acceptodds
Withdrawn from ICLR 2027

Inhibitor Evolution: How Drug Design Evolved to Overcome Mutation Mechanisms

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Targeting the epidermal growth factor receptor (EGFR) in cells with tyrosine kinase inhibitors (TKIs) is an established strategy to block oncogenic phosphorylation cascades and tumor growth in non-small cell lung cancer (NSCLC). However, tumors frequently develop resistance through the T790M gatekeeper mutation, mitigating the effect of early-generation TKIs. Consequently, later generations of TKIs have been developed to combat these mutations. Our experiment computationally compared five TKIs spanning three drug generations (Erlotinib, Gefitinib, Afatinib, Dacomitinib, and Osimertinib) against wild-type and T790M mutant EGFR to provide an effective evaluation of TKI-generation evolution. Standardized molecular docking via AutoDock Vina was utilized for each TKI-receptor combination, revealing minimal free energy changes across TKI generation and receptor state (ΔΔG = −0.30 to +0.46 kcal/mol), demonstrating that docking scores fail to explain inhibitor generation efficacy differences. Consequently, Python Molecule 3D Rendering (PyMOL) and Protein-Ligand Interaction Profiling (PLIP) analysis determined Third-generation TKIs optimally increased gatekeeper clearance distance when met with the mutated receptor (3.07 Å to 4.31 Å) to remain in an ideal position to block unwanted cell signals while increasing hydrophobic contact (2 to 4 contacts) to block ATP and oncogenic phosphorylation cascades, which the first-generation TKIs failed to achieve. Ultimately, this study demonstrates the specific structural shifts that Third-generation TKIs utilize to remain more effective than first-generation TKIs when faced with mutated receptors.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
inhibitor, TKI, receptor, EGFR, PLIP, PyMOL, docking, generation, adaptation

Reviews

Among:
Shape:
1234this paper, Bottom 1%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.00 over 3 reviews, above 0% of the others.
Reviewer_K4vCOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

This paper investigates the structural differences in the binding of five EGFR inhibitors across three generations to wild-type and T790M mutant EGFR. The authors employ conventional molecular docking (AutoDock Vina), PLIP, and PyMOL to compare docking scores, binding poses, and interaction patterns. The results suggest that docking scores poorly reflect the experimentally reported differences in resistance across inhibitor generations.

Critical strengths

The paper addresses an important and practically relevant problem in drug discovery, namely the mechanisms underlying EGFR inhibitor resistance. If accepted, this work may help draw more attention to this important yet relatively underexplored research area.

Critical weaknesses
  • A more comprehensive dataset could be curated, for example, by including resistance mutations in kinases other than EGFR. More computational models could also be evaluated, particularly to investigate how existing CADD and AI-based models perform in predicting mutation-induced changes in inhibitor activity.

  • The contributions of this work are minor, particularly from the perspective of the machine learning community. I am not entirely convinced that the current findings align well with the scope and research themes of ICLR, as the study primarily relies on conventional computational chemistry tools and offers limited methodological or technical contributions to the field. Overall, the work seems more suitable for a medicinal chemistry or computational chemistry journal than for a machine learning conference such as ICLR.

  • Figure 1 is of insufficient resolution and should be provided at a higher DPI to improve readability.

Questions for authors

There are multiple PDB structures available for the same EGFR mutation, and the choice of receptor structure can significantly affect docking results. How did the authors select the PDB structures used in this study? Have they evaluated whether their findings are consistent across different experimentally determined structures of the same mutant?

Reviewer_CQ13OpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper examines how five EGFR tyrosine kinase inhibitors—erlotinib, gefitinib, afatinib, dacomitinib, and osimertinib—behave in docking calculations against wild-type and T790M-mutant EGFR. The authors use AutoDock Vina to estimate score changes across receptor states, compare these changes with literature IC50 ratios, and inspect top-ranked poses with PLIP and PyMOL. The reported docking-score shifts are small (−0.30 to +0.46 kcal/mol). The paper argues that gatekeeper clearance and hydrophobic contacts, particularly for osimertinib, explain resistance to earlier inhibitors better than docking scores do.

Critical strengths

1. A focused, inspectable comparison. The experiment covers ten inhibitor–receptor combinations using the same docking-box dimensions and search settings. Table 1 reports variation over 15 docking seeds rather than a single score per pair. This gives the descriptive comparison a useful degree of transparency.

2. The score–potency mismatch is worth examining. Presenting literature IC50 ratios beside docking-score changes (Table 2), and showing local interaction summaries and poses (Tables 3–4; Figures 3–4), makes the central interpretive difficulty visible. The paper can serve as a compact case study of why a docking score should not be treated as a direct measure of cellular efficacy.

