Objective 3 · sequence
Can IR read an amino-acid sequence?
The honest answer is layered. A bulk infrared spectrum cannot recover residue order — that is mass-spectrometry territory. But IR can identify individual residues by their side-chain bands. So the real question is not "can IR sequence?" but "how much of the sequence can IR resolve if it reads one residue at a time?" This page answers that — quantitatively and interactively.
The realistic path: residue-resolved IR
Three routes give IR access to individual residues rather than the whole-protein average:
Isotope-edited 2D-IR
Site-specific ¹³C / ¹³C¹⁸O labelling shifts one backbone carbonyl out of the amide I band, reading out a single chosen position's local structure. Scan positions to walk the chain.
Processive / nanopore IR
A hypothetical single-molecule readout that records each residue's vibrational signature as the chain threads past — the IR analogue of nanopore sequencing.
Edman + IR identification
Cleave residues one at a time (Edman) and identify each released amino acid by its IR fingerprint — sequencing by serial residue ID.
The hard limit, stated plainly
Every route reduces sequencing to the same core problem: identify a single residue from its IR spectrum. And there IR has a ceiling — many residues differ only in the number of aliphatic C–H groups, which IR barely distinguishes. So IR yields a reduced-alphabet read: some residues are uniquely identifiable, others collapse into indistinguishable sets. That reduced alphabet is exactly what real single-molecule protein-sequencing methods also work with — so it is a genuine, useful result, not a failure.
How separable are the 20 residues by IR?
We build each residue's reconstructed IR spectrum (backbone + side-chain bands) and cluster residues whose spectra are closer than a chosen resolving power. Slide it to see the alphabet IR can read.
Interactive: read a sequence by IR
Enter a one-letter amino-acid sequence. Each position is coloured green = uniquely read by IR or amber = ambiguous (only its IR group is known). Hover a residue for its IR group.
Per-residue IR distinguishability
How isolated each residue is in IR space — its nearest other residue and the spectral distance to it. Large distance = uniquely identifiable; small = easily confused.
| Residue | Most-confusable with | Spectral distance | IR verdict |
|---|
Validation on real computed spectra
We computed first-principles GFN2-xTB IR for all 20 residues in the standard blocked-residue model (Ace–Xaa–NHMe: backbone amides + side chain) and measured their actual separability — the same question, on quantum-chemical spectra instead of reconstructions.
| Spectral resolving power | IR groups | Uniquely resolved | Persistent confusions |
|---|---|---|---|
| High (Δ = 0.03) | 17 | 14 / 20 | Ile≈Val · Ala≈Gly · Phe≈Trp |
| Medium (Δ = 0.05) | 11 | 9 / 20 | + aromatic/aliphatic merges |
| Low (Δ = 0.08) | 6 | 4 / 20 | Asp≈Asn; most collapse |
Read — IR tops out near a 14-letter alphabet. From a clean per-residue spectrum, IR uniquely identifies 14 of 20 residues at high precision; the irreducible confusions are chemically near-identical pairs — β-branched Ile≈Val (spectral distance 0.005), small aliphatic Ala≈Gly (0.015), aromatic Phe≈Trp (0.013). Arginine is the most uniquely identifiable residue (guanidinium bands, nearest neighbour at distance 0.23); the aliphatics the least. As precision degrades, residues merge into a smaller alphabet. The fingerprint region alone (<1500 cm⁻¹) resolves ~8/20.
Closing the loop: model & verify
IR gives a reduced-alphabet read — several candidate sequences fit one spectrum — plus an independent secondary-structure call from amide I. So we close the loop: model each candidate in the IR-indicated fold, energy-minimise, and keep only the structurally feasible ones. If no candidate consistent with the IR read can form that fold, the read is self-inconsistent → re-run the IR.
Worked example (run on the GFN2-xTB / RDKit backend). Amide I says α-helix; the side-chain read is ambiguous at position 8. Each resolution is built, MMFF-minimised, and judged:
| Position-8 resolution | Model builds | Helix propensity | Loop verdict |
|---|---|---|---|
| …EAAAKA (Ala) | ✓ | 0.11 | feasible — accept |
| …ELAAKA (Leu) | ✓ | 0.13 | feasible — accept |
| …EGAAKA (Gly) | ✓ | 0.19 | feasible — accept |
| …EPAAKA (Pro) | ✓ | 0.37 | ✗ internal Pro breaks the helix → re-run IR |
Read. All four are valid molecules (they build), but only those whose residue is compatible with the IR-reported α-helix survive the loop — the Pro resolution is rejected because an internal proline cannot sit in a helix. If the IR side-chain read at a position were consistent only with structure-incompatible residues, the loop returns re-run IR: the spectrum and the fold contradict each other.
