Molecular formula prediction

Molecular Formula Prediction from NMR Spectra

Summary

Predict the molecular formula (elemental composition) of an unknown compound from 1D ¹H and/or ¹³C NMR spectra using a multitask transformer-based machine learning model. This skill enables rapid structure elucidation traditionally performed by manual spectroscopic interpretation.

When to use

Apply this skill when you have 1D ¹H or ¹³C NMR spectra from an unknown organic compound and need to predict its molecular formula (e.g., C₆H₁₂O₂). The compound must contain ≤19 heavy (non-hydrogen) atoms. Use this as the first stage of a structure elucidation pipeline before predicting molecular connectivity.

When NOT to use

• Input compound contains >19 heavy atoms — model was only validated on molecules up to C/H/N/O/S/halide with ≤19 non-hydrogen atoms.
• Input is a 2D NMR spectrum (HSQC, HMBC, COSY) — this skill is trained exclusively on 1D spectra.
• The spectrum is from a mixture or unknown impurities — the model assumes a single pure compound.

Inputs

• 1D ¹H NMR spectrum (token-encoded spectral features or raw peak list)
• 1D ¹³C NMR spectrum (token-encoded spectral features or raw peak list)
• Ground-truth molecular formulas (for training/validation)

Outputs

• Predicted molecular formula (element count vector or SMILES-like formula string)
• Prediction confidence scores (softmax probabilities per element)
• Accuracy metric on held-out test set

How to apply

Encode the input 1D NMR spectrum as a token sequence representing spectroscopic features (chemical shifts, intensities, multiplicities). Feed this sequence through a transformer encoder that learns to map spectral embeddings to a chemical formula space. The model is trained end-to-end using supervised classification loss on ground-truth molecular formulas. At inference, the transformer decoder outputs a predicted molecular formula as a vector encoding element counts (e.g., [C_count, H_count, O_count, N_count, ...]). Validate predictions by comparing predicted formula to ground truth using exact-match accuracy or elemental composition metrics. The model integrates multi-head self-attention to capture long-range spectral correlations and learns which NMR peaks are most diagnostic for formula prediction.

Related tools

• PyTorch (Neural network framework for implementing transformer encoder-decoder architecture, training supervised loss on formula classification, and inference) — https://pytorch.org
• Transformer (encoder-decoder architecture) (Core neural architecture with multi-head self-attention and cross-attention layers to map NMR spectral embeddings to molecular formula space)

Evaluation signals

• Exact-match accuracy: predicted formula matches ground-truth formula (e.g., C₆H₁₂O₂)
• Element-wise accuracy: each predicted element count (C, H, O, N, etc.) matches ground truth
• Schema invariant: output vector has fixed length equal to number of tracked elements and sums to expected molecular weight range
• Held-out test set accuracy reported separately for ¹H-only, ¹³C-only, and combined spectra inputs
• No negative element counts or implausible formulas (e.g., H₀ or C₀)

Limitations

• Validated only on molecules with ≤19 heavy atoms; performance unknown for larger molecules.
• Requires high-quality, well-resolved 1D NMR spectra; noisy, overlapped, or sparse spectra may degrade predictions.
• Model is trained on a specific dataset; generalization to novel chemical scaffolds not in training set is not characterized.
• Does not handle isotope information (e.g., ¹³C labeling) — treats all carbons uniformly.

Evidence

• [intro] A multitask machine learning framework can predict molecular structure (formula and connectivity) from 1D ¹H and/or ¹³C NMR spectra: "we introduce a multitask machine learning framework that predicts the molecular structure (formula and connectivity) of an unknown compound solely based on its 1D ¹H and/or ¹³C NMR spectra"
• [methods] Training uses supervised loss on formula prediction as classification task: "Train the transformer end-to-end using supervised loss on formula prediction (classification) and bond connectivity prediction (link prediction on molecular graph)"
• [methods] Validation involves comparing predicted formula to ground truth: "Validate the trained model on held-out test molecules by comparing predicted formula and connectivity to ground truth, computing accuracy metrics for both outputs"
• [intro] Framework is effective for molecules with up to 19 heavy atoms: "We demonstrate the effectiveness of this framework on molecules with up to 19 heavy (non-hydrogen) atoms"