Critical weaknesses

1. Static poses do not establish the proposed resistance mechanism. Section 5.4 correctly notes that Vina does not model Cys797 covalent bond formation, although this chemistry is central to the later-generation drugs discussed in Section 2. Nevertheless, Sections 5.2–5.3 infer water displacement, ATP exclusion, and suppression of signaling from top-ranked noncovalent poses, PLIP contact counts, and nearest-atom distances. The reported 3.67 Å erlotinib and 4.31 Å osimertinib distances in the mutant receptor (Figures 3–4) do not by themselves establish a steric clash in one case or productive inhibition in the other. Nor does the contact-count argument separate the generations: Table 4 reports an increase from 1 to 5 contacts for gefitinib, versus 2 to 4 for osimertinib. These observations generate a structural hypothesis, but they do not validate the claimed mechanism.

2. The study has neither a matched efficacy validation nor a clear learning contribution. Section 3.1 identifies the wild-type receptor as PDB 1M17 but does not identify the mutant structure, making it difficult to assess how receptor-state differences were controlled. Table 2 lists literature IC50 values without specifying the assay systems, underlying EGFR mutation backgrounds, or a source for each pair of values. As a result, the five-point correlation does not establish that the docking comparisons predict resistance: Figure 2 gives Pearson r = −0.12 with p = 0.85, while Section 5.1 reports p = 0.20 for the same r. More broadly, the workflow combines established docking and visualization packages without introducing or evaluating a new machine-learning method. This limits both the scientific validation.

Questions for authors

1. Which PDB entry was used for the T790M receptor, and what other EGFR mutations, if any, are present in that structure? How were receptor preparation and initial binding-site conformations made comparable to PDB 1M17?

2. For each IC50 pair in Table 2, what are the original sources, assay or cell systems, and mutation backgrounds? Were the wild-type and T790M values obtained under directly comparable conditions?

3. Are the gatekeeper distances and PLIP contact counts taken from a single best-scoring pose per drug–receptor pair? If multiple seed outputs were retained, how stable are these descriptors across the available poses?

4. Which atoms define the reported 3.67 Å and 4.31 Å contacts, and what atom-specific steric criterion supports interpreting the former as repulsive and the latter as compatible? Is any pre-reaction Cys797 geometry available for the covalent inhibitors?

Reviewer_uoCjOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

The manuscript compares five EGFR inhibitors across three generations: erlotinib, gefitinib, afatinib, dacomitinib, and osimertinib. Each compound is docked against wild-type and T790M-mutant EGFR using AutoDock Vina. The docking scores change very little between receptor states and correlate poorly with literature IC₅₀ ratios. The authors therefore examine the top-scoring poses using PLIP and PyMOL. They argue that osimertinib accommodates the mutant gatekeeper residue through greater atomic clearance and additional hydrophobic contacts, whereas earlier inhibitors experience steric repulsion or lose favorable contacts. They conclude that these structural features, rather than docking scores, explain third-generation inhibitor efficacy.

Critical strengths

Looking beyond a single docking score toward interpretable structural features is a reasonable idea.

The authors use multiple docking seeds, provide an anonymized code repository, and acknowledge some limitations of non-covalent docking.

Critical weaknesses

The paper introduces no learning method, dataset, benchmark, or broader insight for machine learning. It is a small computational-chemistry case study using established tools, with five compounds and only one third-generation inhibitor. The modeling omits a central part of the mechanism.

Afatinib, dacomitinib, and osimertinib are covalent inhibitors, yet Vina models only their pre-reaction, non-covalent poses. Correct positioning may be necessary for covalent binding, but it is not sufficient to explain potency or mutant selectivity. For osimertinib, covalent engagement of Cys797 is central to the binding mechanism. The conclusions therefore extend well beyond what the simulations capture. The structural interpretation is insufficiently validated.

No redocking or crystallographic-pose recovery is reported. All PLIP and distance conclusions rely on single top-scoring poses, without assessing pose variability. A nearest-atom distance alone cannot demonstrate repulsion, ATP accessibility, or improved inhibition; the proposed 3.8–4.0 Å and 5.0 Å thresholds are not justified and appear to ignore atom identity. The T790M PDB identifier is also missing, making it difficult to assess whether differences arise from the mutation or from unrelated differences between receptor structures.

Several results conflict with the narrative. Page 5 states that erlotinib clearance “shrinks to 3.67 Å,” whereas Section 5.3 reports an increase from 3.38 to 3.67 Å. Gefitinib gains more hydrophobic contacts than osimertinib, increasing from 1 to 5 versus 2 to 4. Describing the gefitinib result as “displacement” does not resolve this inconsistency without a defined displacement metric. These examples make the generational explanation appear post hoc.

Questions for authors

What is the T790M structure’s PDB identifier, and how were the wild-type and mutant structures matched for construct, conformation, and activating-mutation background?

Does the docking protocol recover known co-crystallized poses? How sensitive are the interaction counts and clearance distances to alternative poses and seeds?

How can the authors attribute generational efficacy to spatial accommodation without modeling covalent binding or reaction kinetics?

What are the exact sources and experimental contexts of the IC₅₀ values in Table 2?

What quantitative evidence distinguishes “binding optimization” from “displacement,” particularly for gefitinib’s increase from one to five hydrophobic contacts?

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Discussion (0)

Sign in to comment.