PaperlayerView exact revision on arXiv

Source-linked AI summary

Machine learning on DNA-encoded libraries: A new paradigm for hit-finding

Kevin McCloskey, Eric A. Sigel, Steven Kearnes, Ling Xue, Xia Tian, Dennis Moccia, Diana Gikunju, Sana Bazzaz, Betty Chan, Matthew A. Clark, John W. Cuozzo, Marie-Aude Guié, John P. Guilinger, Christelle Huguet, Christopher D. Hupp, Anthony D. Keefe, Christopher J. Mulhern, Ying Zhang, Patrick Riley

arXiv:2002.02530v1q-bio.QMcs.LG

TL;DR

Finding diverse small-molecule hits is costly and lengthy, and effective screening of large virtual libraries has long been a goal. This paper applies machine-learning models trained on DEL selection data to low-cost chemical libraries; prospectively, GCNNs generalized to new chemical spaces and outperformed simpler baselines.

  • Problem

    Discovering small-molecule therapeutics is expensive and time-consuming, while effective computational screening of large virtual libraries has long been a goal.

    Source: Introduction · p. 1

  • Method

    The approach trains machine-learning models using DEL selection data and applies them with automated or automatable filters to large, low-cost chemical libraries.

    Source: Introduction · p. 1, Discussion · p. 3

  • Results

    For GCNN models across three targets, the best-performing target yielded hit rates of 72% at 30 µM, 33% at 10 µM, and 29% at 1 µM, outperforming simpler baselines prospectively.

    Source: Introduction · p. 2

  • Takeaways & Limitations

    Model generalization gives practitioners flexibility to choose virtual libraries based on properties such as synthesizability, commercial availability, or desired molecule-property ranges.

    Source: Discussion · p. 3

    Hit rates varied across targets and may correlate with positive training-example counts; c-KIT had the fewest positives and lowest hit rates.

    Source: Discovering potent ligands · p. 2

Abstract

from arXiv · show

DNA-encoded small molecule libraries (DELs) have enabled discovery of novel inhibitors for many distinct protein targets of therapeutic value through screening of libraries with up to billions of unique small molecules. We demonstrate a new approach applying machine learning to DEL selection data by identifying active molecules from a large commercial collection and a virtual library of easily synthesizable compounds. We train models using only DEL selection data and apply automated or automatable filters with chemist review restricted to the removal of molecules with potential for instability or reactivity. We validate this approach with a large prospective study (nearly 2000 compounds tested) across three diverse protein targets: sEH (a hydrolase), ER{\alpha} (a nuclear receptor), and c-KIT (a kinase). The approach is effective, with an overall hit rate of {\sim}30% at 30 {\textmu}M and discovery of potent compounds (IC50 <10 nM) for every target. The model makes useful predictions even for molecules dissimilar to the original DEL and the compounds identified are diverse, predominantly drug-like, and different from known ligands. Collectively, the quality and quantity of DEL selection data; the power of modern machine learning methods; and access to large, inexpensive, commercially-available libraries creates a powerful new approach for hit finding.

Introduction

The paper trains machine-learning models on DEL selection data to screen large, low-cost chemical libraries for hits outside the DEL. This approach addresses the scale and bias limits of human analysis of DEL selections.

Introduction

  • The work addresses the costly, lengthy process of discovering small-molecule therapeutics, for which diverse, tractable hits are foundational.Source: Introduction · p. 1
  • DEL screening can read out target binding by millions to billions of molecules in parallel, expanding the chemical space explored in a single experiment [2–4].The passage also notes that DEL-identified hits have advanced into clinical trials [6,7].Source: Introduction · p. 1
  • Human analysis of DEL selections limits the number of molecules considered, introduces bias, and can miss subtle patterns obscured by yield variation and random sampling [8,9].Source: Introduction · p. 1
  • The approach applies machine learning to DEL selection data to identify hits in large libraries of inexpensive or easily synthesizable compounds.The workflow includes automated diversity and reactive-substructure filters, with chemist review limited to removing potentially unstable or reactive molecules.Source: Introduction · p. 2
  • For the best-performing target, hit rates were 72% at 30 µM, 33% at 10 µM, and 29% at 1 µM, versus ~1% for traditional HTS without ML.The authors also report stronger prospective performance than simpler baseline models.Source: Introduction · p. 2

Results

Prospective testing across three protein targets yielded many potent ligands, with GCNN models outperforming RF models. The identified hits included compounds dissimilar to the training data and structurally novel relative to known ligands.

