D2ML: Models for Direct-to-Biology Small Molecule Screening
Authors: AkshatKumar Nigam, Neri Amara, Hillary Dequina, Dyana Kenanova, Seth Malmersjo, Fiorella Ruggiu, Raymond V. Fucini, Jack Sadowsky, Stig K. Hansen, Imran S. Haque
Organizations: Kimia Therapeutics, Berkeley, CA 94710
Direct-to-biology (D2B) screening enables rapid testing of newly synthesized molecules without purification. However, uncertain concentrations in crude mixtures and limited purified-compound measurements complicate the use of these data in machine learning. Here we show that models tailored to the structure of D2B data can improve prediction and guide the discovery of more potent and selective compounds. Our assay-aware models predict activity from single-point screens, relative potency across targets from crude dose-response assays, and absolute potency from purified-compound measurements. For crude assays, potency ratios between targets tested from the same well can retain selectivity information despite uncertain concentrations. A joint model that shares information across stages and balances their contributions to training improved predictions over separate neural networks. The largest gain was in crude dose-response selectivity prediction, where rank correlation increased from 0.27 to 0.40. In a separate experiment, independently trained models guided the selection of 84 new compounds to improve ROS1 potency while limiting activity against the anti-target TRKB. Synthesis and testing identified compounds that extended the measured Pareto front, achieving better combinations of potency and selectivity than those in the training data. The most selective new compound achieved a ROS1 IC50 of 14.7 nM and 233-fold selectivity over TRKB. By comparison, the most selective training compound had a ROS1 IC50 of 20.3 nM and 79-fold selectivity. ATLAS provides the experimental setting for this approach, supporting synthesis and screening of 1,000-10,000 model-selected compounds in two- to three-week cycles. Together, these results show how assay-aware learning can use the full D2B assay sequence to guide multi-objective molecular optimization.
Figures & tables
Figure 1: Direct-to-biology synthesis, screening stages, and preservation of selectivity across purification. (A) Kimia’s ATLAS platform combines a target-directed bit with libraries of blocks in one-reaction-per-well format. Two-component reactions explore one molecular vector; three-component reactions explore two, yielding up to n2 products from n blocks. (B) The D2B cascade progresses from SP ( O(104) compounds) to DR ( O(103) ) and purified MSP ( O(102) ). DR and MSP support potency ratios across targets, while only MSP supports quantitative comparison across compounds. (C) Purification shifts ROS1 and LTK IC 50 values from the same crude well by nearly identical factors ( 25.2× and 25.3× ), preserving the ∼122 -fold potency ratio and showing that crude DR can retain selectivity even when absolute potency is uncertain. Supplementary Figure 7 shows a systematic paired analysis of DR/MSP potency ratios.
Figure 2: Model architectures across the D2B assay cascade. All models take Morgan fingerprints and differ in supervision and information sharing. (A) L2-regularized linear regression, one coefficient matrix: “Linear” is fitted to MSP alone, and “Linear (all stages)” to all three stages. (B) Single-task: independent networks for each stage. (C) Multi-task: a shared backbone with stage-specific heads; “MT − loss balancing” removes stage-wise loss normalization and weighting. Solid arrows denote the forward pass; dashed arrows denote supervision.
Compounds assayed per target
Compounds per split
Stage
LTK
ROS1
TRKB
ALK
train
val.
test
total
SP
26,538
15,368
–
–
21,228
2,654
2,656
26,538
DR
1,462
1,332
1,163
–
1,273
163
180
1,616
MSP
163
163
163
163
114
25
24
163
MSP-prospective
84
84
84
84
–
–
84
84
Table 1: Dataset used in this work. Left: per-target assay coverage; right: train–validation–test counts. Each compound has one split shared across assay stages, preventing cross-stage leakage (Supplementary Section 5.2 ). The 24-compound MSP test split is MSP-held (Table 2 ); MSP-prospective is a separate prospective set used only for evaluation.
