Negative Sampling

Negative samples must mirror the chemical property distribution of the positive set to prevent the model from exploiting superficial biases.

• Property-matched decoys: Select non-binders with similar molecular weight, logP, and hydrogen bond donor/acceptor counts as active ligands
• Avoids trivial learning: Prevents the classifier from simply predicting 'non-binding' for all high-molecular-weight or highly hydrophobic molecules
• DUD-E methodology: The Directory of Useful Decoys—Enhanced provides a rigorous framework for generating property-matched negatives for benchmarking

Without this matching, a model can achieve artificially high AUC scores by learning physicochemical shortcuts rather than true binding interactions.

Hard negatives are non-binding pairs that the current model confidently misclassifies as positive, residing close to the decision boundary.

• Iterative retraining: After an initial training epoch, mine the top-scoring false positives and add them to the negative set for the next epoch
• Contrastive learning integration: Frameworks like TransformerCPI use hard negatives to sharpen the latent space, forcing the model to learn subtle discriminative features
• Bootstrapping: This process progressively refines the boundary between true binders and structurally similar non-binders

Hard negative mining is essential for distinguishing between a true ligand and a close structural analog that lacks a critical binding interaction.

The sampling strategy for selecting negatives from the vast unlabeled chemical space dramatically impacts model calibration.

• Random sampling from screened libraries: Select negatives uniformly from compounds experimentally confirmed as inactive in high-throughput screening assays
• Stratified sampling by protein family: Ensure negatives are drawn proportionally across kinases, GPCRs, and nuclear receptors to prevent target-class bias
• Temporal splitting: For time-split validation, all negatives must be selected from compounds screened before a specific cutoff date to avoid data leakage

Random sampling from the full chemical universe often yields negatives that are trivially distinguishable, inflating performance metrics without real-world utility.

A 'negative' label is not a universal truth but a function of the assay threshold and experimental conditions.

• Activity cutoff calibration: A compound with an IC50 of 15 µM may be a negative in a 10 µM cutoff assay but a weak positive in a 30 µM screen
• Assay interference artifacts: Compounds flagged as PAINS or aggregators must be explicitly excluded from the negative set, as their non-interaction is due to assay interference, not true biological inactivity
• Dose-response confirmation: Single-concentration primary screens produce noisy negatives; confirmed negatives from dose-response curves provide a higher-fidelity training signal

Treating all unconfirmed inactives as equivalent negatives introduces systematic label noise that degrades the model's upper performance bound.

When training with pairwise or listwise ranking objectives, the negative sampling strategy directly defines the optimization landscape.

• Bayesian Personalized Ranking (BPR): For each positive drug-target pair, sample a fixed number of negatives and optimize the margin between their predicted scores
• LambdaRank with negative sampling: For each query protein, sample a subset of non-binding drugs to approximate the full listwise loss gradient efficiently
• In-batch negatives: In contrastive frameworks, treat all other drug-target pairs within a mini-batch as negatives, dramatically increasing the effective negative count without additional sampling overhead

These strategies shift the model's objective from binary classification to learning a relative ordering of binding affinities.

The 'closed-world assumption'—that all unlabeled pairs are negative—is a fundamental flaw in DTI modeling that must be explicitly addressed.

• PU learning (Positive-Unlabeled): Train classifiers that treat the negative set as unlabeled, learning from positive examples and the unlabeled mixture without assuming all unlabeled data is negative
• Spy technique: Insert known positives into the unlabeled set as 'spies' to estimate the reliable negative proportion before training
• Self-paced learning: Begin training with only the most confidently predicted negatives, gradually introducing harder examples as the model stabilizes

These approaches acknowledge that many 'negatives' in standard benchmarks are simply untested pairs that may, in reality, exhibit binding.

A machine learning paradigm that simultaneously uses descriptors from both the ligand chemical space and the target protein sequence space to predict bioactivity across large interaction matrices. This approach explicitly models the non-interacting pairs as negative examples.

• Kernel-based PCM: Uses similarity kernels for both ligands and proteins
• Deep PCM: Neural networks learning joint embeddings from SMILES and protein sequences
• Interaction matrix: Rows are proteins, columns are ligands, cells are binding affinities

Negative sampling in PCM requires careful selection of non-interacting pairs from the vast combinatorial space, as random pairing creates an extreme class imbalance that can overwhelm the learning signal.

The quantitative strength of the non-covalent interaction between a biomolecule and its ligand, expressed via thermodynamic dissociation constants (Kd) or inhibition constants (Ki). This continuous value defines the boundary between positive and negative examples in DTI classification.

• Typical threshold: Kd < 10 µM considered a binding interaction
• Units: Molar concentration (M), with lower values indicating stronger binding
• Range: From picomolar (10⁻¹² M) for tight binders to millimolar (10⁻³ M) for weak interactions

Setting the affinity cutoff for binary classification is a critical negative sampling design decision—too strict and weak true binders become false negatives; too lenient and non-binders contaminate the positive set.

A binary or count-based vector representation encoding the specific intermolecular contacts between a protein's residues and a bound ligand. These fingerprints serve as ground truth for validating whether predicted negative pairs truly lack binding interactions.

• Bit types: Hydrogen bonds, hydrophobic contacts, pi-stacking, salt bridges, halogen bonds
• Residue-level: Each bit corresponds to a specific amino acid position
• Interaction type-level: Bits represent the presence of any contact of a given type

PLIFs help identify subtle false negatives—compounds that dock successfully but are labeled non-binders due to assay limitations rather than genuine lack of affinity.