Knowing the parts, missing the whole: does machine learning understand new molecules built from what it already knows?

Knowing the parts, missing the whole

Amir Masoud Nourollah

Stefano Leoni

Steven Schockaert

Overview

1. Introduction: how machine learning learns, and how chemists and computers read molecules
2. Motivation: where the data runs out, and what counts as new
3. GMD: the benchmark: familiar parts, new molecules
4. Results: where today’s models break
5. Takeaways

One sentence: chemists know a molecule from its parts; we ask whether ML models do too.

So how does a computer learn anything at all?

Teaching a computer from examples

• We give the computer examples: inputs and the correct answers
• It tunes a flexible formula until its predictions match those answers
• “Learning” just means finding the parameters with the smallest error

You already do this: fit a curve to two or three experimental points, then predict other amounts of reactant without running more experiments.

C. M. Bishop, Pattern Recognition and Machine Learning, Springer (2006), ch. 1.

Keep it friendly. Model = summary of the data it has seen.

Can we trust that curve everywhere, even far from the points we measured?

Inside the data versus outside it

• Between known points the model interpolates, and is usually reliable
• Beyond them it extrapolates, and nothing in the training checks it
• A very flexible model can pass every training point and still go wild just outside

A small error on the training data says little about data the model has never seen.

C. M. Bishop, Pattern Recognition and Machine Learning, Springer (2006), §1.1, §3.2.

Plant the seed: “outside the data” is the theme of the talk. Don’t say “out-of-distribution” yet.

Let’s see how different our interpretation of a molecule is from a computer’s.

How a chemist reads a molecule

G. Socrates, Infrared and Raman Characteristic Group Frequencies, 3rd ed., Wiley (2001).

Each part of ethyl acetate has its own peak: C–H stretch near 2980 cm⁻¹, C=O near 1740 cm⁻¹, C–O–C near 1240 and 1050 cm⁻¹. We read the whole molecule from its parts. Schematic spectrum.

How is that different from how a computer reads a molecule?

What does a molecule look like to a computer?

• Input: each atom’s element and 3D position, exactly what an .xyz file holds
• Output: the energy E, and the force on every atom
• There are many ways to give a molecule to a computer: 3D coordinates, adjacency matrices, molecular graphs, SMILES strings, fingerprints

The model never sees a label like “carbonyl” or “ring”: only elements and distances.

J. Behler, M. Parrinello, Phys. Rev. Lett. 98, 146401 (2007).

Avoid “graph”, “message passing”. Neighbourhood + sum over atoms is enough.

Then does machine learning read molecules the way a chemist does?

Does machine learning read molecules this way?

• The model learns from whole molecules, one geometry at a time
• Nothing forces it to learn “parts” the way a chemist does
• It might learn rules, or it might memorise the molecules it was shown

If every part of a new molecule is familiar, does the model find the whole familiar?

B. Lake, M. Baroni, Proc. ICML (2018).

Analogy: a student who memorised answers for 20 molecules vs one who learned group frequencies. Both score well on the homework; only one passes the exam.

Where to get these data, and how?

Any method can supply the data and become a teacher

• From CCSD(T) to GFN2-xTB, every method gives energies and forces for a geometry
• So any of them can supply training data
• Datasets range from ten small molecules (MD17) to over 100 million calculations (OMol25)

A model is at best as accurate as the method that produced its data.

S. Chmiela et al., Sci. Adv. 3, e1603015 (2017); C. Bannwarth, S. Ehlert, S. Grimme, J. Chem. Theory Comput. 15, 1652 (2019); D. S. Levine et al., arXiv (2025).

If these methods already give us energies and forces, why not just keep using them?

Why we invented semi-empirical methods

• Cost grows steeply with size: roughly N⁴ for Hartree–Fock, N⁵ for MP2, N⁷ for CCSD(T)
• Molecular dynamics needs energies and forces at millions of time steps
• Semi-empirical methods replaced expensive integrals with fitted parameters, trading some accuracy for speed

Every step down the ladder makes the same trade: accuracy for time and computer resources.

M. J. S. Dewar, W. Thiel, J. Am. Chem. Soc. 99, 4899 (1977); W. Thiel, WIREs Comput. Mol. Sci. 4, 145 (2014).

Let the room answer “why not keep using them?” before revealing the bullets. They know the answer: cost.

Machine learning makes a similar trade, by fitting to data. So what kind of method is it?

