion model without explicitly considering ADME during generation, REINVENT4-MPO, which is fine-tuned through reinforcement learning to optimize potency and ADME properties, and Mol2Mol, a seed-based approach that generates analogues of existing molecules Both REINVENT4-MPO and Mol2Mol are implemented using the REINVENT4 framework.

Ines: All generated molecules undergo the same filtering and ranking procedure, which involves applying hard filters for drug-likeness, synthetic accessibility, physicochemical properties, predicted ADME, structural alerts, mutagenicity, and predicted PARP1 activity The surviving molecules are then ranked using a weighted geometric mean (WGM) of static molecular desirability terms.

Paper summary: Ines: We just talked about how the paper investigates this gap between static property prediction and actual drug exposure requirements, focusing on molecules selected from three different PARP1-targeted generative pipelines Ines

Marcus: They found that out of sixty shortlisted molecules, seven failed to get enough target exposure, and twenty-eight percent of the rest showed predicted safety risks at their required dose Marcus

Yuki: It seems the core argument is that conventional property-based screening overlooks how those properties translate into actual drug exposure when you consider the necessary dose Yuki

Ines: The methodology involves using physiologically based pharmacokinetic simulations to estimate drug exposure after oral administration for each candidate and determining the minimum effective dose, or MED Ines

Marcus: They define this MED as the lowest oral dose where the simulated free drug concentration stays above a certain threshold during dosing intervals, which depends on both potency and pharmacokinetics Marcus

Yuki: And they also assess safety at that specific required dose using an exposure-adjusted safety margin, which they define as M j equals IC50 j divided by C j (d star) Yuki

Ines: This margin tells you if the predicted exposure is below the off-target inhibition threshold, and a larger margin is better Ines

Marcus: They used endpoint specific PBPK errors for uncertainty because they evaluate hERG and PARP2 using different parts of the concentration time profile Marcus

Yuki: They also show that eighty-five percent of candidates with an achievable dose were outperformed by molecules that were overlooked during the static selection phase Yuki <ref:2611.13398#pg1,of candidates with an achievable dose>

Ines: So, in short, the paper demonstrates that accounting for dose and exposure can significantly change which generated molecules are considered promising for drug development Ines

Marcus: It’s a shift from just looking at whether a molecule is potent to understanding its behavior under real-world dosing conditions Marcus

Conclusion: Ines: The paper, "The Dose Makes the Poison: Exposure-Aware Evaluation of Generative Drug Design," by Franz and Erichson, really pushes the idea that static selection criteria are insufficient when we talk about drug discovery Ines

Marcus: It shows that simply relying on predicted potency and toxicity doesn't capture how those properties combine to determine exposure at the dose needed for treatment, which is a crucial detail for any computational selection process Marcus

Yuki: What this means for us in population genetics is that we can’t just focus on the best single-molecule traits; we need to consider how those traits interact within a dosing scenario Yuki

Ines: The authors highlight that the usefulness of static off-target predictions depends entirely on how well they reflect risk at the exposure required for target engagement Ines

Marcus: So, future work needs to explore whether this disagreement between static ranking and exposure-aware outcomes persists across other targets and scoring strategies Marcus

Yuki: We need to see if this concept of dose-dependent risk holds up when we look at different biological systems or different experimental measurements Yuki

Maria Franz, N. Benjamin Erichson

International Computer Science Institute

q-bio.QM, cs.AI, cs.LG

Submitted: 2026-11-10

Updated: 2026-11-10

Comments: 14 pages, 3 figures. Accepted at the AI4DD workshop (AI for Drug Discovery: Bridging the Translation Gap), NeurIPS 2026

License: http://arxiv.org/licenses/nonexclusive-distrib/1.0/

Importance score: 92/100

The gist: Generative models are increasingly used to propose molecules for drug discovery They can produce far more candidates than can be experimentally tested…

