How Claude performs on NMR
prediction and structure
elucidation​
​
Assessing Claude Opus 4.7 against ChemDraw 25.0.2 and
MestReNova 17.0.0

Published                     Author

June 5, 2026                  David Kamber
Summary​

One of Anthropic’s goals is to accelerate scientific discovery. This involves training Claude
to take on researchers’ more time-intensive tasks and to give scientists capabilities they
didn’t previously have. A bench chemist’s day is full of bottlenecks: literature searches,
experimental writing, data interpretation, and spectral analysis. Spectral analysis, in
particular, is one of the most time-consuming steps in synthetic chemistry, and improving
its efficiency enables researchers to spend their time elsewhere. For this paper, we tested
whether a frontier general-purpose AI model can perform spectral analysis as well as the
dedicated Nuclear Magnetic Resonance (NMR) software chemists rely on today. We
measured three Claude models (Opus 4.7, Opus 4.6, Sonnet 4.6) against two dedicated
software (ChemDraw and MestReNova) on 20 compounds curated from synthetic
chemistry preprints. We found that for routine data prediction Opus 4.7 is now as good as
or better than ChemDraw and MestReNova on average. Dedicated structure-elucidation
software has existed for decades, but it typically requires 2D NMR (a spectrum with two
axes, and the output is a contour map rather than a row of peaks), specialized training, and
licensed tools. We also found that Claude can work the problem in reverse, proposing a
structure from NMR data alone. Claude works from exactly what a chemist would paste
into a chat: the standard readouts from a routine mass spec and NMR run, with no setup.

                                                                                            2
Summary​                                                                                                 2
Introduction​                                                                                            3
    Evaluation​                                                                                          5
    Selection protocol​                                                                                  5
    Forward prediction: structure → NMR​                                                                 6
    Inverse prediction: NMR → structure​                                                                 11
    Takeaway​                                                                                            12
    Limitations​                                                                                         13
    References​                                                                                          13

Introduction
Chemical reactions rarely produce one product cleanly, and confirming that the resultant
product is what the chemist had intended consumes the bulk of characterization time. For
chemists, the standard confirmation of products from a chemical reaction is an NMR
spectrum: a list of peaks, one per chemically distinct hydrogen or carbon, each shifted by
its surrounding atoms.

Software helps interpret these peaks in two directions. Forward prediction is the more
routine of these: draw the expected structure, predict its spectrum, and compare against
what was measured. ChemDraw and MestReNova are two tools on essentially every
chemist’s desktop that are commonly used for forward prediction of NMR spectra.

The harder direction is inverse prediction (structure elucidation): given a spectrum,
determine the structure. This requires expert reasoning about which fragments are present
and how they connect. While ChemDraw does not have this capability at all, MestReNova
helps assign peaks to a known structure but does not generate candidates from a peak list.

We assessed Claude in both directions: 20 compounds for forward shift prediction against
ChemDraw and MestReNova, and 15 for inverse structure elucidation. The forward task
tests whether a general-purpose model is competitive with the dedicated predictors
chemists already use; the inverse task tests whether it can take on the expert-reasoning
step those tools were not designed to perform. ​
​
We found that:

    ●​    ¹H NMR shift prediction (Figure 4) and multiplicity reporting (Figure 7): Opus 4.7 is
         the most accurate tool tested1
1
 Chemical shift (ppm): where a signal sits in the spectrum, reflecting an atom’s chemical environment;
predicted for ¹H and ¹³C.​

                                                                                                         3
    ●​ ¹³C NMR shift prediction. Opus 4.7’s accuracy is comparable to MestReNova’s (Figure
       2).
    ●​ J-coupling accuracy (Figure 8): Claude models uniformly outperform ChemDraw and
       MestReNova.
    ●​ Peak coverage. ChemDraw’s main strength is breadth: it maintains the widest peak
       coverage across both nuclei (Figure 2), even where its coupling accuracy lags.

