Core research line

Machine learning on the fingerprint region

The hypothesis: below 1500 cm⁻¹, the coupled skeletal modes encode molecular identity, while above 1500 cm⁻¹ the local modes encode functional class. If true, the fingerprint region — usually treated as too complex to read — is the richest channel for learned compound identification and interaction sensing. Phase 1 tests exactly that, on computed spectra.

The hypothesis, stated testably

H1 — Class from the high region. A model predicting functional groups should do well using only >1500 cm⁻¹.
H2 — Identity from the fingerprint. For nearest-spectrum compound retrieval, the <1500 cm⁻¹ region should outperform the >1500 cm⁻¹ region.
H3 — Discrimination of look-alikes. Among molecules with identical functional-group labels, the fingerprint region should separate them better than the high region.

Phase 1 — proof of concept

Component	Setup
Dataset	146 curated small molecules; GFN2-xTB IR computed on the platform's quantum-chemistry backend (0 failures).
Representation	Lorentzian-broadened, 181-pt vector over 400–4000 cm⁻¹ (20 cm⁻¹ bins).
Labels	21 functional-group classes from RDKit SMARTS (deterministic ground truth).
Models	Random forest for group inference; cosine nearest-neighbour for compound retrieval.
Regions compared	FULL (400–4000) · FINGERPRINT (<1500) · HIGH (>1500).

Result 1 — functional-group inference (H1)

Metric	Score	Strong per-class examples
Macro-F1	0.61	acid 1.00 · carbonyl 0.97 · alcohol 0.85 · aromatic 0.78 · nitrile 0.71
Micro-F1	0.76

Read: functional class is recoverable from a computed spectrum even at this small scale. Macro < micro reflects rare classes with few positives — expected, and a scaling target.

Result 2 — compound retrieval by region (H2)

Region	Top-1	Top-3
FULL (400–4000)	0.57	0.71
FINGERPRINT (<1500)	0.42	0.62
HIGH (>1500)	0.36	0.52

Read: the fingerprint region beats the high region for identifying which molecule a spectrum belongs to (0.42 vs 0.36 top-1; 0.62 vs 0.52 top-3) — H2 supported. The full spectrum is best, as expected, since it carries both channels.

Result 3 — separating look-alikes (H3)

Across 400 molecule pairs with identical functional-group labels, mean cosine distance between spectra:

Region	Mean cosine distance
FINGERPRINT (<1500)	0.46
HIGH (>1500)	0.40

Read: when functional groups are the same, the fingerprint region pulls the spectra further apart than the high region (0.46 vs 0.40) — i.e. it carries the residual identity signal that group frequencies miss. H3 supported.

Honest caveats — this is Phase 1

N is small (146 molecules); spectra are computed, not measured, so absolute accuracy is bounded by GFN2-xTB and the harmonic approximation, and a model could partly learn the generator's systematics. Retrieval is on held-out computed spectra, not lab data. The results establish direction and feasibility, not deployable accuracy. Every number above is reproducible from the generation/training scripts on the compute node.

Domain-shift validation — computed vs measured

The decisive test for a synthetic-data programme: does a model built on GFN2-xTB computed spectra recognise real measured ones? We paired 999 of the 1,000 molecules in the NIST experimental database with a GFN2-xTB calculation of the same molecule (SMILES pulled from PubChem and formula-validated against NIST — 1 mismatch dropped, 0 SMILES/calc failures), put both on a common 181-pt absorbance grid, and measured transfer.

Test	In-domain (measured)	Computed → measured
Functional-group F1 (macro / micro)	0.57 / 0.81	0.14 / 0.39
Retrieval — same-class top-1 (full)	—	0.36
Retrieval — same-class (fingerprint / high)	—	0.23 / 0.24
Retrieval — exact-molecule top-1 (full)	—	0.05
Best after frequency-scaling sweep	—	0.39 (×0.98)

Read — the gap is real but not catastrophic at scale. Growing the measured set from 141 → 999 molecules more than doubled the in-domain ceiling (measured-only macro-F1 0.27 → 0.57) — the database growth paid off directly. Cross-domain transfer is still weak (0.14 macro) but a measured spectrum retrieves a same-functional-class computed molecule 36% of the time — real class signal across the gap. Exact-molecule retrieval is low (0.05) because a 1,000-molecule library is dense with near-identical homologs (pentane/hexane/heptane). A global frequency scaling factor still doesn't close the gap (optimum near native, ×0.98) — the mismatch is in intensities, anharmonicity and rotational structure, not a simple offset.

