The meeting presented Dr. Yilai Li’s seminar on using generative AI to improve 3D reconstruction in cryo-electron microscopy (cryo-EM). Dr. Li reviewed the growth and importance of cryo-EM in structural biology, outlined the experimental workflow (plunge freezing, 2D projection imaging, particle picking, and 3D density reconstruction), and characterized reconstruction as... ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ 👋 You've been invited to view this report because Robin D Terwilliger added Read to the meeting and wanted to share the recap with you. Yilai Li - Seminar Purdue AI‑Driven Biomolecular Design, Biotech & Computational Biology Position March 9, 2026 ( https://u25608997.ct.sendgrid.net/ls/click?upn=u001.CrUrehIev9dAOd9CS5LlcqT6... ) map[action_items:[] billing_plan:<nil> chapters:[map[chapter:Opening and Speaker Introduction topics:[]] map[chapter:Seminar Topic and High-level Motivation topics:[Dr. Li’s seminar focused on using generative AI for 3D cryo-EM reconstruction. Structural biology relies on experimental techniques like cryo-EM to determine biomolecular structures, which are essential because structure determines function.]] map[chapter:Cryo-EM Impact and Adoption topics:[]] map[chapter:Experimental Workflow and Data Collection topics:[Cryo-EM output consists of many noisy 2D particle projections that can be used to reconstruct 3D density maps after estimating orientations and translations. A gap exists: limited AI models operate directly in density-map (voxel) space compared with atomic-coordinate models.]] map[chapter:Modeling and Unknowns in the Forward Model topics:[Reconstruction is an ill-posed inverse problem with unknown rotations, translations, imaging parameters, noise, and potential heterogeneity.]] map[chapter:Inverse Problems, Priors, and Bayesian Framing topics:[Priors and Bayesian approaches can guide reconstruction toward physically plausible solutions.]] map[chapter:Prior-FM Project, Data, and Model Architecture topics:[The project will train a foundation model (prior-FM) on ~40,000 public high-resolution cryo-EM densities curated from EMDB. The team selected flow matching as the generative framework and a 3D U-Net architecture for the voxel-space model.]] map[chapter:Flow-matching prior and unconditional sampling topics:[Cryo-FM is trained to map a Gaussian distribution to cryo-EM data distribution using flow matching. The model supports unconditional sampling by iteratively following a learned vector field from Gaussian noise to the data distribution.]] map[chapter:Bayesian formulation and flow posterior sampling topics:[Bayes’ theorem was re-expressed in the flow-matching context to combine a learned prior and a forward-operator-derived likelihood for posterior sampling. Posterior sampling can be performed with the pre-trained FlowFM without any fine-tuning or retraining of the prior.]] map[chapter:Synthetic missing-wedge test and quantitative result topics:[A synthetic missing-wedge posterior sampling test increased average FSC in the masked region from 0 to 0.42.]] map[chapter:Integration into EM iterative refinement and anisotropy correction topics:[Integrating FlowFM into iterative EM reconstruction resolved anisotropic deficits and recovered structural details that standard EM missed.]] map[chapter:Spatially non-uniform noise modeling and multi-dataset results topics:[FlowFM was extended to model spatially varying noise via bandpass/wavelet decomposition across frequency and real-space locations.]] map[chapter:Supervised fine-tuning for map modification and benchmarking topics:[Fine-tuning the prior as a supervised model for map modification produced better FSC0.5 results than EM-READY on the provided dataset. The model and code are open-source and accompanied by two publications and documentation.]] map[chapter:Related work, open source, and future directions topics:[Future work priorities include resolution-aware priors, faster/stabler posterior sampling, richer likelihoods, and autonomous end-to-end workflows.]] map[chapter:Presentation closing and acknowledgements topics:[The presenter concluded the talk and invited questions after thanking their postdoc lab and collaborators.]] map[chapter:Applicability to other nanoparticle systems topics:[Applying the reconstruction approach to large nanoparticles (e.g., 50–150 nm LNPs) should be straightforward for homogeneous samples and more challenging