Richard Michael
Current Organizations / Institutes
The organizations/institutes from the scholar's latest article
Richard Michael has a total of 12 co-authors. Here are the top co-authors by number of shared publications:
- Wouter Boomsma
Department of Computer Science, University of Copenhagen, Copenhagen, Denmark
4 shared publications - Peter Mørch Groth
University of Copenhagen
3 shared publications - Jacob Kæstel-Hansen
Center for Optimized Oligo Escape and Control of Disease, Center for 4D Cellular Dynamics, Department of Chemistry University of Copenhagen Copenhagen DK‐2100 Denmark
2 shared publications - Jesper Salomon
Novonesis A/S, Biologiens Vej 2, 2800 Kongens Lyngby, Denmark
2 shared publications - Nikos S. Hatzakis
Department of Chemistry and Nano-Science Center, University of Copenhagen
2 shared publications - Pengfei Tian
National Institutes of Health
2 shared publications
Latest Publications
- Open Access
APOSM: Pairwise preference learning improves generative small-molecule design
Small-molecule lead refinement is constrained by the cost of synthesizing and assaying candidates, making the surrogate models that prioritize compounds for experimental testing central to the design process. The reliability of such surrogates is limited by the noise and sparsity of screening measurements. We show that training the surrogate on pairwise comparisons between candidate molecules, rather than on absolute predicted scores, yields a substantially more reliable signal for active candidate selection in this regime. We develop APOSM, an active-learning algorithm that combines a fragment-based generator, a pairwise message-passing graph neural network surrogate, and probabilistic ranking inside a batched acquisition loop. On the Practical Molecular Optimization benchmark and a GPCR ligand rediscovery task, APOSM improves target attainment and sampling efficiency over unguided fragment-based optimization, the Graph-GA genetic algorithm, and a pointwise-regression ablation, with the largest gains on tasks where absolute scores are hardest to calibrate.
Open AccessAPOSM: Pairwise preference learning improves generative small-molecule design
Small-molecule lead refinement is constrained by the cost of synthesizing and assaying candidates, making the surrogate models that prioritize compounds for experimental testing central to the design process. The reliability of such surrogates is limited by the noise and sparsity of screening measurements. We show that training the surrogate on pairwise comparisons between candidate molecules, rather than on absolute predicted scores, yields a substantially more reliable signal for active candidate selection in this regime. We develop APOSM, an active-learning algorithm that combines a fragment-based generator, a pairwise message-passing graph neural network surrogate, and probabilistic ranking inside a batched acquisition loop. On the Practical Molecular Optimization benchmark and a GPCR ligand rediscovery task, APOSM improves target attainment and sampling efficiency over unguided fragment-based optimization, the Graph-GA genetic algorithm, and a pointwise-regression ablation, with the largest gains on tasks where absolute scores are hardest to calibrate.
- Open Access
- 3 May 2024
- Published by Public Library of Science (PLoS) in PLOS Computational Biology
- Vol. 20 (5) , e1012061
A systematic analysis of regression models for protein engineering
To optimize proteins for particular traits holds great promise for industrial and pharmaceutical purposes. Machine Learning is increasingly applied in this field to predict properties of proteins, thereby guiding the experimental optimization process. A natural question is: How much progress are we making with such predictions, and how important is the choice of regressor and representation? In this paper, we demonstrate that different assessment criteria for regressor performance can lead to dramatically different conclusions, depending on the choice of metric, and how one defines generalization. We highlight the fundamental issues of sample bias in typical regression scenarios and how this can lead to misleading conclusions about regressor performance. Finally, we make the case for the importance of calibrated uncertainty in this domain.