SP
DR
MSP-held
MSP-prospective
Model
ROC-AUC
ρ
r
ρ
r
ρ
r
Linear
–
–
–
0.624
0.616
−0.029
−0.021
Linear (all stages)
0.732
0.241
0.174
−0.011
−0.117
0.349
0.357
Single-task
0.890
±
.006
0.271
±
.049
0.259
±
.063
0.420
±
.026
0.365
±
.033
0.120
±
.034
0.150
±
.027
Multi-task (MT)
0.892
±
.011
0.402
±
.024
0.400
±
.025
0.443
±
.047
0.443
±
.062
0.175
±
.015
0.188
±
.022
MT -loss balancing
0.874
±
.006
0.336
±
.020
0.334
±
.020
0.427
±
.048
0.357
±
.044
0.133
±
.025
0.132
±
.025
Table 2: Predictive performance across the D2B assay cascade. SP is scored by macro AUROC (LTK, ROS1); DR and MSP by Spearman ( ρ ) and Pearson ( r ), computed per target and macro-averaged (three pairwise ratios at DR, four kinases at MSP). MSP-held is the retrospective out-of-distribution set ( 24 molecules), MSP-prospective the independent prospective set ( 84 ). The MSP-prospective compounds were selected using single-task models; multi-task results on this cohort assess prediction accuracy, not the ability of multi-task models to select compounds prospectively. DR and MSP evaluation is restricted to resolved potency estimates ( pinact<0.30 ). Deep models report mean ± s.d. over five initializations; linear baselines are single fits.
Figure 3: Model-guided prospective selection occupies favorable MSP prediction space and extends the measured ROS1 potency–selectivity frontier. (A,B) Predicted (A) and measured (B) ROS1 potency and ROS1/TRKB selectivity ( IC50 ratio). A: MSP-train prediction density (gray); MSP-held (blue) and -prospective (orange) predictions. B: DR/MSP training measurements (one per compound; n=1,137 , blue); MSP-prospective ( n=84 , orange). Open circles denote Pareto-optimal compounds in either set; black markers identify compounds in panel C. (C) The most selective prospective (1) and training (2) compounds and the closest training analog of compound 1 (3). Changes are largest in the left-side substituent; (1) is not a local change from (2), and its closest neighbor is weak, indicating exploration beyond local transformations of top hits (T = ECFP4 Tanimoto similarity).
Figure 4: Prospective selection enriches measured ROS1 potency and selectivity. Percentage of compounds satisfying measured criteria for ROS1 potency (IC 50<100 nM), ROS1-over-TRKB selectivity ( ≥10× ), or both, in MSP-train, MSP-held, and MSP-prospective. Only compounds with resolved potency estimates for both targets ( pinact<0.30 ) are included. Annotations report fold enrichment of MSP-prospective relative to MSP-train. The corresponding distributions in MSP prediction space are shown in Figure 3 A.
Figure 5: Bayesian model for dose-response curve fitting. Robust Bayesian fits and parameter-sharing at the plate level avoid common numerical instabilities arising in least-squares fitting of assay data that may range from highly active to totally inactive, and enable downstream propagation of uncertainties.
Figure 6: Stage contributions to training loss in an exploratory multi-task model. Total training loss (black dashed) and its SP, DR, and MSP contributions without (A) and with (B) epoch-wise geometric-mean loss balancing. Thin curves show three paired initializations and thick curves their means over all 300 epochs. Loss contributions are recorded during training and sum to the total; the vertical axis is logarithmic. The model uses soft-label binary cross-entropy for SP and standardized Huber losses for DR and MSP, distinct from the likelihood-based models in the main results.
Hyperparameter
Range
Scale
Fingerprint radius
{2,3}
categorical
Fingerprint length
{2048,4096}
categorical
Backbone width
{512,1024,2048}
categorical
Residual blocks
{1,2,3}
categorical
Dropout
[0.05,0.35]
uniform
Head width
{0,128,256}
categorical
Table 3: Hyperparameter search space for the neural models. Stage weights are optimized only for the multi-task model; the MT − loss balancing configuration uses equal stage weights by construction. The linear warm-up period was fixed at 8 epochs for all neural models.
Model
Radius
Bits
Width
Blocks
Head
Dropout
Single-task (SP)
3
4096
512
3
256
0.086
Single-task (DR)
2
4096
1024
3
256
0.099
Single-task (MSP)
3
4096
512
3
128
0.217
MT − loss balancing
2
2048
1024
1
256
0.196
Multi-task (MT)
3
4096
2048
2
256
0.125
Table 4: Selected molecular-representation and architecture hyperparameters for the neural models. “Head” denotes the width of the stage-specific prediction head.
Model
Learning rate / WD
λσ
λact
Stage weights
Single-task (SP)
8.5×10−4 / 1.1×10−6
0.6
0.7
—
Single-task (DR)
1.4×10−3 / 5.1×10−5
0.4
0.0
—
Single-task (MSP)
2.0×10−3 / 4.2×10−4
0.5
0.1
—
MT − loss balancing
7.6×10−4 / 3.6×10−5
0.5
0.2
1.0/1.0/1.0
Multi-task (MT)
2.6×10−4 / 3.5×10−4
0.5
0.5
3.0/1.5/2.0
Table 5: Selected optimization and loss hyperparameters for the neural models. “WD” denotes weight decay. Stage weights are reported as wSP/wDR/wMSP ; MT − loss balancing uses equal stage weights by construction.
