Quantitative Estimate of Drug-Likeness

The foundational heuristic for oral bioavailability that QED builds upon. A molecule is likely to have poor absorption if it violates two or more of these criteria:

• Molecular Weight > 500 Da
• LogP > 5
• Hydrogen Bond Donors > 5
• Hydrogen Bond Acceptors > 10 QED refines this binary pass/fail into a continuous probabilistic score by modeling the desirability functions of these and other descriptors against a reference set of known oral drugs.

The mathematical backbone of the QED metric. Each molecular descriptor (e.g., molecular weight, logP, HBD, HBA, TPSA, rotatable bonds) is transformed into a dimensionless desirability value between 0 and 1 using asymmetric double sigmoidal functions. A value of 1.0 indicates the descriptor falls within the optimal range observed in approved oral drugs. The final QED score is the geometric mean of all individual desirability scores, ensuring that a poor score in any single parameter significantly penalizes the overall assessment.

A drug-like molecule is useless if it cannot be synthesized. The Synthetic Accessibility (SA) Score complements QED by estimating the ease of laboratory synthesis based on:

• Fragment frequency in PubChem: common substructures lower the score
• Structural complexity: presence of non-standard ring systems, chiral centers, and macrocycles increases the score A balanced candidate exhibits both a high QED (drug-like) and a low SA Score (easy to make). Multi-objective optimization often treats these as competing objectives on a Pareto frontier.

Drug design is inherently a Pareto optimization problem. QED is rarely maximized in isolation; it must be balanced against potency, selectivity, metabolic stability, and synthetic tractability. Algorithms like NSGA-II or Bayesian optimization with scalarization functions evolve populations of molecules toward the Pareto frontier. A common composite objective is: Score = w1 * QED + w2 * Bioactivity - w3 * SA_Score This ensures generated molecules are simultaneously drug-like, active, and synthesizable.

QED serves as a topological constraint during the navigation of the vast chemical universe (estimated at 10^60 synthesizable molecules). Generative models like variational autoencoders or reinforcement learning agents use QED as a reward signal to bias exploration toward drug-like regions of latent space. Without such guidance, models waste computational resources generating chemically invalid, toxic, or non-drug-like structures that would never survive preclinical development.