On the inverse task (NMR/HRMS → structure), Opus 4.7 was asked to propose candidate
SMILES for 15 published targets drawn from the same ChemRxiv preprint pool, spanning a
deliberate range of scaffold density. At the simpler end: anilino-chloropyrimidines, a
Boc-N-aryl maleimide, an oxazolidinone-iodobenzamide, and a triethylsilyl sulfonamide. At
the more complex end: fused chloropyridazines, an N-acyl oxazinane, spirocyclic ketones
with phenacyl or acetyl pendants, an α-aryl triethylsilyl sulfonamide, and a gem-disilyl
sulfonamide. ​
​
From 1D NMR and HRMS alone, Opus 4.7 recovered the published target on every attempt
for the simpler scaffolds. For the structurally denser targets, supplementing the spectra
with the starting-material SMILES, with no other reaction context (no reagents, conditions,
mechanism, or product class), was sufficient for Opus 4.7 to reach the correct structure on
most or all attempts across the set. This extends automated structure elucidation from 1D
data into the scaffold-density regime where chemists currently default to 2D experiments
or manual interpretation, using only information already in hand at the point of a
confirmatory NMR. Opus 4.7 spans both forward and inverse: it leads or ties for the lead on
shift accuracy, multiplicity, and J-coupling, and it makes inverse elucidation tractable in a
setting where ChemDraw and MestReNova cannot operate.

Evaluation
To set up this assessment, we drew compounds from ChemRxiv synthetic chemistry
preprints and selected them by hand per the Section 3 protocol: a chemist read each
preprint, kept compounds whose reported NMR and HRMS data were complete and
self-consistent, and transcribed the peak lists manually (section 3). We produced two sets
of compounds: 20 for forward shift/coupling prediction (4 scaffold classes × 5 compounds;
Figure 1) and 15 for inverse structure elucidation. The four forward-task scaffolds are

Multiplicity: a ¹H signal's splitting pattern (singlet, doublet, triplet, etc.), set by its number of neighboring
hydrogens.

J-coupling: the coupling constant (J, in Hz), the spacing between split sub-peaks, measuring how strongly two
nuclei interact through bonds.

Peak coverage: the fraction of observed atoms a tool actually predicts.

                                                                                                                    4
referred to as P1 (chloropyridazines), P2 (Boc-N-aryl maleimides + N-Boc ynamides), P3
(spiroketones), and P4 (α-silyl methanesulfonamides) below.

Figure 1. Four scaffold classes covered by the forward-prediction benchmark. Each probes a different
category of NMR challenge. P1 chloropyridazines have a slow-exchange NH on aminopyridazine in
DMSO-d₆; P2 Boc-N-aryl maleimides and N-Boc ynamides exercise α-vinyl-imide carbonyls and the
rare ynamide α/β-carbon pair; P3 spiroketones are spirobicyclic ketones with phenacyl or acetyl
pendants and diastereotopic CH₂; P4 α-silyl methanesulfonamides have shielded silicon-α carbons. Five
compounds per class, n = 20 total.

Selection protocol
To avoid selection bias, we chose and locked the compounds before generating any
predictions from Claude or the comparison tools. From ChemRxiv synthetic chemistry
preprints with full ¹H and ¹³C characterization in the supporting information (SI), we took
the first fully characterized novel compounds per preprint, excluding compounds reported
as mixtures of rotamers, extracting SMILES, ¹H and ¹³C peak lists, and NMR solvent.
Forward (n=20) and inverse (n=15) sets were drawn separately under the same protocol.

Forward prediction: structure → NMR

Each tool was given a SMILES and asked to predict ¹H and ¹³C shifts—with multiplicity and
scalar spin-spin coupling constants (J, in Hz) where applicable—in the solvent of the
original report. The three Claude models were queried three times per compound to
characterize run-to-run variability; ChemDraw and MestReNova are deterministic, so were
run once per compound. Predicted peaks were matched to experimental atoms one-to-one
by minimum |Δδ| (Hungarian assignment), with ¹H multiplets reported as ranges wider than

                                                                                                    5
0.3 ppm were dropped from the shift-error calculation but retained for multiplicity. We
report MAE, RMSE, and median |Δδ| against fixed denominators of 401 ¹H and 225 ¹³C atoms;
the headline tolerance metric is the fraction of atoms predicted within ±0.20 ppm (¹H) or
±1.0 ppm (¹³C), and the per-compound win rate counts compounds where each tool gives
the lowest MAE, computed once per LLM replicate and reported with its min–max range.