Figure 7: Paired crude and purified selectivity. Compound-matched DR and MSP selectivity for (A) ROS1/TRKB, (B) ROS1/LTK, and (C) LTK/TRKB ( n=121 , 114, and 99). Points are pooled compound-level posterior summaries from all retrospective splits; both targets must be resolved at both stages. Dashed lines indicate equality and shading a descriptive threefold band. Pearson r and Spearman ρ are shown.
Figure 8: MSP-only counterpart to Figure 3 B. Blue and orange contours show the historical ( n=163 ) and prospective ( n=84 ) purified MSP cohorts, respectively. Open circles and connecting steps mark the Pareto frontiers, which remain unchanged after excluding crude DR measurements.
High-throughput drug screening uses inexpensive primary screens to select compounds for more costly follow-up dose-response experiments, which estimate compound potency (pXC50) by determining the concentration required to produce a specified biological effect. However, many promoted compounds fail to yield a reportable potency, even when activity is subsequently confirmed. We investigated whether inexpensive three-point primary-screening data could predict whether a more costly 11-point dose-response experiment would yield a reportable pXC50. Using a dataset of 32,971 compound-assay pairs with three-point primary-screening and 11-point dose-response measurements, we trained random-forest models to predict successful potency quantification. Our model achieved an AUROC of 0.939 +/- 0.002 and generalized to previously unseen chemical scaffolds and assay-mechanism families. We found that primary-screen response features were more predictive of dose-response quantifiability than molecular structure-based features. Successful potency quantification was also strongly associated with primary-screen response amplitude. Across eight chronological test releases, the model retrospectively retained 90% of reportable potencies while selecting only 55.7% of historically promoted compound-assay pairs for follow-up profiling. These findings suggest that predicting potency quantifiability from primary-screening data could improve the efficiency of compound prioritization and reduce unnecessary dose-response experiments.
Sean Lim
Department of Natural Science, Rice University, Houston, TX, USA
Designing small molecule ligands that bind with high affinity to specific protein pockets is a fundamental goal in drug discovery, as small molecules constitute a major fraction of approved therapeutics. Recent breakthroughs in structure prediction, such as AlphaFold-3 and Boltz-2, enable accurate biomolecular interaction prediction and show promise as foundation models for downstream tasks, including binding affinity prediction. We propose to leverage these models and introduce DBMol, a new structure predictor-guided framework for de novo small molecule design. DBMol formulates an alternating optimization and projection process. In the optimization stage, DBMol starts from an initial molecule and uses gradient-based optimization to improve pocket-specific interactions and predicted binding affinity using a structure prediction model. In the projection stage, a flow-matching model maps the optimized molecular graph to discrete and chemically valid molecules. Experiments show that DBMol effectively optimizes the Boltz-2 affinity proxy and generates molecules with strong predicted affinity and specificity under Boltz-2 evaluation. To reduce self-confirmation bias, we further evaluate generated molecules using held-out metrics, including AF3-based evaluation. DBMol substantially improves pocket coverage while maintaining molecular diversity over unconditional generation, and is competitive under held-out metrics despite the absence of reference-ligand supervision. These results support the promise of structure prediction models as effective optimization signals for de novo molecular design.
Yiming Qin, Kai Yi, Miruna Cretu +3
EPFL, Switzerland · MRC-LMB, UK · University of Cambridge, UK
Designing small-molecule drugs that selectively bind to a target protein while avoiding unintended interactions with other proteins is essential for reducing side effects that contribute substantially to clinical attrition. Although structure-based drug design (SBDD) has advanced rapidly with deep learning, most generative models still optimize binding to a single target and disregard selectivity due to limited training data and unknown off-target interactions. Motivated by this gap, we introduce SELECTBENCH, a comprehensive benchmark assessing the selectivity of generative SBDD models. SELECTBENCH provides three curated evaluation tasks reflecting different real-world drug discovery scenarios: literature-based case studies, safety screening panels used in pharmaceutical development, and large-scale structure-based evaluation sets. We further introduce a suite of metrics to quantify selectivity and conduct a systematic evaluation of nine representative SBDD models, including methods explicitly trained for selectivity and widely-used inference-time guidance techniques. Our results show that even selectivity-trained models achieve only modest off-target discrimination and inference-time guidance provides limited gains, underscoring the need for method improvement. SELECTBENCH provides both a characterization of where current generative SBDD stands on selectivity and an open, extensible platform to accelerate progress. The benchmark is available at https://github.com/lovetaro/selectbench.
Ella Miray Rajaonson, Jungyoon Lee, William Chau +5
University of Toronto · Vector Institute · Université de Montréal +6