Results

  • Models were trained on DEL selection data alone, without known-ligand data beyond the competitive inhibitors used in selections, explicit target representations, or 3D data.The authors applied diversity, logistical, and structural filters plus restricted chemist review to predictions; all successfully acquired or synthesized compounds were experimentally validated.Source: Discovering potent ligands · p. 2
  • Two sEH DEL selections performed months apart produced training sets that were equivalent for model training.Source: Discovering potent ligands · p. 2
  • The prospective study covered sEH, c-KIT, and ERα, using single-point assays followed by dose–response confirmation across 1885 compounds from Mcule and a virtual library.The validations were biochemical activity or ligand-displacement assays.Source: Discovering potent ligands · p. 2
  • Across three targets, 304 ligands had better than 10 µM potency and 165 had better than 1 µM potency; GCNN models achieved higher hit rates and better potencies than RF models.Hit rates varied across targets; c-KIT was least productive, with 78 GCNN hits below 30 µM. Mcule was generally more productive than XVL, and 568/583 (97%) confirmed hits had ≤1 Lipinski Rule of 5 violations.Source: Discovering potent ligands · p. 2
  • A similarity search against selected ERα training examples tested 107 compounds and yielded no hits with detectable activity.Source: Discovering potent ligands · pp. 2–3
  • The study frames generalization as important because DELs sample only a minute fraction of the estimated universe of small, drug-like molecules, while models often fail on different data distributions.Source: Analysis of confirmed hits discovered… · p. 3
  • Below 0.4 ECFP Tanimoto similarity to training data, the study still found 22 sEH, 28 ERα, and 5 c-KIT hits with better than 30 µM potency.The least-similar ERα hit had an ECFP Tanimoto similarity of 0.29 and an IC50 of 20 nM; confirmed-hit scaffold overlap with training was 42.7% for GCNN and 60.8% for RF.Source: Analysis of confirmed hits discovered… · p. 3
  • Only 2.2% of GCNN hit scaffolds and 3.0% of RF hit scaffolds had previously been reported in ChEMBL for these targets.Source: Analysis of confirmed hits discovered… · p. 3

Discussion

The discussion presents DEL-trained machine learning with filters as a way to find diverse, novel hits outside DELs. The authors emphasize low-cost library access and identify data, models, and make-on-demand compounds as factors in the approach.

Discussion

  • The approach combines DEL selection data, machine learning, and automated or automatable filters to identify diverse, novel hits outside the DEL across three protein targets.The authors describe screening purchasable or easily synthesizable molecules, avoiding the expense and time of adding new matter to DEL or HTS libraries.Source: Discussion · p. 3
  • The authors attribute success to more powerful machine-learning methods, high-quantity and high-quality DEL data, and large, low-cost, structurally diverse make-on-demand libraries.Source: Discussion · p. 3
  • The authors suggest exploring applications complementary to traditional HTS and combining the approach with lead generation and optimization.They also anticipate applying DEL selections to properties beyond competitive on-target binding, including some ADMET properties.Source: Discussion · p. 4
  • The authors expect the approach's utility to grow as machine-learning models improve and make-on-demand libraries expand.They hope broader application across targets will affect later stages of drug discovery.Source: Discussion · p. 4

Methods

The methods classify DEL disynthons from selection data, train and evaluate random-forest and graph-convolutional models, then select compounds for experimental testing. A separate similarity-search baseline and biochemical assays assess prospective activity.

Disynthon classification

  • Competitive-hit labels required target-condition enrichment, no enrichment with matrix-only or competitor conditions, and a low promiscuity ratio.Disynthons were assigned to five classes: competitive hit, non-competitive hit, promiscuous binder, matrix binder, or non-hit.Source: Classification of Disynthons from… · p. 4

Model training and evaluation

  • Random-forest models used ECFP4 fingerprints, while GCNN models used a weave architecture and were trained to predict five disynthon classes.GCNN cross-validation averaged approximately 0.8 ROC-AUC; ensemble holdout performance for c-KIT and ERα was approximately 0.99 ROC-AUC.Source: Random forest models · p. 4, GCNN models · pp. 4–5

Similarity-search baseline

  • The similarity-search baseline selected 107 tested Mcule compounds, and it produced no molecules with detectable activity.The search used similarity to enriched ERα training examples; compounds were filtered and clustered before testing.Source: Compounds selected by similarity search · p. 5

Compound selection

  • Predicted compounds were selected by score thresholds, diversity clustering, and molecular-property filters, followed by chemist review limited to potential instability or reactivity.Selection also removed duplicate scaffolds and applied library-specific filters, including molecular-weight and delivery-time restrictions for some Mcule orders.Source: Selection of diverse predicted compounds · p. 5

