ALF: An Active Learning Framework for Scientific Discovery

• Large combinatorial search spaces and costly measurements make exhaustive screening impractical, motivating targeted data acquisition under budget constraints.Source: §1 · p. 1
• Active learning iteratively fits a surrogate, scores unlabelled candidates for information value, obtains labels from an oracle, and adds them to training data.The same loop can support Bayesian optimisation by scoring candidates for their potential to maximise an unknown function.Source: §1 · p. 1
• Existing tools typically serve offline benchmarking or online oracle-based deployment, although both are parts of scientific data acquisition.Source: §1 · pp. 1–2
• ALF provides one API for offline and online campaigns, with five interchangeable, data-agnostic components spanning datasets, surrogates, acquisition, search, and oracles.The paper states that the shared API supports detailed benchmarking and rapid real-world deployment.Source: §1 · p. 2

• Ax targets continuous parameter and hyperparameter optimisation, while modAL provides model wrappers and queryable strategies for tabular data; neither supplies the surrounding multi-round data-handling and labelling loop.Source: §2 · p. 2
• FLEXS, MolPAL, and ALDE provide sequential active-learning loops but are limited to biological sequences or small-molecule screening, whereas ALF is domain-agnostic.Source: §2 · p. 2
• GAUCHE focuses on surrogate kernels, while TorchDrug and DeepChem lack acquisition-function and oracle abstractions; MADE offers composable closed-loop discovery but is material-specific.Source: §2 · p. 2
• The paper distinguishes ALF by combining modular components across the full discovery loop for use across modalities and in both offline and online settings.Source: §2 · p. 2

3.1 Framework

• Each DesignTask round fits a surrogate, scores candidates with an acquisition function, evaluates top-scoring candidates through an oracle, and adds their labels to the dataset.The State object tracks dataset splits, surrogate, round index, acquisition history, and recorded metrics.Source: §3.1 · p. 2

3.2 Components

• Datasets manage train, validation, test, and candidate-pool splits, while surrogates featurise and model labelled training data to predict candidate values.Source: §3.2 · p. 3
• Search functions populate each round's candidate pool, using a fixed dataset offline or proposing new candidates online through protocol or generative approaches.Provided search implementations include sequence, molecule, and materials methods.Source: §3.2 · p. 3
• Acquisition functions score candidates by estimated value of acquiring ground truth, and oracle classes return recorded labels offline or compute labels online.Source: §3.2 · p. 3
• Figure 2 illustrates an offline DesignTask using a pretrained ESM2 surrogate, with only search and oracle lines changed to switch online.Source: Figure 2 · p. 4

4.1 Experimental Setup

• Campaigns use protein, small-molecule, and materials design tasks, with 5,000 initial labels, 10 rounds, 25 acquisitions per round, a GP surrogate, and five random seeds.AL objectives report mean RMSE; BO objectives report the maximum score acquired so far.Source: §4.1 · p. 4

4.2 Offline Benchmarking

• Offline benchmarking evaluates seven acquisition strategies for active learning and Bayesian optimisation using an existing dataset as both the initial training data and candidate pool.Figure 3 shows the small-molecule task; protein and materials experiments are reported in Appendix C.Source: §4.2 · p. 4, Figure 3 · p. 4

4.3 Online Data Acquisition

• Online campaigns run the same seven strategies unchanged, using an RDKit oracle and single-edit SMILES search for small molecules.Figure 4 presents the small-molecule results; protein and materials experiments use PyRosetta and MACE oracles, respectively.Source: §4.3 · p. 5, Figure 4 · p. 5

4.4 Impact

• In reactive-chemistry interatomic-potential work, an NTK acquisition signal obtained lower energy and force error per round than prior state-of-the-art methods.Source: §4.4 · p. 5
• In protein sequence design, fine-tuning a protein language model with a ranking-based loss on model likelihoods retrieved higher-scoring candidates per round than traditional MSE losses.Source: §4.4 · p. 5

5 Conclusion

• ALF combines offline validation and online oracle-based campaigns through one API and five extendible, data-agnostic components.The paper compares acquisition strategies under active learning and Bayesian optimisation across protein sequences, small molecules, and materials.Source: §5 · p. 5
• ALF is open source, and its authors hope it will make deciding what data to acquire next easier across scientific domains.Source: §5 · p. 6

A.1 Package structure

• ALF separates lightweight interfaces and tasks in alf-core from concrete implementations in alf-tools, whose modality-specific dependencies are optional extras.alf-core avoids a machine-learning framework dependency; alf-tools uses dependencies such as torch, gpytorch, and botorch.Source: §A.1 · p. 10

A.2 Base Classes

• The component documentation describes extension requirements and base-class functionality; Oracle and Surrogate are configured as thin wrappers rather than subclassed.A single BaseModel can serve as a surrogate, oracle, or generative search model through its fit/predict or sample role.Source: §A.2 · p. 10, Table 1 · p. 10

A.3 ALF Tasks

• DesignTask runs an ask/tell active-learning loop, while SupervisedTask fits once and ZeroShotTask evaluates a pretrained surrogate without training.DesignTask asks for a batch, labels it with the oracle, refits the surrogate, logs metrics each round, and writes an aggregate summary.Source: §A.3 · p. 10