Key concepts

Exposure-Aware Evaluation
This method assesses a molecule not just on its potency, but on the dose needed to reach the target and whether it remains safe at that dose. It combines predicted molecular properties with PBPK simulations to estimate actual drug exposure after oral dosing, allowing for a more realistic comparison of candidates.
Physiologically Based Pharmacokinetic (PBPK) Simulations
These are computer models that simulate how a drug moves through the body, including absorption, distribution, metabolism, and excretion. They estimate the free (unbound) drug concentration over time after oral administration. This simulation is crucial for determining the minimum effective dose and assessing safety at that specific exposure.
Minimum Effective Dose (MED)
The MED is the lowest oral dose required to keep the drug concentration above a certain threshold needed for target engagement throughout the dosing interval. It depends on both how potent a molecule is and its pharmacokinetic properties like absorption and clearance, meaning potency alone isn't enough.
Static Selection vs. Exposure-Aware Evaluation
Static selection ranks molecules based only on predicted molecular properties like potency and synthetic accessibility. Exposure-aware evaluation uses PBPK data to re-rank these candidates, revealing that static rankings can miss superior molecules because they don't account for how dose and exposure jointly determine success.

Summary

Generative models are increasingly used to propose molecules for drug discovery They can produce far more candidates than can be experimentally tested, making computational selection essential This work investigates this gap by using physiologically based pharmacokinetic (PBPK) simulations to re-evaluate molecules selected from three PARP1-targeted generative pipelines Of 60 shortlisted molecules, seven fail to achieve sufficient target exposure, while 28% of the remaining candidates show predicted safety risks at their required dose Moreover, 85% of candidates with an achievable dose are outperformed by molecules overlooked during static selection These findings suggest that accounting for dose and exposure can substantially change which generated molecules are considered promising for drug development. p. 0

Exposure-Aware Evaluation Methodology

We evaluate generated molecules based on the dose needed to achieve sufficient target exposure and the associated safety margins For each candidate, we combine predicted molecular properties with physiologically based pharmacokinetic (PBPK) simulations to estimate drug exposure following oral administration We then determine the dose required for target engagement and assess safety at that dose, accounting for uncertainty in both the molecular predictions and simulated exposure The minimum effective dose (MED) is determined by searching for the minimum oral dose at which the simulated free (unbound) drug concentration remains above this threshold during the dosing interval This MED depends on both potency and pharmacokinetics, as even a highly potent molecule may require a large dose if it is poorly absorbed or rapidly eliminated Safety at the required dose is assessed using the exposure-adjusted safety margin Mj = IC50j / Cj (d∗) This allows us to compare candidates based on their predicted behavior at a relevant dose.

Pipeline Diversity and Static Selection

We investigate three PARP1-targeted generative pipelines that differ in how they generate and optimize molecules These pipelines include MOOD, which is a docking-guided diffusion model without explicitly considering ADME during generation, REINVENT4-MPO, which is fine-tuned through reinforcement learning to optimize potency and ADME properties, and Mol2Mol, a seed-based approach that generates analogues of existing molecules Both REINVENT4-MPO and Mol2Mol are implemented using the REINVENT4 framework. All generated molecules undergo the same filtering and ranking procedure, which involves applying hard filters for drug-likeness, synthetic accessibility, physicochemical properties, predicted ADME, structural alerts, mutagenicity, and predicted PARP1 activity The surviving molecules are then ranked using a weighted geometric mean (WGM) of static molecular desirability terms. p. 2

Discrepancies Between Static Ranking and Exposure-Aware Outcomes

We demonstrate that accounting for dose and exposure can change candidate selection Among the 53 shortlisted molecules with an achievable dose, 45 are outperformed in dose requirement and both safety margins by at least one molecule overlooked by the static ranking This suggests that improving molecular generation alone may have limited benefit if candidate ranking does not account for how potency and pharmacokinetics jointly determine exposure at the dose needed for treatment For instance, predicted hERG risk is essentially uncorrelated with its exposure-adjusted margin across all 53 candidates

Limitations and Future Directions

Full PBPK simulation is too expensive to apply indiscriminately to large generated libraries A practical workflow could use static filters to narrow the candidate pool, exposure-aware estimates to guide ranking, and detailed PBPK simulations for final evaluation Simply adding clearance or protein binding to a static score may not be sufficient because their effects depend on other molecular properties and the required dose These findings remain model-based rather than experimentally validated, as predicted potency thresholds and PBPK inputs are uncertain The RED, ORANGE, and GREEN categories indicate predicted risk, not experimentally established toxicity or safety.