Molecular similarity analysis

  • Molecular similarity was quantified with count-based ECFP6 or FCFP6 fingerprints and Tanimoto similarity, with nearest neighbors found by exact search.The count-based vectors were used to better capture molecular-size differences from repeated substructures.Source: Molecular similarity comparisons · pp. 5–6

Training-data reproducibility

  • ROC-AUCs of 0.97 and 0.99 cross-predicted competitive-hit labels between sEH DEL selections performed months apart.The first selection’s enrichment values predicted the second selection’s labels at 0.97 ROC-AUC; the reverse prediction achieved 0.99.Source: Reproducibility of training data · p. 6

Virtual-library synthesis

  • The 83.2-million-compound XVL was enumerated from compatible in-house building blocks, and selected compounds were synthesized on a micromole scale for assay testing.Reaction mixtures were filtered and analyzed by LC-MS before dilution for primary biochemical assays.Source: On-demand synthesis of virtual library… · p. 6

Experimental validation

  • The study tested compounds using single-point inhibition assays followed by dose–response assays, reporting potency as IC50.Potency and hit-rate reporting aggregated single-point and dose–response results, with specified curve-validity thresholds and resource-based limits on retesting.Source: Assay cascade and reported potency… · pp. 7–8

Code availability

The authors used scikit-learn’s RandomForestClassifier for RF predictions and a custom TensorFlow implementation of the W2N2 graph convolutional neural network for GCNN predictions.

  • RF predictions used scikit-learn’s RandomForestClassifier, while GCNN predictions used a custom TensorFlow implementation of the W2N2 graph convolutional neural network.Source: Code availability · p. 8

Author contributions statement

The statement assigns study conception and direction, model and classification development, model training, experiments, analysis, and writing across the author team.

Additional information

The authors disclose employment ties, company interests, and patent filings related to DEL technology and machine learning.

  • All authors were current or former employees of X-Chem or Google, and the passage lists company interests and patent filings associated with the work.Source: Competing interests · p. 10

Figures

The figures and tables map the DEL-to-machine-learning workflow and document prospective hit testing across targets, models, and compound sources. Additional visuals examine similarity to training data, model performance, hit examples, retesting thresholds, and validation costs.

Figures

  • Figure 1 illustrates the pipeline from DEL selections and aggregated disynthon labels to model predictions on virtual or commercial libraries, followed by experimental testing.Selections covered about 40 libraries and ∼10^11 unique molecules; the example also notes that counter-selections can provide richer labels.Source: Figures · p. 11, Figure 1 · p. 11
  • Figure 2 summarizes tested compounds, hit rates, and potencies across three targets, two models, and Mcule and XVL; overlapping buckets yield 1,885 unique molecules from 1,900 entries.Darker colors represent lower concentrations, and a black vertical line marks 1 µM in each bar chart.Source: Figure 2 · p. 11
  • At sEH, the 1 µM hit rate falls from 29.7% across 347 compounds to 22.2% among 36 compounds with training-set similarity ≤0.40.Figure 3 plots cumulative hit rates by similarity to the training DELs and includes scatter plots of hits; error bands are 95% Clopper–Pearson intervals.Source: Figure 3 · p. 12
  • Table 1 pairs potent-hit examples with their closest known ChEMBL hits, while Extended Data Table 1 lists nearest training-set neighbors for selected low-similarity hits.The captions specify ECFP6-counts Tanimoto similarity, with FCFP6-counts values also given in parentheses.Source: Table 1 · pp. 12, 17
  • Extended Data Figures 1–2 break out hit rates and potencies by compound source and plot cumulative hit rates against training-set similarity across models and fingerprint types.Figure 2 includes GCNN and RF predictions using ECFP6-counts or FCFP6-counts fingerprints.Source: Figure 1 · p. 13, Figure 2 · p. 14
  • Extended Data Figure 3 compares GCNN with DNN cross-validation performance and shows holdout-set evaluations for the models used in experimental predictions..Source: Figure 3 · p. 15
  • Extended Data Figure 4 plots confirmed-hit similarity to the nearest training compound and nearest positive training example, with each distribution mean marked..Source: Figure 4 · p. 15
  • Extended Data Tables 2–3 give dose-response retesting thresholds and approximate validation costs; the cost caption attributes the main difference to compound acquisition, with custom synthesis 20x–30x more expensive.The cost estimate excludes human-analysis time.Source: Table 2 · p. 18, Table 3 · p. 19
Loading 2002.02530v1…