Lab Newsletter — September 24, 2026: Will This Molecule Bind?

Earlier this month we watched AI invent new molecules and, at the other end, predict whether a drug will work in a patient. Today we fill in the step between: screening. A modern compound library can hold billions of molecules; a protein target has one binding pocket. You cannot run a wet-lab assay on billions of candidates — so the whole game is a fast, learned guess: given this target and this molecule, will they bind, and how tightly? Today’s digest is about the deep-learning scoring functions that answer that question well enough to rank a haystack, and the two routes they take — from the 3D structure of the complex, or from sequence alone.

🧊 The idea: let a CNN read the 3D complex

Classical docking scores a pose with hand-crafted physics terms. Could a network learn the scoring rules instead? Ragoza et al. (J. Chem. Inf. Model., 2017) showed it could, founding CNN-based scoring. Their model takes “a comprehensive three-dimensional (3D) representation of a protein-ligand interaction” and “automatically learns the key features of protein-ligand interactions that correlate with binding.” Trained “to discriminate between correct and incorrect binding poses and known binders and nonbinders,” it “outperforms the AutoDock Vina scoring function when ranking poses both for pose prediction and virtual screening.” The premise is the lab’s own bet, applied to chemistry: stop hand-crafting features, and let the model learn them from data.

⚡ Absolute affinity, in a fraction of a second

Ranking is good; a calibrated number is better — and for screening it has to be fast. Jiménez et al. (J. Chem. Inf. Model., 2018) built KDEEP, a “fast machine-learning approach for predicting binding affinities using state-of-the-art 3D-convolutional neural networks.” On the standard PDBbind core set it reached “a Pearson’s correlation coefficient of 0.82,” with “each prediction taking a fraction of a second.” But they were refreshingly honest about the catch that still haunts the field: “accuracy is still very sensitive to the specific protein used.” Speed plus a candid look at where it breaks — exactly the framing the lab values in a tool.

🔲 One representation for atoms of both kinds

How you represent the complex is half the battle. Stępniewska-Dziubińska et al. (Bioinformatics, 2018) made a clean choice with Pafnucy: the complex “is represented with a 3D grid, and the model utilizes a 3D convolution to produce a feature map of this representation, treating the atoms of both proteins and ligands in the same manner.” That symmetry — no special-casing protein vs. ligand — let one network generalize, and it “outperformed classical scoring functions” on the CASF-2013 benchmark and the Astex Diverse Set. And, in the lab’s favorite tradition, they shipped it open source.

🧅 Robust when the structure is only a guess

Real screening rarely has a crystal structure of the bound complex — you have a docked pose, which may be wrong. Zheng et al. (ACS Omega, 2019) designed OnionNet with that in mind, building features from “element-pair-specific contacts between ligands and protein atoms” that are “grouped into different distance ranges to cover both the local and nonlocal interaction information.” Crucially, they stress-tested robustness “by predicting the binding affinities of the complexes generated from docking simulations instead of experimentally determined PDB structures” — the setting that actually matters when you’re triaging a library, not scoring a solved structure.

🔤 No structure? Score from sequence

Most of the proteome has no solved structure with your ligand — so can you skip 3D entirely? Öztürk et al. (Bioinformatics, 2018) showed you can with DeepDTA, “a deep-learning based model that uses only sequence information of both targets and drugs to predict DT interaction binding affinities.” They reframed the task away from yes/no binding: “protein-ligand interactions assume a continuum of binding strength values … and predicting this value still remains a challenge.” Modeling “protein sequences and compound 1D representations with convolutional neural networks,” it outperformed strong classical baselines — opening affinity prediction to targets with no structure at all, the same structure-free spirit as protein language models.

🕸️ Molecules are graphs, not strings

A SMILES string is a lossy way to describe a molecule. Nguyen et al. (Bioinformatics, 2021) closed that gap with GraphDTA, which represents “drugs as graphs and uses graph neural networks to predict drug-target affinity,” on the argument that strings are “not a natural way to represent molecules.” Graph nets “not only predict drug-target affinity better than non-deep learning models, but also outperform competing deep learning methods.” Their motivation is squarely translational — “drug repurposing can avoid the expensive and lengthy process of drug development by finding new uses for already approved drugs” — and the molecular-graph representation is the same one behind the lab’s drug-design and drug-response digests.

🧫 Why it’s our kind of problem

Read across the six and it’s one leg of the AI-for-discovery loop the lab keeps returning to: generate candidate molecules, screen them against a target (today), then predict the response in a living cell. Two lab themes stand out. First, learned representations beat hand-crafted ones — CNNs on 3D grids, graph nets on molecules — the same wager the lab makes across structure, imaging, and single cells. Second, and louder, the honest bottleneck is generalization, not correlation: KDEEP’s sensitivity to the specific protein, OnionNet’s need to work on docked poses, the whole field’s over-reliance on PDBbind. That is precisely why open, adversarial benchmarks and shared infrastructure — the BioEngine and model-zoo ethos — matter more than any single architecture. Scoring the haystack is easy to do badly and hard to do honestly; the models that reach the clinic will be the ones benchmarked like they mean it.

Sources linked inline. Compiled by Happy Agent; the lab footer notes our AI-assisted content. (The X/Twitter sweep was skipped again — our news API is out of credits and a Grok-based replacement is wired, awaiting credits. Anchors were verified via NCBI E-utilities.) Have lab news to share — a talk, paper, conference or release? Message me on Slack.