Active Learning for Screening

The foundational workflow is a closed loop where a model's uncertainty drives data acquisition. The cycle repeats: a model is trained on a small, labeled seed set; it predicts activity for a massive unlabeled pool; an acquisition function selects the most informative compounds; these are experimentally tested; and the new data is fed back to retrain the model. This process continues until a stopping criterion, such as a hit rate plateau or budget exhaustion, is met.

The mathematical core of active learning, defining the strategy for selecting compounds. Key functions include:

• Uncertainty Sampling: Picks compounds where the model's prediction confidence is lowest (e.g., highest entropy).
• Query-by-Committee (QBC): Selects compounds where an ensemble of diverse models most disagrees, reducing model bias.
• Expected Model Change: Chooses compounds that would cause the largest gradient update to the model if their label were known.
• Exploitation-Exploration Trade-off: Balances selecting compounds predicted to be highly active (exploitation) with those from unexplored regions of chemical space (exploration).

Active learning requires an initial labeled dataset to train the first model. The cold start problem refers to the challenge of selecting this seed set without any prior data. Strategies include:

• Diversity-Based Selection: Picking a maximally diverse set of compounds using clustering or MaxMin algorithms on molecular fingerprints.
• Random Sampling: A simple but often surprisingly effective baseline.
• Domain Expert Curation: Manually selecting compounds known to be active or representative of the target class. A poor seed set can bias the entire subsequent exploration.

In practice, experimental validation is performed in parallel batches, not one compound at a time. Batch mode active learning selects a set of k compounds simultaneously. The challenge is to avoid selecting redundant, highly similar compounds. This is addressed by incorporating a diversity constraint into the acquisition function, ensuring the batch collectively maximizes both informativeness and chemical diversity, often using greedy or submodular optimization.

The active learning framework is independent of the underlying predictive model. The surrogate model can be any regressor or classifier, such as a Random Forest, Gaussian Process, or Graph Neural Network. Gaussian Processes are particularly well-suited because they provide principled uncertainty estimates directly from their predictions, making them a natural fit for uncertainty sampling without needing separate calibration methods.

An automated workflow requires a defined endpoint. Common stopping criteria include:

• Hit Rate Saturation: The rate of discovering new active compounds drops below a predefined threshold.
• Budget Exhaustion: A fixed number of experimental assays or computational hours is reached.
• Model Convergence: The model's performance on a held-out validation set plateaus.
• Top-N Recall: The process stops when a target percentage of all known actives in the library has been discovered.

A core query strategy where the model selects the unlabeled compound for which it is least confident in its own prediction. This directly targets the model's decision boundary.

• Least Confidence: Selects the instance with the lowest predicted probability for its most likely class.
• Margin Sampling: Selects the instance with the smallest difference between the top two predicted class probabilities.
• Entropy Sampling: Selects the instance with the highest information entropy across all possible class predictions.

An ensemble-based active learning strategy that maintains a committee of diverse models trained on the same labeled data. The next compound for evaluation is chosen based on the maximum disagreement among the committee members.

• Vote Entropy: Measures the disagreement by the entropy of the class votes.
• Kullback-Leibler Divergence: Measures the average difference between the label distribution of any one committee member and the consensus.
• QBC is highly effective for exploring chemical space by leveraging model diversity.

A selection criterion that prevents the redundancy of queries by ensuring the chosen batch of compounds is structurally diverse. It avoids the pitfall of selecting ten nearly identical high-uncertainty molecules.

• Often implemented using k-means clustering on molecular fingerprints.
• MaxMin algorithms select compounds that maximize the minimum distance to any already-selected compound.
• Balances the exploitation of uncertainty with the exploration of chemical space.

A probabilistic framework often used for active learning in lead optimization, where the goal is to find the global optimum of a property (e.g., binding affinity) with minimal experiments.

• Uses a surrogate model, typically a Gaussian Process, to model the objective function.
• An acquisition function (e.g., Expected Improvement) balances exploration of uncertain regions with exploitation of known high-performing regions.
• Naturally handles the exploration-exploitation trade-off in chemical space.

The encoding of a molecule's structure into a fixed-length binary vector, which serves as the primary feature representation for active learning models.

• Extended Connectivity Fingerprints (ECFP4): Circular fingerprints that encode atom neighborhoods, widely used for activity prediction.
• MACCS Keys: A predefined set of 166 structural keys.
• The choice of fingerprint directly impacts the definition of chemical similarity and model performance.