Enter machine-learning potentials

• Learn from reference calculations once
• Then predict energies and forces in milliseconds, close to the accuracy of the reference
• Nanosecond MD at near-reference accuracy becomes routine

Is machine learning a numerical, a semi-empirical, or an empirical method?

J. Behler, M. Parrinello, Phys. Rev. Lett. 98, 146401 (2007); J. Behler, Chem. Rev. 121, 10037 (2021).

Don’t answer the question. Pause and let them think, then move on.

Whatever it is, it can learn from huge datasets. What if one model learned all of chemistry?

Universal models

• Train once on millions of DFT calculations across chemistry
• One model for many molecules, with no new DFT for each system
• Examples: MACE-OFF, MACE-MP-0, PET-MAD, models trained on SPICE and OMol25

The promise: download a model and simulate your molecule. But what has the model actually learned about chemistry?

I. Batatia et al., NeurIPS (2022); D. P. Kovács et al., J. Am. Chem. Soc. (2025); P. Eastman et al., Sci. Data 10, 11 (2023); A. Mazitov et al., arXiv (2025).

End on the question. This is the bridge to Motivation.

Even with millions of molecules, what happens where the data runs out?

What the data covers

• A model is only checked where its training data are
• MD at room temperature samples the energy well densely, and the steep repulsive wall almost never
• On the wall the model has to guess, and it guesses wrong

High-energy geometries the data does not cover: one kind of out-of-distribution.

J. Behler, Chem. Rev. 121, 10037 (2021); X. Fu et al., Trans. Mach. Learn. Res. (2023).

Schematic Morse-like curve. The tick marks along the bottom show where the MD samples fall.

Now think of a reaction. Which point on the path does MD visit least?

Reactions: the rarest point matters most

• MD spends almost all its time in the reactant and product wells
• The transition state is visited rarely, so it is barely in the data
• At 298 K, every 1.4 kcal/mol of barrier error is a factor of 10 in the rate

Rare events the data barely covers: a second kind of out-of-distribution.

H. Eyring, J. Chem. Phys. 3, 107 (1935).

From k ∝ exp(−ΔG‡/RT): RT ln 10 ≈ 1.36 kcal/mol at 298 K. Schematic profile.

Steep walls, rare events… what else counts as out-of-distribution?

Many ways to be out-of-distribution

Out-of-distribution: anything the training data does not represent, where the model has to go beyond its data rather than fill in between it.

• High-energy geometries; rare events
• Higher temperature than in training
• Elements, charges or spin states never seen
• Solvent or surfaces, when trained in vacuum
• A new combination of known groups

S. Chmiela et al., Sci. Adv. 3, e1603015 (2017); X. Fu et al., Trans. Mach. Learn. Res. (2023).

All but the last change the conditions. The last one changes only the composition: every part is covered, the whole is not.

That last one is the least tested. What would a test of it look like?

Our choice: scaffold out-of-distribution

• Fusion: –OH and C=O seen apart; tested together on one carbon, as –COOH
• Duplication: one –COOH seen; tested with two

Every group is in training; only the molecule is new.

A. M. Nourollah, I. Khalid, S. Leoni, S. Schockaert, NeurIPS (2026), Evaluations & Datasets Track; arXiv:2605.08988.

Two of the five GMD tasks, shown as examples; the full set comes in the next section. The question: does a model understand molecules the way we do, from their parts?

To run this test, we need a dataset built for it. That is GMD.

GMD: Generalisation for Molecular Dynamics

• Carbon chains carrying –OH, –CHO, –COOH or –NH₂
• MD for each molecule, DFT energies and forces for every snapshot
• Training and test sets contain different molecules, not just different snapshots

Usually models are tested on new shapes of known molecules. GMD tests them on new molecules.

A. M. Nourollah, I. Khalid, S. Leoni, S. Schockaert, NeurIPS (2026), Evaluations & Datasets Track; arXiv:2605.08988.

Data and code are public: huggingface.co/datasets/AmirMasoud/GMD. Each molecule’s trajectory has about 2000 snapshots; 140 molecules in total.

So which new molecules do we test them on?

Five ways to build a new molecule from known parts

A. M. Nourollah, I. Khalid, S. Leoni, S. Schockaert, NeurIPS (2026), Evaluations & Datasets Track; arXiv:2605.08988.

Walk down the rows: abstract task on the left, real molecules on the right. Why small molecules: each task changes exactly one thing, so a failure has one cause; every part of a test molecule is in training; chemists can check what a sensible answer looks like; failing the smallest possible step is a strong result.

