Molecular AffinityHKMU AI4S Lab
← All four methods

02 / DL-FSG

Molecular screening

Get a relative screening score for your molecule and a protein target.

DL-FSG combines molecular sequence and spatial shape. Compare scores for molecules against the same target.

Try your own input Explore the research

Screen a molecule

Submit one SMILES and choose a target. Calculate a relative score, explore its attribution and download the completed run.

Preparing analysis…

One model per protein target

Screening asks which compounds to prioritise from a collection. Each available target uses its own model, trained on compounds labelled for that protein. It scores a molecule using its SMILES and 3D shape. Visitors do not need to supply a protein sequence, structure or complex.

Target-specific screening. Pair-based models use protein and compound features together. Target-specific models learn to recognise candidate compounds for one protein. Figure 1, Ke and Wang (2026).

Target-specific screening

Protein–compound pair prediction on the left and target-specific compound screening on the right.

Pair-based models use protein and compound features together. Target-specific models learn to recognise candidate compounds for one protein. Ke and Wang (2026).

Why use two descriptions?

A SMILES sequence records chemical connectivity, while a molecular grid provides spatial context. Our question was whether learning interactions between these descriptions improves the ranking of active compounds.

Read the DL-FSG paper ↗

DL-FSG: learning across two molecular views

DL-FSG encodes the SMILES sequence and ligand grid in parallel. Bidirectional cross-attention lets the features from each branch inform the other. The combined features then produce a target-specific screening score.

Sequence–grid fusion. DL-FSG combines sequence and spatial features through cross-attention. Figure 2, Ke and Wang (2026).

Sequence–grid fusion

DL-FSG architecture with parallel SMILES and molecular grid encoders, bidirectional cross-attention and a screening prediction head.

DL-FSG combines sequence and spatial features through cross-attention. Ke and Wang (2026).

Screening across 102 DUD-E targets

The paper evaluates separately trained models across 102 DUD-E targets and reports aggregate AUROC 0.9910 and AUPRC 0.9725. In the live tool, each available target has its own model. It scores a ligand from its SMILES and generated 3D shape; scores should be compared within one target.

Screening performance across targets. Screening performance across 102 DUD-E targets. Figure 3, Ke and Wang (2026).

Screening performance across targets

AUC and AUPR comparisons between DL-FSG and baseline methods across 102 DUD-E targets.

Screening performance across 102 DUD-E targets. Ke and Wang (2026).

What changes when a component is removed?

Replacing cross-attention with self-attention reduced AUC to 0.9855. Using either SMILES or grids alone also reduced performance. The two ablation studies support learning from both representations and their interaction.

Attention comparison. Cross-attention improves both AUC and AUPR over self-attention and the configuration without attention. Table 4, Ke and Wang (2026).

Attention comparison

Attention comparison: cross-attention AUC 0.9910 and AUPR 0.9725; self-attention 0.9855 and 0.9665; no attention 0.9798 and 0.8448.

Cross-attention improves both AUC and AUPR over self-attention and the configuration without attention. Ke and Wang (2026).

Representation comparison. Combining sequence and grid features improves screening performance over either representation alone. Table 5, Ke and Wang (2026).

Representation comparison

Representation comparison: combined features AUC 0.9910 and AUPR 0.9725; SMILES alone 0.9683 and 0.8323; grids alone 0.9646 and 0.8266.

Combining sequence and grid features improves screening performance over either representation alone. Ke and Wang (2026).

Screening prioritises compounds for a target. Our ETDock method examines a related structural question: how a ligand might fit within its protein.

Reading the grid branch with Grad-CAM

Grad-CAM highlights spatial regions associated with a higher screening score. It uses score gradients to weight the grid branch’s feature maps. The gradients pass through the sequence–grid fusion model while the molecule and conformer remain fixed.

Follow the score back to the grid

The DL-FSG architecture joins sequence and grid features. Differentiate that joint score with respect to the grid branch’s spatial activations, before pooling and projection.

L = ReLU(Σk αkAk)

Ak is a spatial feature map. Its weight αk is the mean score gradient over that map. Only the positive part of the weighted sum is retained.

For this ethanol conformer, the MAO-B score is 0.5371.

a Locate the plane

Ethanol conformer and the blue plane at the z coordinate of the Grad-CAM cross-section.

z = 0.21 Å in the model’s spatial frame.

b Read the spatial response

Ethanol Grad-CAM cross-section on a fixed zero to 0.00001 scale, with projected heavy-atom positions and a broad response across the plane.

Atoms are projected onto the plane. The 24 Å field contains a 12 × 12 cross-section.

The response extends across the plane. This ethanol example produces a broad Grad-CAM response. Atom markers locate the molecule within the map, whose shading shows positive Grad-CAM intensity. Grad-CAM: Selvaraju et al. (2017), applied to DL-FSG.

Each voxel summarises a spatial neighbourhood, giving a coarse view of the model’s response. Read the map at this scale rather than as a contribution from an individual atom. Colours are scaled to the peak of this prediction. The map shows model attribution, not proof of a binding mechanism; colour strength should not be compared across different runs.

Papers & resources

  1. Rongji Ke and Debby D. Wang. Deep learning on the fusion of chemical sequences and molecular grids for ligand-based virtual screening. BIO Web of Conferences 232, 02002 (2026).

  2. Ramprasaath R. Selvaraju, Michael Cogswell, Abhishek Das, Ramakrishna Vedantam, Devi Parikh and Dhruv Batra. Grad-CAM: Visual Explanations From Deep Networks via Gradient-Based Localization. Proceedings of the IEEE International Conference on Computer Vision, pp. 618–626 (2017).

← All four methods