Beyond Proxy Objectives in De Novo Molecular Design

18 concepts6 papers1 proposal

Imagine trying to design a key for a highly secure, complex lock. If your only metric for success is whether the key is made of brass and weighs roughly ten grams, you can easily manufacture thousands of "successful" keys. However, when you take them to the lock, none of them will turn.

For years, the field of machine learning for de novo molecular design has operated under a similar illusion. Generative models have achieved seemingly spectacular results by optimizing simple mathematical formulas like logP (octanol-water partition coefficient) and QED (Quantitative Estimate of Drug-likeness). Yet, when these AI-generated molecules are scrutinized by chemists, they frequently turn out to be unstable, impossible to synthesize, or biologically inert.

A growing body of research is exposing the limitations of these oversimplified proxy objectives and shifting the field toward benchmarks that capture the true, rugged complexity of physical chemistry and biology.

The Trivial Optimization Trap

The fundamental issue with early molecular design benchmarks is that they lack discriminative power: they are too easy to solve.

In A Fresh Look at De Novo Molecular Design Benchmarks, Austin Tripp et al. systematically evaluated standard molecular design tasks. They discovered that simple baseline algorithms could optimize properties like logP and QED just as effectively as highly sophisticated deep generative models. This echoed earlier findings by Nathan Brown et al. in GuacaMol: Benchmarking Models for de Novo Molecular Design, which noted that many standard benchmarking tasks failed to differentiate advanced generative architectures from basic virtual screening methods.

When a benchmark can be maxed out by a simple heuristic or a random search, it ceases to be an effective tool for measuring algorithmic progress. It creates a false sense of accomplishment while masking the true capabilities—or limitations—of advanced machine learning models.

Why Proxy Objectives Fail

Why do these proxy functions fail to translate into real-world success? The answer lies in the topology of the optimization landscape.

Oversimplified Proxy Landscape (e.g., QED, logP)
Smooth, continuous, easily exploitable by simple baselines
▲
│     ┌─┐
│    ┌┘ └┐
│   ┌┘   └┐
└───┴────┴─────────────────► Chemical Space

Real-World Biochemical Landscape
Rugged, sparse, multi-modal, constrained by physics and synthesis
▲
│       ║       ║
│   ┌┐  ║  ┌┐   ║
│  ┌┘└┐ ║ ┌┘└┐  ║  ┌┐
└──┴──┴─╩─┴──┴──╩──┴───────► Chemical Space

As Austin Tripp et al. argue in An evaluation framework for the objective functions of de novo drug design benchmarks, real drug discovery landscapes are incredibly rugged. They feature infrequent optima, multiple distinct solution modes, and intricate structure-activity relationships. Proxy functions smooth out this reality into a continuous, easily exploitable hill. Generative models quickly learn to exploit the mathematical loopholes of these formulas, generating bizarre, unrealistic molecular structures that maximize the score without fulfilling the actual underlying objective.

Furthermore, these heuristic metrics suffer from saturation. In Tartarus: A Benchmarking Platform for Realistic And Practical Inverse Molecular Design, AkshatKumar Nigam et al. point out that beyond a certain threshold, higher scores in water solubility or drug-likeness cease to correlate with actual physical performance. A proxy objective is only valuable if it closely mirrors the intrinsic, real-world goal. Once that correlation breaks down, optimizing the proxy becomes a meaningless exercise.

Grounding Evaluation in Physical Reality

To bridge the gap between benchmark performance and real-world utility, the research community is replacing toy metrics with high-fidelity, multi-dimensional evaluation frameworks. This transition is happening across three main fronts:

1. Standardizing Real-World Constraints

Platforms like the Therapeutics Data Commons (TDC), introduced by Kexin Huang et al. in Therapeutics Data Commons: Machine Learning Datasets and Tasks for Drug Discovery and Development, provide AI-ready datasets that force models to grapple with complex, practical constraints. Instead of optimizing a single property, models must account for a suite of real-world challenges, including a molecule's actual synthesizability and safety profiles.

2. Simulating Physical Interactions

Instead of relying on static, property-based formulas to guess whether a molecule might make a good drug, newer benchmarks evaluate how a molecule physically interacts with its biological target. In Generative Models Should at Least Be Able to Design Molecules That Dock Well: A New Benchmark, Tobiasz Cieplinski et al. proposed using molecular docking simulations—calculating the 3D binding energy between a generated molecule and a target protein—as a baseline requirement. When evaluated on this realistic task, many popular generative models failed to outperform simple chemical databases, highlighting how severely older proxies overestimated model capabilities.

3. Domain-Specific Physics Simulations

Moving beyond drug discovery, the Tartarus platform (Nigam et al.) introduces standardized workflows for materials science, such as solar cells and OLEDs, using true physics-based simulations. Their work demonstrates that generative models perform vastly differently depending on the specific chemical domain, proving that a single, generalized proxy cannot predict real-world utility.

The Path Forward

The shift toward harder, more realistic objectives fundamentally changes how generative models must be designed. As Alexia Jolicoeur-Martineau et al. note in Any-Property-Conditional Molecule Generation with Self-Criticism using Spanning Trees, future architectures can no longer rely on easily computable shortcuts. They must be built to handle computationally intensive, biologically significant properties directly within their generation loops.

By abandoning oversimplified proxies and embracing the messy, rugged reality of physical chemistry, the field is moving away from gaming metrics and toward discovering genuine scientific solutions.

Key Foundational Work

A Fresh Look at De Novo Molecular Design Benchmarks (2021) — contributed 4 concept elements to this direction.
GuacaMol: Benchmarking Models for de Novo Molecular Design (2019) — contributed 2 concept elements to this direction.
Therapeutics Data Commons: Machine Learning Datasets and Tasks for Drug Discovery and Development (2021) — contributed 1 concept elements to this direction.
History & citationVersion 1 · 2026-10-04

Current · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_oversimplified_objective_functions_cbff370d,
  title = {oversimplified objective functions},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/oversimplified-objective-functions/rdr_cbff370d0d16426ebf2966e983be5996},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-19},
}

This exact version (version 1, frozen text):

@misc{lacuna_oversimplified_objective_functions_cbff370d_v1,
  title = {oversimplified objective functions},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/oversimplified-objective-functions/rdr_cbff370d0d16426ebf2966e983be5996/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-19},
}
About

These directions overlap with this one through shared papers.

See something wrong? Report a mistake