for heterogeneous ones.]] map[chapter:Complex compositions and model generality topics:[The model is trained on diverse EMDB density maps (including complexes with RNA/DNA) and focuses on local features, so it is not restricted to proteins.]] map[chapter:Rationale for density-map reconstruction and structural workflows topics:[Density maps are treated as a necessary validation step for cryo-EM and are part of the established workflow from data to density to atomic model.]] map[chapter:Cryo-ET vs cryo-EM, orientations, and hallucination risk topics:[Cryo-ET provides known orientations and richer 3D information but distributes electron dose across tilts, whereas cryo-EM concentrates dose per particle, creating trade-offs between methods. With extremely limited or single-orientation inputs, the model can produce prior-driven reconstructions that may include hallucinated features, making denoising the primary realistic goal.]] map[chapter:Extensions to medical imaging and inverse problems topics:[The presenter identified forward-process modeling combined with structural priors as a promising approach for ill-posed inverse problems in MRI, CT, ultrasound, and spectroscopy.]] map[chapter:Meeting close, logistics, and administrative questions topics:[]]] content_tease_experiment_variant:<nil> distributor_names_string:Robin D Terwilliger email_access_invite_token:eyJhbGciOiJIUzI1NiIsInR5cCI6IkpXVCJ9.eyJraW5kIjoic2Vzc2lvbl9lbWFpbF9hY2Nlc3NfaW52aXRlIiwiZXhwaXJlc19hdCI6MTgwNDYxOTM2MjQwMywic2Vzc2lvbl9pZCI6IjAxS0s5VlFZRkhRVkpaMVlHOFhXMFZFWEhIIiwiZW1haWwiOiJibWVncmFkc3R1ZGVudHMtbGlzdEBlY24ucHVyZHVlLmVkdSJ9.cTUiqufzjLaHnvJnFzjJHpuBmIKLk40_OZe2yNGtlAs email_domain:<nil> email_verification_token:eyJhbGciOiJIUzI1NiIsInR5cCI6IkpXVCJ9.eyJraW5kIjoiZW1haWxfdmVyaWZpY2F0aW9uIiwiZXhwaXJlc19hdCI6MTgwNDYxOTM2MjQwNCwiZW1haWwiOiJibWVncmFkc3R1ZGVudHMtbGlzdEBlY24ucHVyZHVlLmVkdSJ9.JDKc1Nv3J-MNH5rvvDOLTfHIOrlPhvVyUAExniEQN8g end_time_string:<nil> engagement_score:%!s(float64=64) has_thumbnail:%!s(bool=false) is_meeting_owner:%!s(bool=false) is_read_user:%!s(bool=false) is_subject_line_emoji_experiment:<nil> is_upsell_experiment:<nil> is_upsell_preview:<nil> is_zoom_upsell:<nil> key_questions:[Could the methods used for protein reconstruction be applied to other nanoparticle-based systems such as LNPs or coil-like particles? How large are the particles discussed and what size range would be targeted for application? Can the model reconstruct structures composed of mixed biomolecules (lipids, carbohydrates, proteins) rather than proteins alone? Why reconstruct a density map first instead of going directly from images to atomic models? Does this approach offer advantages over cryo-ET, and would one still use cryo-ET if available? How does the model handle very limited orientations or scarce data, and what is the risk of hallucination? Could this framework be applied to medical imaging modalities (MRI, CT, ultrasound) and spectroscopy inverse problems like SAXS? What are the processes for handling conflicts of interest for co-op internships and working externally during the school year?] late_by_string:0s live_notes_url:<nil> meeting_date_string:March 9, 2026 meeting_platform:teams meeting_title:Yilai Li - Seminar Purdue AI‑Driven Biomolecular Design, Biotech & Computational Biology Position metrics_level:full num_domain_users:<nil> num_recommendations:<nil> on_time:%!s(bool=true) overall_score:%!s(float64=72) raw_meeting_end_time:2026-03-09T19:05:04 raw_meeting_start_time:2026-03-09T18:00:00 raw_session_data_expiration_date:Apr 8 recipient_has_meeting_recap_enabled:%!s(bool=false) remaining_reports:<nil> report_quota:<nil> report_url:https://u25608997.ct.sendgrid.net/ls/click?upn=u001.CrUrehIev9dAOd9CS5LlcqT6... rss_vanity_id:<nil> search_copilot_upsell_content:Could the methods used for protein reconstruction be applied to other nanoparticle-based systems such as LNPs or coil-like particles? search_copilot_upsell_link_placeholder:Could%20the%20methods%20used%20for%20protein%20reconstruction%20be%20applied%20to%20other%20nanoparticle-based%20systems%20such%20as%20LNPs%20or%20coil-like%20particles%3F search_copilot_upsell_variant:sc_upsell_5 send_recap_to_participants:<nil> sentiment_score:%!s(float64=79) session_id:01KK9VQYFHQVJZ1YG8XW0VEXHH sgParameters:map[replyTo:rterwill@purdue.edu templateID:d-27268193e8204d13bd4207d81aabe6c2 templateName:Participant Meeting Recap 3/6/2026 - Revert