On ¹H, Opus 4.7 has an MAE of 0.079 ppm, the lowest of any tool tested, and ranks first on
the fraction of atoms within the ±0.20 ppm tolerance window (Figures 2–4). On ¹³C, Opus
4.7 (1.37 ppm MAE) and MestReNova (1.48 ppm) are comparable across the 20-compound
set; the remaining tools hold the same rank order as on ¹H. The per-compound win rate
makes the same point at the sample level: Opus 4.7 wins on most compounds on both
nuclei, while MestReNova's wins are split across chloropyridazines and spiroketones, with
spiroketones the largest single bucket on ¹³C and the only scaffold on which MestReNova
wins any ¹H compounds (Figure 4, bottom). Note that MestReNova's ¹H MAE is computed on
partial coverage (267/401 atoms; the >0.3 ppm experimental multiplets it skips are dropped
from its score), so it is not directly comparable to fully covered tools.

The slow-exchange NH proton sits in a narrow 6.8–7.9 ppm window experimentally; it
drifts upfield in Opus 4.7, scatters across several ppm in Opus 4.6, and is misplaced into the
10–13 ppm regime by Sonnet 4.6. The aromatic ring carbons are predicted systematically
low by both Opus models, while Sonnet 4.6’s errors on those atoms are large but
undirected. Outside chloropyridazines, the bias picture is benign: Boc-N-aryl maleimides,
spiroketones, and α-silyl methanesulfonamides give Opus 4.7 either the lowest scaffold
MAE or the lowest residual spread (Figures 5–6). One common-mode artifact is worth
flagging: every tool predicts carbonyl carbons slightly low, so a two-tool sanity check that
requires agreement on δ(C=O) will not detect it.

Beyond shift accuracy, Claude models hold their edge on coupling-pattern reporting. Opus
4.7 returns the highest exact-multiplicity rate of any tool under one-way leniency (Figure
7); ChemDraw’s low exact rate is a scoring artifact rather than a capability gap, since it
routinely emits concrete multiplicity calls where the experimentalist reported only ‘m’. On
J-values (Figure 8), all three Claude models cluster around 0.5 Hz mean |ΔJ| and reach
80–84% of pairs within ±0.5 Hz; the classical tools sit at 1.9–2.0 Hz mean |ΔJ| and reach only
26–35% within ±0.5 Hz, with ChemDraw’s template defaults inflating its error on aromatic
systems—most visibly the 12.4 Hz geminal coupling, which appears in 5 of its 31 J calls;
aromatic vicinal couplings are emitted near 7.0–7.1 Hz.

Across replicates, Opus 4.7 is the most stable model: its run-to-run MAE varies by less than
the gap to the next tool (Figure 2), while the per-compound winner shifts often enough
across the 20-compound set that single-replicate assessments would substantially misstate
win rates (Figure 4, bottom).

                                                                                             6
Figure 2. Per-tool MAE (solid) and RMSE (hatched) for ¹H (left) and ¹³C (right) shift errors across 20
compounds for forward prediction, with coverage shown beneath each tool. Claude bars: mean across
three replicates with min–max range and overlaid replicate dots. Classical tools: single-point
predictions (no range).

Figure 3. Distribution of absolute shift error |Δδ| per tool across all matched pairs, shown for ¹H (left)
and ¹³C (right) on a log scale. Box = IQR; whiskers = 1.5×IQR; small grey circles = outliers; gold diamond
= mean; thick black bar = median. Claude models pool 3 replicates; classical tools are a single
observation set.