So how do today’s models do?

Known molecules versus new molecules

A. M. Nourollah, I. Khalid, S. Leoni, S. Schockaert, NeurIPS (2026), Evaluations & Datasets Track; arXiv:2605.08988.

Each bar runs from a model’s error on known molecules (|) to its error on new ones, one marker per task; right column is the average OOD/ID ratio. MACE-Medium zero-shot shows ×1.0 only because it is already about ten times worse on known molecules. Energies are worse still: over 1000× on new molecules.

But where inside the molecule do these errors actually appear?

Where the failure happens: inside the molecule

• Hexanoic acid: new in the fusion task
• Inner disc: error in force strength; ring: error in direction
• Backbone fine; error on –COOH, mostly strength

–OH and C=O are known. –COOH is not.

A. M. Nourollah, I. Khalid, S. Leoni, S. Schockaert, NeurIPS (2026), Evaluations & Datasets Track; arXiv:2605.08988.

Each atom’s error averaged over 200 test snapshots. NequIP spreads the error over the whole group; others concentrate it on one or two atoms. In the paper, directional errors also localise on the heteroatoms of each task’s new environment.

Combining parts fails badly. How does the simplest task, a longer chain, compare?

Chain extension: a smaller drop

• Train on ethane to hexane; test on heptane to tridecane
• Errors still rise on longer chains, but far less than in the other four tasks
• A longer chain adds more CH₂, nothing chemically new

Repeating a known part costs some accuracy. Combining parts in a new way costs far more.

SchNet and DimeNet++ are the exceptions: their errors jump by orders of magnitude past hexane.

A. M. Nourollah, I. Khalid, S. Leoni, S. Schockaert, NeurIPS (2026), Evaluations & Datasets Track; arXiv:2605.08988.

So what do we take home from all this?

Takeaways

• More of the same (longer chains) costs some accuracy; known parts in new combinations cost far more
• Wrong force strengths on a functional group mean wrong vibrations, wrong conformers, wrong barriers
• More data, bigger models and pre-training do not fix it
• Accuracy on known molecules does not predict accuracy on new ones

Before trusting a model on a new molecule, ask: has it seen this combination, or only these parts?

A. M. Nourollah, I. Khalid, S. Leoni, S. Schockaert, NeurIPS (2026), Evaluations & Datasets Track; arXiv:2605.08988.

Call back to the first slides: every molecule you care about is a new combination of familiar parts. Details on data and model size are in the appendix.

Thank you

Amir Masoud Nourollah

Stefano Leoni · Steven Schockaert

Paper: arXiv:2605.08988 (NeurIPS 2026)

Data: huggingface.co/datasets/AmirMasoud/GMD

Code: github.com/Nourollah/GMD-26

Questions?

Not a data or model-size problem

More data

Foundation models pre-trained on millions of structures (MACE-MP-0, UMA) do not consistently beat MACE trained from scratch on GMD.

Bigger models

A larger MACE-MP-0 helps on some tasks, but the gap stays.

More training coverage

Adding related molecules to the training set does not close the gap for fusion.

Another level of theory

GFN2-xTB labels instead of DFT give the same pattern.

Fine-tuning fixes the molecules the model was shown, not the ones it wasn’t.

A. M. Nourollah, I. Khalid, S. Leoni, S. Schockaert, NeurIPS (2026), Evaluations & Datasets Track; arXiv:2605.08988.

Paper Section 4.3 and Appendices A.3 and C.2. Foundation models were fine-tuned for 2000 epochs per task, an upper bound on what fine-tuning can do.

Future work

What I am working on

• A model whose learning is guided by chemical structure
• It is forced to see a molecule the way a chemist does: as functional groups and how they connect

Goal: a model that understands a new molecule from its parts.

A. M. Nourollah, I. Khalid, S. Leoni, S. Schockaert, NeurIPS (2026), Evaluations & Datasets Track; arXiv:2605.08988.

Fitting well is not generalising well

• Two kinds of points: the model learns a border between them
• A straight line misses a few training points, but gets new points right
• A wiggly border gets every training point right, and new points wrong

We want a model that works on data it has never seen, not one that remembers its training data. That is generalisation.

C. M. Bishop, Pattern Recognition and Machine Learning, Springer (2006), §1.1, §4.1.

Point at the circled new points on the right panel. The over-fitted model “learned” the noise.

What about points that land outside everything the model has seen, or outside what our data covers?