Dose Requirement Variability

The minimum effective dose (MED) varies substantially within and across pipelines The median MED is 1,785 mg/day for Mol2Mol (range 1.1–22,658), 2,340 mg/day for REINVENT4-MPO (range 310–15,326), and 15,487 mg/day for MOOD (range 2,357–34,793) This variation indicates that molecules selected using similar static criteria can require very different doses

Uncertainty in Safety Assessment

Our safety assessment accounts for uncertainty in both predicted off-target potency and simulated drug exposure We use endpoint-specific PBPK errors because hERG and PARP2 are evaluated using different parts of the concentration–time profile For off-target potency, we use the held-out RMSE of the corresponding Chemprop model as a source of uncertainty The RED, ORANGE, and GREEN categories describe risk under our predictive models, not experimentally established toxicity or safety.

Dominance of Unselected Molecules

We found that 45 of the 53 candidates (85%) are outperformed by at least one unselected molecule These alternatives are not simply repeated variants of compounds already represented in the shortlist Every confirmed dominator has a Murcko scaffold absent from its corresponding shortlist, indicating structural diversity that the static shortlist does not capture.

Static Threshold Limitations

A simple risk threshold based on off-target potency cannot fully replace evaluation at the required exposure, particularly when predictions are uncertain A fixed threshold would miss some candidates with high exposure-related risk while excluding others with more favorable predicted safety margins This illustrates why a single potency threshold cannot fully replace evaluation at the required exposure, particularly when predictions are uncertain.

Individual Parameter Insights

The seven candidates without an achievable MED tend to have lower unbound fractions and higher clearance than candidates with an achievable dose Neither property consistently falls outside a plausible drug-like range, suggesting that their combined effects limit the free concentration that can be maintained between doses The raw hERG IC50 is not monotonic across risk tiers, showing that exposure-adjusted risk is not explained by unusually strong predicted hERG inhibition alone

Conclusion on Selection Strategy

The results demonstrate that static selection can differ substantially from exposure-aware evaluation The usefulness of static off-target predictions depends on how well they reflect risk at the exposure required for target engagement Future work should examine whether this disagreement persists across other targets, scoring strategies, and experimental measurements.

Table 2 Summary

Table 2 reports pooled Spearman rank correlations between PARP1 potency and MED, hERG risk and its exposure-adjusted safety margin, and PARP2 risk and its corresponding margin PARP1 potency is strongly associated with MED (ρ = −0.81), as more potent molecules generally require lower doses The two safety endpoints show different patterns, with predicted hERG risk being essentially uncorrelated with its exposure-adjusted margin (ρ = −0. p. 303) and PARP2 risk showing a strong negative correlation (ρ = −0. p. 370).

Table 4 Insights

Table 4 shows the MED and predicted off-target IC50 values for the 15 candidates classified as RED, where bold indicates the endpoint responsible for the RED classification For example, six of these 13 hERG-RED candidates exceed the 10 µM hERG threshold and would therefore appear less concerning under that static criterion. The raw hERG IC50 is not monotonic across risk tiers, as evidenced by the difference between GREEN and RED candidates. p. 2

Figure 3 Interpretation

Figure 3 illustrates the difference using three MOOD candidates with similar predicted hERG potency but different exposure-aware classifications The exposure-aware evaluation reverses the interpretation of a static threshold, showing that RED exceeds the central IC50 while GREEN remains below it even after accounting for uncertainty.