A.4 Offline to Online

• Moving a campaign online changes only its search function and oracle; offline search draws from a fixed pool with dataset lookup, while online search proposes candidates scored by a model.Source: §A.4 · p. 11

A.5 Extending ALF

• A custom acquisition function extends AcquisitionFunction by predicting candidates with the surrogate, computing utility scores, and returning scored candidates.Source: §A.5 · p. 11

A.6 Tutorials and Documentation

• ALF provides runnable offline and online tutorials, component how-to recipes, domain-dataset tutorials, and hosted API documentation and installation guidance.Source: §A.6 · p. 11

• The exact-inference GP uses a Matérn-5/2 kernel with ARD for molecules and materials, and a linear k-mer spectrum kernel without ARD for sequences.Source: B Experimental Details · p. 11
• 20 Adam iterations at learning rate 0.5 refit the GP each round; inputs are min-max normalized for molecules and materials, while labels are z-scored for fitting and inverse-transformed for evaluation.Sequence inputs are left unnormalized; reported RMSE is in original label units.Source: B Experimental Details · p. 11
• Each campaign acquires 25 candidates per round for 10 rounds, totaling 250 per cell across five seeds.Each seed fixes data splits, seed draw, and stochastic acquisition; experiments use Nvidia H100 GPUs.Source: B Experimental Details · p. 12
• Active-learning strategies are label-blind CoreSet and posterior-standard-deviation uncertainty sampling; Bayesian-optimisation strategies are greedy, UCB, expected improvement, and Thompson sampling.UCB uses α = 1.0; random uniform selection is an additional baseline.Source: B Experimental Details · p. 12
• Online search mutates sequences, SMILES strings, or material compositions by one domain-specific edit, using top-K training-set candidates as parents and removing candidates already in train or validation.Source: B Experimental Details · p. 12

• Information-based strategies stay close to random on active-learning RMSE, while random is strongest across domains; utility-based strategies acquire higher-scoring candidates than random on every modality.Source: Figure 5 · p. 12, C Offline Benchmarking Results · p. 12
• The figure plots test-set RMSE against acquisition round for active learning, where lower values are better.Source: Figure 5 · p. 12
• For Bayesian optimisation, the figure plots the best score acquired so far, where higher values are better; bands show means and standard deviations across five seeds.Source: Figure 5 · p. 12

D Online Data Acquisition Results

• Sequence active-learning RMSE does not fall, which may reflect single mutants’ near-identical k-mer spectra and the linear kernel’s limited separation within a batch.Molecule and materials batch RMSE falls over the campaign.Source: D Online Data Acquisition Results · pp. 12–13
• Every strategy improves the best scoring candidate under Bayesian optimisation, while active-learning batch RMSE falls for molecules and materials but not sequences.Figure 6 summarizes single online campaigns across sequence, molecule, and materials tasks.Source: D Online Data Acquisition Results · pp. 12–13, Figure 6 · p. 13

E Further Case Studies Results

• The neural-tangent-kernel signal reaches a given force error in fewer rounds than committee, distance-based, and random baselines, with the gap widening over the campaign.For protein design, ranking-loss fine-tuning retrieves more top sequences than MSE at every round; ProteinNPT and language-model embedding baselines pull ahead of one-hot encodings and random.Source: E Further Case Studies Results · p. 13, Figure 7 · p. 13
• Figure 7 presents multi-round campaigns from cited work, including interatomic-potential active learning and protein-design Bayesian optimisation.Panel (a) shows force error per round; panels (b) and (c) show top-candidate recall.Source: Figure 7 · p. 13, E Further Case Studies Results · p. 13

F Limitations

• The demonstrations use one exact-inference GP, one dataset per modality, and a 10-round budget of 25 candidates over five seeds; the authors make no claim that strategy orderings generalize to other settings.Five seeds support error bands but not rankings when curves overlap; exact GP inference scales cubically in the number of labels.Source: F Limitations · p. 14
• Online oracles are cheap computational stand-ins, and ALF assumes synchronous return of every batch label, leaving noise, attrition, and delayed feedback unmodelled.The real assays and simulations motivating the work may be noisy, unavailable for proposed candidates, or delayed across rounds.Source: F Limitations · p. 14
• Search uses single-step edits and top-k batches, limiting exploration around current high-scoring candidates and offering no mechanism to escape a local optimum or prevent near-duplicates.GeneratorSearch is available, but no campaign using it is reported.Source: F Limitations · p. 14

G Broader Impacts

• Better active-learning tooling is intended to reduce measurements needed to reach a result, while ALF’s shared offline-online API can avoid reimplementation between benchmarking and deployment.The authors also state that open release lowers the engineering cost of running campaigns.Source: G Broader Impacts · p. 14
• Because the acquisition loop is objective-agnostic, it can optimize harmful as well as useful properties; ALF contributes orchestration rather than capability.The wrapped surrogates, oracles, datasets, and scoring functions are already publicly available, and defining a harmful objective does not require ALF.Source: G Broader Impacts · p. 14
• A miscalibrated surrogate can spend an experimental budget in the wrong direction, so the authors recommend treating round-by-round surrogate metrics as a stop condition.They identify this as the practical risk most likely to occur.Source: G Broader Impacts · p. 14
• Model fine-tuning, oracle runs, and search carry non-trivial energy costs, although target-setting simulations and wet-lab measurements tend to cost more than model training and inference.Source: G Broader Impacts · p. 14