                                                                                                             7
Figure 4. Top: % of experimental atoms within ±0.20 ppm (¹H, left) and ±1.0 ppm (¹³C, right). Bottom:
per-compound win rate (compounds where the tool had the lowest per-compound MAE, out of 20).
Claude bars: mean across three replicates with min–max range; classical tools: single-point
predictions.

                                                                                                        8
Figure 5. Per-tool signed shift bias by scaffold (Δ = pred − exp). One panel per (nucleus, scaffold)
combination. Translucent dots are individual matched atoms; diamonds (Claude models) and squares
(classical tools) mark the mean signed Δ, with error bars showing ±1 SD across the three LLM replicates.
Mean ± SD values are printed below each tool. Colored bands flag bias regions (¹H |Δ| > 0.5 ppm; ¹³C |Δ| >
3.0 ppm); points with |Δ| beyond the displayed y-axis range (¹H 0.7 ppm; ¹³C 4.0 ppm) are summarized
as triangles at the panel edges with their counts. For LLMs, error bars and ±SD labels show the spread
of the per-replicate mean across 3 runs; classical tools are deterministic single-point predictions and
carry no replicate SD.

Figure 6. Atom-level prediction-error spread per (tool × scaffold). σ = standard deviation of signed Δ
across atoms within the scaffold (LLM cells pool 3 replicates). σ is the headline metric for spread: a tool
with low σ predicts the scaffold consistently (tight residuals around its mean bias), even if the mean bias
itself is non-zero. Two panels (¹H, ¹³C); cell color is normalized within each column to its maximum σ
(darker = more spread relative to the worst tool on that scaffold), with absolute σ printed in each cell.

                                                                                                         9
Figure 7. ¹H multiplicity agreement across 20 compounds. Stacked bars per tool show how each
predicted multiplicity was scored against the experimental call, in four categories: exact (the predicted
multiplicity matches experiment), lenient (experiment was reported as “m” but the tool committed to a
specific call), wrong (concrete disagreement, or experiment was concrete and the tool predicted "m"),
and missing (no predicted peak aligned to this experimental atom). Scoring uses a strict denominator of
401. Every anchored ¹H atom across the 20 compounds is counted, including atoms a tool failed to
predict (which fall into missing), so coverage gaps cannot be hidden by simply omitting hard atoms.
Brackets above the LLM bars span the min–max range of % exact across the three independent
replicates per compound; classical tools (MestReNova, ChemDraw) are deterministic single predictions
and have no bracket.

Figure 8. J-coupling accuracy across 20 compounds, summarized as a tolerance-band heatmap. Cells
show the percentage of matched experimental–predicted J pairs (33 anchored experimental pair-points
across the 20 compounds; per-tool matched counts are shown beside each tool label and reflect the
3-replicate pool for LLMs and the single deterministic run for classical tools) that fall within
increasingly strict tolerance thresholds, from ≤ 5 Hz on the right to ≤ 0.1 Hz on the left. Color encodes
the percentage on a diverging green-to-red scale (green = high pass rate, red = low). For LLMs,
predictions pool the three replicate runs.

Inverse prediction: NMR → structure

Opus 4.7 was given 15 NMR/HRMS elucidation problems and asked, three times each, to
return up to three ranked SMILES candidates, with stereochemistry excluded since 1D
NMR alone cannot define absolute stereochemistry. The 15 problems were run under two
prompt conditions, chosen by structural complexity. For 8 of the simpler targets (Q1–Q5,
Q9, Q10, Q14), Opus 4.7 received only the HRMS and 1D NMR. For the remaining 7
structurally denser targets (Q6–Q8, Q11–Q13, Q15), the same data plus the starting-material
SMILES were provided, since accurate prediction on these scaffolds required that

                                                                                                       10
additional anchor. No other reaction context (reagents, conditions, mechanism, expected
scaffold) was provided in either condition. The two conditions reflect the two situations a
chemist actually faces: confirming the product of an unspecified reaction, versus
confirming the product of a reaction whose inputs are known.