Figure 2 Interpretation

Figure 2 shows substantial variation within each pipeline, indicating that molecules selected using similar static criteria can require very different doses The median MED is substantially higher for MOOD (range 2,357–34,793) compared to Mol2Mol (range 1.1–22,658).

Figure 1 Interpretation

Figure 1 illustrates two limitations revealed by PBPK simulation: a compound that fails to reach its PARP1 potency threshold even at an extreme test dose, and a compound whose exposure at its MED exceeds the predicted hERG threshold

Table 5 Summary

Table 5 provides ADMET-AI-predicted physicochemical and pharmacokinetic parameters for the 15 RED candidates These parameters, such as logP and clearance, are used to parameterize the PBPK model which determines how the PBPK model represents absorption, distribution, and elimination. The ADMET-AI predictions supply the inputs to our own PBPK simulations. p. 7 p. 7

Table 5 Detailed Data

Table 5 lists the logP, logS, Perm., fu, and CL for the 15 RED candidates across different models For example, MOOD P1 007 has a logP of 1. p. 853 and a clearance (CL) of 4. p. 00. The ADMET-AI predictions supply the inputs to our own PBPK simulations.

Table 4 Detailed Data

Table 4 shows the MED and predicted off-target IC50 values for the 15 candidates classified as RED, showing that raw potency alone does not reliably distinguish exposure-adjusted risk categories For instance, the median PARP2 IC50 decreases from GREEN (3.00 µM) to ORANGE (0. p. 498 µM) to RED (0. p. 426 µM). The raw hERG IC50 is not monotonic across risk tiers, as seen in the median values for GREEN, ORANGE, and RED candidates. p. 2

Appendix A Details

Appendix A details the dose-search procedure used to determine the minimum effective dose (MED) We begin with a single PK-Sim simulation at a reference dose of 70 mg and use linear scaling to estimate an initial dose Each subsequent iteration runs a new oral PBPK simulation, estimates the terminal half-life, and constructs the concentration–time profile after repeated dosing We then compare the free steady-state trough concentration, Css,min,free, with the predicted PARP1 IC50

Appendix B Details

Appendix B describes how safety assessment accounts for uncertainty in both predicted off-target potency and simulated drug exposure We use endpoint-specific PBPK errors because hERG and PARP2 are evaluated using different parts of the concentration–time profile A candidate is classified as RED if its central simulated exposure exceeds the central predicted off-target IC50, ORANGE if the central exposure is below the threshold but uncertainty intervals overlap, and GREEN if the intervals remain separated.

Appendix H Details

Appendix H provides correlations and examines whether the raw off-target potency predictions reveal these risks It shows that the disagreement arises from how pharmacokinetic properties combine with target and off-target potency to determine exposure and safety at the required dose The analysis is limited to PARP1 and three generative pipelines examined as independent case studies rather than a controlled comparison.

Appendix I Details

Appendix I describes how we screen unselected molecules by running a full individualized MED search for every filtered molecule Of these, 5 confirmed dominators from MOOD, 108 from REINVENT4-MPO, and 12 from Mol2Mol Overall, 45 of the 53 candidates (85%) are outperformed by at least one unselected molecule.

Appendix G Details

Appendix G describes how we use Chemprop directed message-passing neural networks trained on experimental bioactivity data to predict PARP1, hERG, and PARP2 bioactivity These models provide the potency thresholds used for target engagement and exposure-adjusted safety assessment

Appendix F Details

Appendix F details the predicted PARP1, hERG, and PARP2 bioactivity using Chemprop directed message-passing neural networks trained on experimental bioactivity data The held-out root mean squared error (RMSE) is 0.79 pIC50 units for PARP1, 0. p. 972 for hERG, and 0. p. 289 for the expanded PARP2 dataset. p. 9

Appendix E Details

Appendix E describes the scoring functions and weights used in the static ranking procedure Predicted PARP1 potency receives the highest weight, 2.0, while CYP3A4 risk serves as a proxy for metabolic drug–drug interaction liability. p. 9 The geometric mean penalizes candidates with poor values for individual properties more strongly than an arithmetic mean

Appendix D Details

Appendix D details the hard filtering criteria used to remove candidates with unfavorable predicted properties before static ranking These filters include criteria such as QED ≥ 0.5, synthetic accessibility score ≤ 6. p. 30, and predicted PARP1 pIC50 ≥ 6. p. 70.

