De Novo Molecular Generation Tasks

Overview

The primary task in our benchmark is de novo molecular generation, which involves generating novel molecular compounds while optimizing a set of desirable properties. This task is fundamental to drug discovery, where the goal is to identify candidate molecules that satisfy multiple criteria simultaneously.

In this setting, no high-scoring molecule is known a priori—the generative model must identify drug candidates with no prior information on which molecules are promising. The task is inherently multi-objective: generated candidates must simultaneously satisfy multiple criteria, such as:

Binding affinity: The ability to bind to a specific protein target
Drug-likeness: Compliance with rules that predict good pharmacokinetic properties
Synthetic accessibility: The feasibility of synthesizing the molecule in a laboratory

Dataset Statistics

Split	Number of Prompts	Description
Training	49,000	Multi-objective optimization prompts
Evaluation	1,000	Held-out prompts for validation (binding targets are OOD, i.e not present in training set)
Test	1,000	Final evaluation (binding targets are OOD, i.e not present in training set or evaluation set)

The dataset is well-balanced across optimization complexity:

Single-property optimization: 29% of prompts
Two-property optimization: 36% of prompts
Three-property optimization: 35% of prompts

Optimization Task Distribution

Docking Simulations

What is Molecular Docking?

Molecular docking animation

(Roy, Aritra (2021). Molecular Docking First Try.gif. figshare. https://doi.org/10.6084/m9.figshare.14676669.v1)

Molecular docking is a computational technique that predicts the preferred orientation of a small molecule (ligand) when bound to a protein target. The docking score quantifies the strength of this interaction—lower scores indicate better binding affinity.

Docking is a standard computational tool in drug discovery to evaluate how well a compound fits into a protein's binding pocket, making it an ideal objective for molecular generation tasks.

Protein Structure Sources

We extracted protein structures from two main sources:

SAIR Dataset

The SAIR dataset consists of over 1 million protein-ligand pairs from approximately 5,000 unique protein structures. This dataset was built upon the structural predictions of Boltz1, a co-folding AI model, and notably contains predicted protein structures for which no known experimental structure exists.

SIU Dataset

We complemented the SAIR structures with proteins from the SIU dataset, which provides labeled binding pockets.

Pocket Identification

In the SAIR dataset, binding pockets are not pre-defined. To identify binding pockets for each protein structure, we performed a clustering of the protein's residues based on the predicted receptor-ligand structure. The pipeline:

Filters structures to retain only high-quality ligand poses (top 50% by pIC50 potency and confidence score)
Identifies pocket residues by finding the closest protein residues to each ligand atom
Clusters pockets across multiple conformations using IoU-based similarity
Selects representative conformations by minimizing pairwise RMSD across pocket residues

After processing all structures with Meeko, we obtained 4,500 protein structures, each associated with a unique UniProt ID.

Description of Molecular Properties to be Optimized

In addition to docking scores, our benchmark includes 12 classical molecular properties computed using RDKit. These properties are essential for evaluating drug-likeness and pharmacokinetic characteristics.

Drug-likeness Properties

QED (Quantitative Estimate of Drug-likeness)

A composite score combining multiple molecular properties to estimate how "drug-like" a molecule is. QED ranges from 0 to 1, with higher values indicating better drug-likeness.

Typical objective: Maximize or keep above a threshold (e.g., > 0.5)
SA Score (Synthetic Accessibility)

Estimates how difficult a molecule is to synthesize, ranging from 1 (easy) to 10 (difficult). The score considers fragment contributions and complexity penalties.

Typical objective: Minimize or keep below a threshold (e.g., < 4)

Physicochemical Properties

LogP (Partition Coefficient)

The logarithm of the octanol-water partition coefficient, measuring lipophilicity. Optimal values for oral drugs typically range from 1 to 5 (Lipinski's Rule of Five).

Typical objective: Keep within a range or minimize (to improve solubility)
Exact Molecular Weight

The molecular mass in Daltons. Drug-like molecules typically have molecular weights below 500 Da.

Typical objective: Keep below a threshold (e.g., < 500 Da)
TPSA (Topological Polar Surface Area)

The surface area occupied by polar atoms (N, O, and their attached hydrogens). TPSA affects membrane permeability and oral absorption.

Typical objective: Keep below a threshold (e.g., < 140 Å² for oral bioavailability)

Structural Properties

Number of Hydrogen Bond Acceptors (NumHBA)

Atoms capable of accepting hydrogen bonds. Lipinski's Rule suggests ≤ 10 acceptors for oral drugs.
Number of Hydrogen Bond Donors (NumHBD)

Atoms capable of donating hydrogen bonds. Lipinski's Rule suggests ≤ 5 donors for oral drugs.
Number of Rotatable Bonds (NumRotatableBonds)

Bonds that allow free rotation, affecting molecular flexibility. Fewer rotatable bonds (≤ 10) generally improve oral bioavailability.
Number of Aromatic Rings (NumAromaticRings)

Count of aromatic ring systems. Too many aromatic rings can reduce solubility and increase toxicity risk.
Fraction of sp³ Carbons (FractionCSP3)

The ratio of sp³-hybridized carbons to total carbons. Higher values indicate more three-dimensionality, which is associated with better selectivity and fewer off-target effects.

Shape and Complexity Descriptors

Hall-Kier Alpha

A molecular connectivity index describing molecular shape and branching.
Phi (Flexibility Index)

A descriptor related to molecular flexibility derived from the Kier shape indices.

Property Distribution in Dataset

Properties are sampled with varying probabilities to encourage generation of feasible, drug-like molecules:

Property	Relative Frequency	Allowed Objectives
QED	High (7.0)	Maximize, Above threshold
SA Score	High (3.0)	Minimize, Below threshold
LogP	Medium (2)	Above or Below threshold, Minimize, Maximize
Phi	Medium (0.5)	Above or Below threshold, Minimize, Maximize
Exact Molecular Weight	Low (0.8)	Above or Below threshold
TPSA	Low (0.6)	Above or Below threshold
Other properties	Low (0.3-0.5)	Various

Reward Function

The reward for molecular generation is computed as the geometric mean of per-property rewards:

\[R(q, \hat{y}) = \left(\prod_{i=1}^{n} r_i\right)^{1/n}\]

Where \(r_i\) is the reward for the \(i\)-th property objective and \(n\) is the total number of properties. Using the geometric mean ensures that only completions effectively optimizing all properties are rewarded—a single per-property reward of 0 results in a total reward of 0.

Per-Property Reward Functions

The reward for each individual property depends on its optimization objective. We call \(\rho\) the normalized property value, scaled to [0, 1] of a generated molecule.

Maximize Objective

\[R = \rho \in [0, 1]\]

The reward is simply the normalized property value. Higher property values receive higher rewards.
Minimize Objective

\[R = 1 - \rho\]

The reward is the inverse of the normalized property value. Lower property values receive higher rewards.
Above Threshold

\[R = \begin{cases} 1 & \text{if } \rho \geq \text{threshold} \\ 0 & \text{otherwise} \end{cases}\]

Binary reward: full reward if the property meets or exceeds the threshold, zero otherwise.
Below Threshold

\[R = \begin{cases} 1 & \text{if } \rho \leq \text{threshold} \\ 0 & \text{otherwise} \end{cases}\]

Binary reward: full reward if the property is at or below the threshold, zero otherwise.

Invalid Molecules

If the generated SMILES string is invalid or cannot be parsed, the reward is automatically set to 0.