- 3 May 2024
- Published by Public Library of Science (PLoS) in PLOS Computational Biology
- Vol. 20 (5) , e1012061
Open AccessA systematic analysis of regression models for protein engineering
To optimize proteins for particular traits holds great promise for industrial and pharmaceutical purposes. Machine Learning is increasingly applied in this field to predict properties of proteins, thereby guiding the experimental optimization process. A natural question is: How much progress are we making with such predictions, and how important is the choice of regressor and representation? In this paper, we demonstrate that different assessment criteria for regressor performance can lead to dramatically different conclusions, depending on the choice of metric, and how one defines generalization. We highlight the fundamental issues of sample bias in typical regression scenarios and how this can lead to misleading conclusions about regressor performance. Finally, we make the case for the importance of calibrated uncertainty in this domain.
- Open Access
FLOP: Tasks for Fitness Landscapes Of Protein wildtypes
Protein engineering has the potential to create optimized protein variants with improved properties and function. An initial step in the protein optimization process typically consists of a search among natural (wildtype) sequences to find the naturally occurring proteins with the most desirable properties. Promising candidates from this initial discovery phase then form the basis of the second step: a more local optimization procedure, exploring the space of variants separated from this candidate by a number of mutations. While considerable progress has been made on evaluating machine learning methods on single protein datasets, benchmarks of data-driven approaches for global fitness landscape exploration are still lacking. In this paper, we have carefully curated a representative benchmark dataset, which reflects industrially relevant scenarios for the initial wildtype discovery phase of protein engineering. We focus on exploration within a protein family, and investigate the downstream predictive power of various protein representation paradigms, i.e., protein language model-based representations, structure-based representations, and evolution-based representations. Our benchmark highlights the importance of coherent split strategies, and how we can be misled into overly optimistic estimates of the state of the field. The codebase and data can be accessed via https://github.com/petergroth/FLOP .
Open AccessFLOP: Tasks for Fitness Landscapes Of Protein wildtypes
Protein engineering has the potential to create optimized protein variants with improved properties and function. An initial step in the protein optimization process typically consists of a search among natural (wildtype) sequences to find the naturally occurring proteins with the most desirable properties. Promising candidates from this initial discovery phase then form the basis of the second step: a more local optimization procedure, exploring the space of variants separated from this candidate by a number of mutations. While considerable progress has been made on evaluating machine learning methods on single protein datasets, benchmarks of data-driven approaches for global fitness landscape exploration are still lacking. In this paper, we have carefully curated a representative benchmark dataset, which reflects industrially relevant scenarios for the initial wildtype discovery phase of protein engineering. We focus on exploration within a protein family, and investigate the downstream predictive power of various protein representation paradigms, i.e., protein language model-based representations, structure-based representations, and evolution-based representations. Our benchmark highlights the importance of coherent split strategies, and how we can be misled into overly optimistic estimates of the state of the field. The codebase and data can be accessed via https://github.com/petergroth/FLOP .
- Open Access
Assessing the performance of protein regression models
To optimize proteins for particular traits holds great promise for industrial and pharmaceutical purposes. Machine Learning is increasingly applied in this field to predict properties of proteins, thereby guiding the experimental optimization process. A natural question is: How much progress are we making with such predictions, and how important is the choice of regressor and representation? In this paper, we demonstrate that different assessment criteria for regressor performance can lead to dramatically different conclusions, depending on the choice of metric, and how one defines generalization. We highlight the fundamental issues of sample bias in typical regression scenarios and how this can lead to misleading conclusions about regressor performance. Finally, we make the case for the importance of calibrated uncertainty in this domain.
Open AccessAssessing the performance of protein regression models
To optimize proteins for particular traits holds great promise for industrial and pharmaceutical purposes. Machine Learning is increasingly applied in this field to predict properties of proteins, thereby guiding the experimental optimization process. A natural question is: How much progress are we making with such predictions, and how important is the choice of regressor and representation? In this paper, we demonstrate that different assessment criteria for regressor performance can lead to dramatically different conclusions, depending on the choice of metric, and how one defines generalization. We highlight the fundamental issues of sample bias in typical regression scenarios and how this can lead to misleading conclusions about regressor performance. Finally, we make the case for the importance of calibrated uncertainty in this domain.