🔁 Feasibility check
Enter a candidate sequence and the secondary structure IR reported (amide I). The loop judges whether the sequence can adopt that fold.
End-to-end: predicting a sequence from IR
The capstone — per-residue IR (computed) → a trained classifier → a predicted sequence. We computed GFN2-xTB spectra for the 20 residues in peptide context, augmented them to mimic measurement variation (frequency jitter + noise), trained the residue classifier, and ran whole sequences through it.
| Test sequence | Exact (20-letter) | Reduced-alphabet |
|---|---|---|
| Insulin B-chain (30 aa) | 29/30 · 96% | 96% |
| Ubiquitin N-terminus (30 aa) | 28/30 · 93% | 93% |
| Random 20-mer | 20/20 · 100% | 100% |
| Overall | 96% | 96% |
The pipeline runs end-to-end and recovers the sequence at 96% — when the classifier is trained on the same spectral context the test residues come from (in-distribution). It proves the information is present and the machinery — per-residue spectrum → residue → sequence — works.
The honest gap — why this isn't "solved" yet
That 96% is in-distribution. The decisive experiment — train on one spectral context, predict residues from a different (unseen-neighbour) context — starts at just ~22%: a residue's IR is dominated by its backbone-amide environment, which swamps the side-chain identity across contexts (the same computed→measured domain gap seen platform-wide).
More data measurably closes it. Training the residue classifier on increasingly many consistent peptide contexts (11 dipeptide contexts, ~220 computed residue spectra) lifts generalisation to genuinely unseen contexts along a clean curve — then plateaus:
| training contexts | generalisation (20-class) | reduced-alphabet |
|---|---|---|
| 1 | 0.22 | 0.23 |
| 2 | 0.27 | 0.30 |
| 3 | 0.31 | 0.35 |
| 5 | 0.35 | 0.38 |
| 7 | 0.35 | 0.39 |
So the trend is real and quantified — more data lifts cross-context accuracy ~60% (0.22 → 0.35) — but it plateaus: neighbour-context diversity alone saturates near 0.35. Closing the rest needs qualitatively more diverse, realistic training data — longer peptide contexts, conformational ensembles, solvent and anharmonic lineshapes calibrated to the computed→measured offset, and ultimately real measured residue spectra. The pipeline is sound and the lever (data scale + diversity) is identified; the remaining distance is a data-generation programme, not a missing idea.
Built for experimental data — the deployable interface
The trained model is saved with a fixed, protocol-agnostic interface: it accepts a stack of per-residue IR spectra (181 points, 400–4000 cm⁻¹, max-normalised) — computed today, experimental tomorrow — and returns a reduced-alphabet sequence with per-residue confidence. The moment residue-resolved experimental IR exists (isotope-edited 2D-IR, a processive readout), it plugs into exactly this pipeline: spectra in → sequence out. Everything upstream — the residue library, the augmentation, the feasibility loop — is already in place.
What this shows — and the honest caveats
IR cleanly resolves residues with strong, distinctive side-chain bands — aromatics (Phe/Tyr/Trp/His), the charged set, the side-chain amides, Cys (S–H), Met — but it conflates the aliphatic residues (Gly/Ala/Val/Leu/Ile/Pro) that differ only in C–H count, and pairs like Asp/Glu and Asn/Gln. So IR delivers a partial, reduced-alphabet sequence, not a full 20-letter read. The interactive tool above uses reconstructed spectra from literature band positions; the computed-spectra section confirms the picture on real GFN2-xTB residue spectra (and is in fact more optimistic — ~14/20 resolvable at high precision). True IR sequencing would further require single-residue spatial/temporal resolution (2D-IR + isotope labels, or a processive device) that does not yet exist as a turnkey method, and even ideal IR caps out near a 14-letter alphabet, not 20. The model-and-verify loop judges fold compatibility from established helix/β-sheet propensity + steric rules (an internal proline cannot be helical) plus a clash-free 3D build — deliberately not from a single vacuum force-field energy, which does not reliably rank secondary structure (peptides collapse in vacuum); a definitive energetic verdict on one candidate would use GFN2-xTB with implicit solvent.
References
- Barth, A. The infrared absorption of amino acid side chains, Prog. Biophys. Mol. Biol. 74 (2000) 141.
- Ganim, Z. et al. Amide I two-dimensional infrared spectroscopy of proteins, Acc. Chem. Res. 41 (2008) 432.
- Alfaro, J. A. et al. The emerging landscape of single-molecule protein sequencing, Nat. Methods 18 (2021) 604 — reduced-alphabet sequencing context.
Related: amide-I → secondary structure · amino-acid side-chain bands · the AI fingerprint program.