Machine learning is rebuilding the chemistry bench, one measurement problem at a time
In new preprints, pairwise learning beats single-value prediction on small peptide datasets, a retrosynthesis benchmark finally targets manufacturing, and a docking filter lifts metal-binding success rates from 18 percent to as high as 40 percent.
- In a preprint, a pairwise-learning framework for cyclic peptide permeability beat single-value regression on datasets with fewer than 1,000 peptides, with the biggest advantage for close analogues differing by a single modification. Self-reported m:…5581
- In a preprint, a new benchmark called OPRD-target collects 2,219 manufacturing chemistry targets from 1,915 papers, addressing a gap the authors identify: existing retrosynthesis benchmarks focus on discovery-scale synthesis, not manufacturing. Self-reported m:…5467
- In a preprint, a model called CHELATE lifted metal-binding success rates in docking from a baseline of 18 percent to 40 percent with one docking program, after training on more than 50,000 metalloprotein-ligand systems. Self-reported m:…9306
- In a preprint, a molecular representation method called Aimetry, trained on 44 million conformers from 2.8 million molecules, kept more geometric detail than standard contrastive learning and improved linear-probe performance across the evaluated datasets. Self-reported m:…5296
Ask which molecule is better, not how good it is
The most interesting idea in this week's batch of chemistry preprints is a change of question. A framework called CycPeptMMP predicts the membrane permeability of cyclic peptides, a class of molecules that drug developers care about because they can hit targets that small molecules cannot. Instead of predicting an absolute permeability value for each peptide, the model learns the difference between matched pairs of molecules, comparing two close relatives and asking which one crosses a membrane more easily. m:…5581
That shift matters because permeability measurements come from many different laboratories and carry systematic biases that make absolute numbers hard to compare. The authors report that pairwise learning outperformed single-value regression on datasets with fewer than 1,000 peptides, and that the advantage was greatest for close analogues differing by a single modification — exactly the situation a medicinal chemist faces when tweaking a lead compound. The framework also uses a censored loss function to handle measurements that were clamped at a detection limit, which the authors say improves accuracy in picking the more permeable member of a pair. m:…5581
A benchmark for the factory, not the bench
A second preprint attacks a different measurement problem: how do you know whether a computer-planned synthesis route is any good? The authors built OPRD-target, a collection of 2,219 unique target molecules drawn from 1,915 papers in Organic Process Research and Development, published between 1997 and 2022. The point is that existing benchmarks for computer-aided synthesis planning are aimed at discovery-scale chemistry, while manufacturing-scale route design has had almost nothing to measure against. m:…5467
The dataset lets developers compare routes their algorithms propose against routes that experts actually designed for industrial manufacturing. That comparison is the missing piece for anyone trying to use retrosynthesis software for scale-up rather than for making a few milligrams in a research lab. A benchmark does not solve the problem, but it does make it possible to tell whether the next tool is better than the last one. m:…5467
Geometry that survives the embedding
Two more preprints work on the representations that sit underneath these tools. Aimetry builds molecular embeddings by feeding quantum-mechanical features from a learned potential into a graph transformer, and it guides the embedding with fingerprint similarity between different molecules and structural deviation between conformers of the same molecule. The authors trained it on 44 million conformers derived from 2.8 million molecules, and report that it improved linear-probe performance across the evaluated datasets compared with standard contrastive learning baselines. m:…5296
The claim that matters is the one about what the embedding keeps. Standard contrastive learning tends to suppress geometric nuance, treating different shapes of the same molecule as near-identical. Aimetry, the authors report, retained more of the structural relationships defined by fingerprints and showed greater sensitivity to geometry. For property prediction that depends on a molecule's shape — which is most of it — that is the difference between a representation that works and one that quietly discards the signal. m:…5296
A companion effort, UEFCCS, applies a similar instinct to a narrower problem: predicting collision cross section values, which help identify molecules in ion mobility spectrometry. It combines traditional molecular descriptors with three-dimensional representations from a fine-tuned Transformer encoder trained on protonated-molecule data, and the authors report improved accuracy and better extrapolation to structurally dissimilar molecules — the cases where current models most often fail. m:…5580
Fixing docking where it actually fails
The most immediately usable of the batch is CHELATE, a post-docking filter for metalloproteins, where the metal-coordination geometry is often wrong even when the rest of a docked pose looks right. The model predicts how far a metal-binding group sits from its correct position and re-ranks docking poses to favour native-like coordination. Trained on a curated dataset of more than 50,000 metalloprotein-ligand systems covering over 20 metal ions, it lifted the success rate for metal-binding groups — defined as within 1 angstrom of the correct position — from a baseline of 18 percent to 32 percent with AutoDock Vina, 34 percent with smina, and 40 percent with GNINA. m:…9306
The practical attraction is that CHELATE slots into existing docking pipelines without retraining the underlying tools, and the authors describe it as computationally inexpensive enough for high-throughput screening. It does not make docking reliable; it makes one specific, common failure less likely, which is usually how these tools earn their keep. m:…9306
What to watch
Read together, the five papers describe a field that has stopped asking whether machine learning can predict a property and started asking whether the prediction can be trusted, compared, and used. Pairwise learning sidesteps biased measurements; a manufacturing benchmark gives synthesis planners something honest to aim at; geometry-aware embeddings try to keep the information that older methods threw away; and a docking filter fixes a known failure instead of promising a new one. None of these is a breakthrough in the usual sense. Each is a piece of plumbing, and plumbing is what decides whether a method gets used twice. m:…5581 m:…5467 m:…9306 m:…5296
What is still unresolved
- Whether pairwise learning's advantage holds on larger peptide datasets and on properties beyond permeability.
- Whether OPRD-target's expert routes become a widely adopted standard for manufacturing-scale retrosynthesis tools.
- Whether CHELATE's gains transfer to docking programs and metal ions outside those tested.
Provenance by Synorb. Every claim links to its source. This story was written by an agent from the Manifests listed below. Read the methodology and limitations, or send a correction.
Free to reuse, including commercially. Credit “Hanging Context, with provenance by Synorb” and link to this story. Keep the cited Manifest ids. Reuse terms v1.