Recovery tracks structural complexity (Figure 9). The single-ring and two-fragment
scaffolds (anilino-chloropyrimidines, Boc-N-aryl maleimide, oxazolidinone-iodobenzamide,
triethylsilyl sulfonamide) were recovered on every attempt from HRMS and spectra alone.
The structurally denser targets (fused chloropyridazines, an N-acyl oxazinane, spirocyclic
carbonyl systems, an α-aryl triethylsilyl sulfonamide, the gem-disilyl sulfonamide) required
the starting-material SMILES; under that condition, Opus 4.7 predicted the correct
candidate structure on all three attempts for Q6, Q8, Q12, and Q15, and on two of three
attempts for Q7, Q11, and Q13.

Figure 9. Structure-elucidation results across the 15 inverse-task problems. Each panel shows the
published target with its success count out of 3 attempts. Border color indicates the prompt condition:
green for spectra and HRMS only with no starting-material context; blue for spectra, HRMS, and the
starting-material SMILES, with no other reaction context.

Takeaway

Two findings stand out. First, an LLM matches the strongest dedicated NMR tool on
forward ¹³C shift prediction (1.37 vs 1.48 ppm MAE) and records the lowest ¹H error of any

                                                                                                          11
tool (0.079 ppm), across 20 compounds in four scaffold classes. Second, the same model
makes inverse elucidation from 1D NMR + HRMS tractable, a problem that has historically
been underserved by software, with dedicated elucidation packages either requiring 2D
experiments or assuming a candidate structure to assign against, leaving the 1D-only case
to expert reasoning. Together these change what a working chemist can ask of a general
reasoning model. Confirming the product of a known reaction, ruling out a regioisomer,
sanity-checking an assignment, and triaging which compounds need 2D experiments are all
tasks that previously required either dedicated software or expert reasoning; a single
model now handles each of them as plain text. The practical effect is fewer steps that
demand expert attention on the routine cases, and more time available for the experiments
and decisions that genuinely need it.

Limitations

Several caveats bound the conclusions. The assessment was small—20 compounds across
four scaffolds for the forward task, 15 for the inverse task—and each scaffold contributes a
single class of failure modes, so the numerical rankings should be read as indicative rather
than precise. Slow-exchange NH heteroaromatics are sampled only through
chloropyridazines, leaving related systems (hydroxypyridines, aminothiazoles, and other
DMSO-d₆ NH-active scaffolds) untested; 2D experiments (COSY, HSQC, HMBC) and
stereochemistry are out of scope by design, since 1D NMR alone cannot fix configuration.
Solvent coverage is limited to DMSO-d₆, CDCl₃ and D₂O, so methanol-d₄, benzene-d₆, and
acetone-d₆ are not assessed.

References

Kordubailo, M. V.; Borysov, O. V.; Vlasov, S. V.; Ryabukhin, S. V.; Volochnyuk, D. M. The
Assembly of Fused Pyridazines via One-Pot Sequence IEDDA Reaction, Scope and
Limitations. ChemRxiv, April 22, 2026. https://doi.org/10.26434/chemrxiv.15002316/v1.

Heymans, T.; Durant-Baudet, L.; Landrain, Y.; Evano, G. Straightforward and Modular
Maleimide Synthesis from Ynamides. ChemRxiv, 2026.
https://doi.org/10.26434/chemrxiv.15002423/v1.

Strong, C. S.; Chen, P.-J.; Bissenali, S.; Goddard, W. A., III; Stoltz, B. M. Enantioselective
Michael Spirocyclization of Palladium Enolates. ChemRxiv, 2025.
https://doi.org/10.26434/chemrxiv-2025-59lfh.

                                                                                                 12
Leitch, M. A.; Ding, R.; Garrigues, S. L.; Chiong, H.; Mwiva, J.; Miller, K. E.; Liu, P.; Wang,
Y.-M. Access to an Elusive Soft Deprotonation Manifold: Direct α-C(sp3) Silylation of
Sulfonyl Compounds. ChemRxiv, April 21, 2026.
https://doi.org/10.26434/chemrxiv.15002274/v1.

                                                                                                  13