Appendix C Details

Appendix C provides details on the implementation of the three generative pipelines MOOD uses a diffusion backbone pretrained on ZINC250k and is guided by the PARP1 docking-score regressor REINVENT4-MPO fine-tunes its recurrent neural network through reinforcement learning using a multi-parameter reward Mol2Mol uses REINVENT4’s molecule-tomolecule generation mode to explore chemical neighborhoods around existing molecules.

Appendix G describes how we use Chemprop directed message-passing neural networks trained on experimental bioactivity data to predict PARP1, hERG, and PARP2 bioactivity These models provide the potency thresholds used for target engagement and exposure-adjusted safety assessment.

Appendix B describes how safety assessment accounts for uncertainty in both predicted off-target potency and simulated drug exposure A candidate is classified as RED if its central simulated exposure exceeds the central predicted off-target IC50, ORANGE if the central exposure is below the threshold but uncertainty intervals overlap, and GREEN if the intervals remain separated.

Appendix I describes how we screen unselected molecules by running a full individualized MED search for every filtered molecule Overall, 45 of the 53 candidates (85%) are outperformed by at least one unselected molecule. The alternatives are not simply repeated variants of compounds already represented in the shortlist. p. 4

Green — the paper says it, and the page is printed. Red — no passage carries it, so treat that sentence as unverified. The same check runs on your own text: check a summary against its source, $5.

Improvements for AI systems

  1. Exposure-aware evaluation can substantially change which generated molecules are considered promising for drug development because accounting for dose and exposure can substantially change which generated molecules are considered promising (Abstract). The improved system will move beyond relying solely on predicted molecular properties by integrating physiologically based pharmacokinetic (PBPK) simulations to estimate drug exposure at the required dose.

  2. The system can identify molecules that fail to reach sufficient target exposure under dosing constraints, as demonstrated by the finding that seven fail to achieve sufficient target exposure among 60 shortlisted molecules (Results). This allows for a more realistic assessment of candidate viability than static selection alone.

  3. The system will incorporate a mechanism to select candidates based on dose and safety margins rather than just static multi-parameter scores, addressing the gap where predicted molecular properties do not always translate into favorable exposure (Abstract). Specifically, it will use the exposure-adjusted safety margin defined as Mj = IC50j / Cj (d∗) (Section 2).

  4. The improved system can be designed to prioritize molecules that have a lower minimum effective dose and larger safety margins, as an unselected molecule dominates a shortlisted candidate if it requires a lower MED and has larger safety margins for both hERG and PARP2 (Results). This addresses the finding that 45 of the 53 shortlisted molecules with an achievable MED are outperformed in dose requirement and both safety margins by at least one molecule overlooked by the static ranking (Abstract).

  5. The system can utilize a two-stage screening procedure to reduce computational costs, first evaluating unselected molecules at a fixed dose of 70 mg before running the full individualized minimum effective dose search (Section 14). This balances thoroughness with scalability for large libraries.

  6. The safety classification will move beyond single-threshold potency cutoffs by employing a risk category system where RED if its central predicted exposure exceeds the offtarget potency threshold, ORANGE if the central exposure is below the threshold but the uncertainty intervals overlap, and GREEN if the intervals remain separated (Section 2). This accounts for uncertainty in both predicted off-target potency and simulated exposure.

  7. The system will incorporate a diversity constraint during ranking by selecting the highest-ranked candidate from each Murcko scaffold before allowing a scaffold to contribute another molecule (Section 10). This ensures that the shortlist includes structural diversity, as confirmed by the finding that every confirmed dominator has a Murcko scaffold absent from its corresponding shortlist (Section 14).

Related papers