What this means for the programme

So the synthetic library is not a drop-in substitute for measured data — but the scaled learning curve below shows it is not useless either. The research directions remain: (1) better physics — anharmonic VPT2 rather than harmonic frequencies; (2) domain adaptation — fine-tune on the measured anchor set; (3) keep growing the measured database, which moved the ceiling more than anything else (0.27 → 0.57 macro-F1 from 141 → 999 molecules).

When does the computed library help? (the cold-start result)

We train a functional-group model on the full computed library plus k measured spectra, and compare against k measured spectra alone, scoring on held-out measured spectra (5-fold). At 999 molecules a clear pattern emerges that the small (141) set could not show:

Measured spectra in training (k)	Measured-only (macro / micro)	Computed + measured
0	—	0.13 / 0.39
10	0.08 / 0.29	0.15 / 0.44
20	0.12 / 0.38	0.16 / 0.46
40	0.17 / 0.50	0.16 / 0.48
80	0.24 / 0.58	0.20 / 0.51
320	0.45 / 0.72	0.37 / 0.67
all (~800)	0.56 / 0.80	0.52 / 0.77

Read — synthetic spectra help exactly when real data is scarce. In the cold-start regime (k ≤ 20) the computed library clearly beats measured-only (k=10: 0.15 vs 0.08 macro) — harmonic spectra are a useful prior when you have almost no labels. The benefit vanishes by k ≈ 40 and reverses beyond it: once enough measured spectra exist, the computed library becomes a distractor (k=320: 0.37 vs 0.45). This refines the earlier small-sample conclusion into an actionable rule: use synthetic IR augmentation when measured data is scarce, drop it once you have plenty — and the fastest way to raise the absolute ceiling is simply more measured spectra (green column climbs to 0.56). Numbers are RandomForest on broadened harmonic spectra; the crossover point will shift with task, representation and better physics.

PDB-grounded structure learning

Experimental protein IR at scale doesn't exist openly — so we ground the learning in the Protein Data Bank: real 3D structures provide true (sequence, secondary-structure) labels, we forward-simulate the amide I band from each fold, and learn the inverse. The structure is the ground truth; the IR is computed. 76 single-domain chains across fold classes (mean 37% helix, 23% sheet).

amide I → secondary-structure content	helix MAE	sheet MAE	R²
Baseline (predict dataset mean)	22%	18%	—
Clean simulated band (SS only)	3.0%	1.2%	0.97 / 0.98
Realistic (+ side-chain interference + noise)	3.7%	3.6%	0.96 / 0.94

Read — when the target is structure, IR + ML works. The model recovers secondary-structure content to ~3–4% even when side-chain absorptions (Asn/Gln/Arg/Asp/Glu/Tyr/His) overlap the amide I band — the genuine confound that complicates real protein FTIR — versus 18–22% for guessing the mean. This reproduces the established FTIR protein-structure analysis computationally, validated on real PDB folds, and is the positive counterpart to the small-molecule domain gap above: exact-identity transfer is hard, but structure/composition is well within reach. It is also the data-generation template for the harder targets — per-residue structure, and coupling to the sequence-feasibility loop.

Honest scope

The forward model is our amide I band-shape (secondary-structure components + side-chain bands + noise); real measured protein FTIR adds the strong water background, ATR/scattering effects and conformational heterogeneity. The result establishes that the information is recoverable and the pipeline is sound; the next iterations scale the PDB set, move to per-residue secondary structure, and replace the analytic forward model with computed amide I from real coordinates.

Objective 2 — noncovalent interactions (design)

The next experiment turns the same pipeline onto interactions. A hydrogen bond red-shifts, broadens and intensifies the donor X–H stretch by an amount that tracks bond strength (the Badger–Bauer relation). The plan: compute spectra of monomers vs H-bonded dimers/complexes and train a classifier to detect and grade the interaction from the band shift — first on O–H···O and N–H···O, then halogen bonding and ion pairing.

Δν̄(X–H) ∝ −ΔH_H-bond // red-shift grows with interaction strength

Roadmap

Phase	Goal	Needs
1 ✓	Feasibility: class, identity-by-region, look-alike separation	done (146 mol, computed)
2	Noncovalent detector: monomer vs complex band-shift classifier	dimer/complex generation set
3 ◐	Scale to 10³–10⁴ molecules; measured anchor set (1,000 NIST spectra live, growing); computed-vs-measured domain-shift study done (see above) → gap quantified, next is domain adaptation + anharmonic spectra	local GPU compute node
4	Deep models (1D-CNN / transformer) over raw spectra; amide-shape → secondary structure	scaled labelled corpus

Compute

Phases 3–4 are compute-bound. The platform's quantum-chemistry backend already runs the GFN2-xTB calculations on demand; scaling to 10³–10⁴ molecules plus deep models is the motivation for a dedicated local GPU node. Spectra are explored on-site in the simulator.