DreaMS-Fluorine

1. Obtain model weights

Please contact us at roman.bushuiev@uochb.cas.cz to request the DreaMS-Fluorine model weight files (dreams_fluorine_epoch=1-step=7000.ckpt and dreams_fluorine_epoch=30-step=111000.ckpt). After receiving them, place both files in /DreaMS/dreams/models/pretrained/.

Note that DreaMS-Fluorine was trained using NIST20, so we can share the weights only with NIST library license holders. Please attach your NIST license or order confirmation to the email.

2. Run DreaMS-Fluorine

Execute the following command, where --in_dir specifies the folder containing .mzML or .mgf files (here, data/MSV000099559). The script will generate dreams_fluorine_predictions.csv in the same folder, containing predicted fluorine-presence probabilities for each MS/MS spectrum in each file. The current model version supports positive-mode data only.

python3 dreams/cli.py dreams_fluorine --in_dir data/MSV000099559

3. Examine the predictions

import pandas as pd
df = pd.read_csv('../data/MSV000099559/dreams_fluorine_predictions.csv')
df

	RT	charge	file_name	polarity	precursor_mz	precursor_target_mz	scan_number	spectrum	window_lo	window_uo	F_preds_111k_steps	F_preds_7k_steps	dformat	tag
0	810.063540	1	MO23S_030.mzML	1	304.891639	304.891632	4153	[[81.07015991210938, 83.0491714477539, 84.9596...	0.5	0.5	0.923519	0.210867	A	Only 111k checkpoint > 0.9 hit
1	34.545791	1	MO23S_030.mzML	1	241.999878	241.999878	133	[[81.07012939453125, 84.95979309082031, 84.964...	0.5	0.5	0.921477	0.217751	A	Only 111k checkpoint > 0.9 hit
2	514.087512	1	MO23S_030.mzML	1	204.138540	204.138535	2640	[[78.58903503417969, 79.05422973632812, 84.044...	0.5	0.5	0.913897	0.604710	A	Only 111k checkpoint > 0.9 hit
3	810.776700	1	MO23S_027.mzML	1	328.915390	328.915405	4179	[[81.06999206542969, 87.28488159179688, 90.947...	0.5	0.5	0.903643	0.140945	A	Only 111k checkpoint > 0.9 hit
4	809.980320	1	MO23S_027.mzML	1	304.891539	304.891541	4175	[[84.95977783203125, 85.2956314086914, 90.0553...	0.5	0.5	0.878663	0.227668	A	Only 111k checkpoint > 0.75 hit
...	...	...	...	...	...	...	...	...	...	...	...	...	...	...
11125	745.386000	2	MO23S_027.mzML	1	615.455358	615.455383	3859	[[79.05448150634766, 80.05440521240234, 81.069...	0.5	0.5	0.000000	0.000000	A	NaN
11126	745.794840	2	MO23S_027.mzML	1	593.442417	593.442444	3861	[[80.05455780029297, 81.06998443603516, 83.085...	0.5	0.5	0.000000	0.000000	A	NaN
11127	747.321360	2	MO23S_027.mzML	1	571.429528	571.429504	3868	[[80.05457305908203, 81.06989288330078, 83.085...	0.5	0.5	0.000000	0.000000	A	NaN
11128	748.533720	2	MO23S_027.mzML	1	549.416251	549.416260	3874	[[80.05448150634766, 81.06990814208984, 81.132...	0.5	0.5	0.000000	0.000000	A	NaN
11129	749.549340	2	MO23S_027.mzML	1	527.403199	527.403198	3879	[[77.15992736816406, 77.27590942382812, 80.054...	0.5	0.5	0.000000	0.000000	A	NaN

11130 rows × 14 columns

The prediction file contains standard metadata parsed from the input files, such as retention time (RT) and precursor m/z (precursor_mz), which can be used to reference the input spectra. In addition, it includes DreaMS-Fluorine predictions in the columns F_preds_111k_steps and F_preds_7k_steps. These columns represent fluorine presence probabilities predicted by two versions of the DreaMS-Fluorine model trained for different numbers of steps.

We consider a prediction to be confident if both models predict a fluorine probability of at least 0.75. Such cases are annotated in the tag column as > 0.75 hit, > 0.9 hit, or > 0.95 hit, depending on the prediction scores. Less confident predictions are labeled as Only 111k checkpoint > 0.75 hit, Only 111k checkpoint > 0.9 hit, or Only 111k checkpoint > 0.95 hit, indicating that only the model trained for 111k steps produced a confident prediction.

We further classify spectra into low- and high-quality categories (see Fig. 2b). If a spectrum is of low quality but still yields a confident prediction, the model may be hallucinating the result (for example, for spectra containing only a single signal). Such cases are therefore marked with a Low quality prefix in the tag column. If the tag value is NaN, the spectrum is not predicted to correspond to a fluorinated molecule.

The plot below summarizes these tags for three files from the example MSV000099559 dataset. While there are no highly confident predictions supported by both models, there are 18 predictions of fluorinated molecules supported by only one of the models (corresponding to the last two rows in the heatmap).

import matplotlib.pyplot as plt
import seaborn as sns

df_plot = df[df['tag'] != '']
tag_file_counts = df_plot.pivot_table(index='tag', columns='file_name', values='RT', aggfunc='count', fill_value=0)
plt.figure(figsize=(5, 5))
sns.heatmap(tag_file_counts, annot=True, fmt='d', cmap='Greens', cbar_kws={'label': 'Count'})
plt.xlabel('File Name')
plt.ylabel('Tag')
plt.title('Counts of fluorine predictions per file')
plt.show()