Can AI reproduce observed chemical diversity?

Benhenda, Mostapha

doi:10.1101/292177

“…Both platforms occasionally generated very small, non-specific structure predictions (for example, a single unspecified amino acid or a single malonyl unit) that did not provide actionable information about the chemical structure of the encoded product; to remove these from consideration, we applied a molecular weight filter to remove structures under 100 Da output by either platform. To evaluate the internal structural diversity of each set of predicted structures, we computed the distribution of pairwise Tcs for each set 45 , taking the median pairwise Tc instead of the mean as a summary statistic to ensure robustness against outliers. Structural similarity to known natural products was assessed using the RDKit implementation of the 'natural product-likeness' score 22 , and by the median Tc between predicted structures and the known secondary metabolite structures deposited in the NP Atlas database 46 .…”

Section: Methodsmentioning

confidence: 99%

Comprehensive prediction of secondary metabolite structure and biological activity from microbial genome sequences

Skinnider

¹

,

Johnston

²

,

Gunabalasingam

³

et al. 2020

View full text Add to dashboard Cite

Novel antibiotics are urgently needed to address the looming global crisis of antibiotic resistance. Historically, the primary source of clinically used antibiotics has been microbial secondary metabolism. Microbial genome sequencing has revealed a plethora of uncharacterized natural antibiotics that remain to be discovered. However, the isolation of these molecules is hindered by the challenge of linking sequence information to the chemical structures of the encoded molecules. Here, we present PRISM 4, a comprehensive platform for prediction of the chemical structures of genomically encoded antibiotics, including all classes of bacterial antibiotics currently in clinical use. The accuracy of chemical structure prediction enables the development of machine-learning methods to predict the likely biological activity of encoded molecules. We apply PRISM 4 to chart secondary metabolite biosynthesis in a collection of over 10,000 bacterial genomes from both cultured isolates and metagenomic datasets, revealing thousands of encoded antibiotics. PRISM 4 is freely available as an interactive web application at http://prism.adapsyn.com.

show abstract

“…• The internal diversity 37 , defined as the mean Tanimoto coefficient between all pairs of molecules generated by the model. Extended connectivity fingerprints 64 with a diameter of 3 and a length of 1,024 bits were used as input to the calculation of the Tanimoto coefficient.…”

Section: Methodsmentioning

confidence: 99%

“…To accomplish this goal, we calculated a suite of 23 different metrics that have previously been proposed for the evaluation of generative models of molecules 18,23,24,34,[37][38][39] . In addition to the proportion of valid SMILES strings, we also computed the proportions of unique and novel molecules generated by the model (Supplementary Fig.…”

Section: Deep Generative Models Learn From Limited Training Datamentioning

confidence: 99%

Deep Generative Models Enable Navigation in Sparsely Populated Chemical Space

Skinnider

¹

,

Rg²,

Ds³

et al. 2021

Preprint

View full text Add to dashboard Cite

Deep generative models are powerful tools for the exploration of chemical space, enabling the on-demand gener- ation of molecules with desired physical, chemical, or biological properties. However, these models are typically thought to require training datasets comprising hundreds of thousands, or even millions, of molecules. This per- ception limits the application of deep generative models in regions of chemical space populated by only a small number of examples. Here, we systematically evaluate and optimize generative models of molecules for low-data settings. We carry out a series of systematic benchmarks, training more than 5,000 deep generative models and evaluating over 2.6 billion generated molecules. We find that robust models can be learned from far fewer examples than has been widely assumed. We further identify strategies that dramatically reduce the number of molecules required to learn a model of equivalent quality, and demonstrate the application of these principles by learning models of chemical structures found in bacterial, plant, and fungal metabolomes. The structure of our experiments also allows us to benchmark the metrics used to evaluate generative models themselves. We find that many of the most widely used metrics in the field fail to capture model quality, but identify a subset of well-behaved metrics that provide a sound basis for model development. Collectively, our work provides a foundation for directly learning generative models in sparsely populated regions of chemical space.

show abstract

“…We then quantified the similarity of the distributions observed for generated molecules and the training set using the Jensen-Shannon divergence. To specifically assess the diversity of the generated molecules, we calculated the mean Tanimoto coefficient between random pairs of generated molecules, or random pairs of generated and training set molecules, to obtain the internal and external diversities, respectively 37 . Finally, we computed the Fréchet ChemNet distance 38 , a metric based on the predicted biological activities of the generated molecules that was developed specifically for the evaluation of chemical generative models.…”

Section: Deep Generative Models Learn From Limited Training Datamentioning

confidence: 99%

Deep Generative Models Enable Navigation in Sparsely Populated Chemical Space

Skinnider

¹

,

Stacey²,

Wishart³

et al. 2021

Preprint

View full text Add to dashboard Cite

Deep generative models are powerful tools for the exploration of chemical space, enabling the on-demand gener- ation of molecules with desired physical, chemical, or biological properties. However, these models are typically thought to require training datasets comprising hundreds of thousands, or even millions, of molecules. This per- ception limits the application of deep generative models in regions of chemical space populated by only a small number of examples. Here, we systematically evaluate and optimize generative models of molecules for low-data settings. We carry out a series of systematic benchmarks, training more than 5,000 deep generative models and evaluating over 2.6 billion generated molecules. We find that robust models can be learned from far fewer examples than has been widely assumed. We further identify strategies that dramatically reduce the number of molecules required to learn a model of equivalent quality, and demonstrate the application of these principles by learning models of chemical structures found in bacterial, plant, and fungal metabolomes. The structure of our experiments also allows us to benchmark the metrics used to evaluate generative models themselves. We find that many of the most widely used metrics in the field fail to capture model quality, but identify a subset of well-behaved metrics that provide a sound basis for model development. Collectively, our work provides a foundation for directly learning generative models in sparsely populated regions of chemical space.

show abstract

Can AI reproduce observed chemical diversity?

Cited by 58 publications

References 23 publications

Comprehensive prediction of secondary metabolite structure and biological activity from microbial genome sequences

Comprehensive prediction of secondary metabolite structure and biological activity from microbial genome sequences

Deep Generative Models Enable Navigation in Sparsely Populated Chemical Space

Deep Generative Models Enable Navigation in Sparsely Populated Chemical Space

Contact Info

Product

Resources

About