Action Items Section versionID:066ab4f9-a894-4a56-b6e5-44efc96f9e75] start_time_string:<nil> subject:<nil> summary:The meeting presented Dr. Yilai Li’s seminar on using generative AI to improve 3D reconstruction in cryo-electron microscopy (cryo-EM). Dr. Li reviewed the growth and importance of cryo-EM in structural biology, outlined the experimental workflow (plunge freezing, 2D projection imaging, particle picking, and 3D density reconstruction), and characterized reconstruction as... summary_paragraphs:[The meeting presented Dr. Yilai Li’s seminar on using generative AI to improve 3D reconstruction in cryo-electron microscopy (cryo-EM). Dr. Li reviewed the growth and importance of cryo-EM in structural biology, outlined the experimental workflow (plunge freezing, 2D projection imaging, particle picking, and 3D density reconstruction), and characterized reconstruction as...] talk_time:<nil> talking_pace:<nil> top_talkers:[map[name:Crystal D O'Neal rank:%!s(float64=1) talk_time_percentage:%!s(float64=100) talk_time_string:1h 5m 52s] map[name:Pengfei Wu rank:%!s(float64=2) talk_time_percentage:%!s(float64=0) talk_time_string:5s]] topics:[Dr. Li’s seminar focused on using generative AI for 3D cryo-EM reconstruction. Structural biology relies on experimental techniques like cryo-EM to determine biomolecular structures, which are essential because structure determines function. Cryo-EM output consists of many noisy 2D particle projections that can be used to reconstruct 3D density maps after estimating orientations and translations. A gap exists: limited AI models operate directly in density-map (voxel) space compared with atomic-coordinate models. Reconstruction is an ill-posed inverse problem with unknown rotations, translations, imaging parameters, noise, and potential heterogeneity. Priors and Bayesian approaches can guide reconstruction toward physically plausible solutions. The project will train a foundation model (prior-FM) on ~40,000 public high-resolution cryo-EM densities curated from EMDB. The team selected flow matching as the generative framework and a 3D U-Net architecture for the voxel-space model. Cryo-FM is trained to map a Gaussian distribution to cryo-EM data distribution using flow matching. The model supports unconditional sampling by iteratively following a learned vector field from Gaussian noise to the data distribution. Bayes’ theorem was re-expressed in the flow-matching context to combine a learned prior and a forward-operator-derived likelihood for posterior sampling. Posterior sampling can be performed with the pre-trained FlowFM without any fine-tuning or retraining of the prior. A synthetic missing-wedge posterior sampling test increased average FSC in the masked region from 0 to 0.42. Integrating FlowFM into iterative EM reconstruction resolved anisotropic deficits and recovered structural details that standard EM missed. FlowFM was extended to model spatially varying noise via bandpass/wavelet decomposition across frequency and real-space locations. Fine-tuning the prior as a supervised model for map modification produced better FSC0.5 results than EM-READY on the provided dataset. The model and code are open-source and accompanied by two publications and documentation. Future work priorities include resolution-aware priors, faster/stabler posterior sampling, richer likelihoods, and autonomous end-to-end workflows. The presenter concluded the talk and invited questions after thanking their postdoc lab and collaborators. Applying the reconstruction approach to large nanoparticles (e.g., 50–150 nm LNPs) should be straightforward for homogeneous samples and more challenging for heterogeneous ones. The model is trained on diverse EMDB density maps (including complexes with RNA/DNA) and focuses on local features, so it is not restricted to proteins. Density maps are treated as a necessary validation step for cryo-EM and are part of the established workflow from data to density to atomic model. Cryo-ET provides known orientations and richer 3D information but distributes electron dose across tilts, whereas cryo-EM concentrates dose per particle, creating trade-offs between methods. With extremely limited or single-orientation inputs, the model can produce prior-driven reconstructions that may include hallucinated features, making denoising the primary realistic goal. The presenter identified forward-process modeling combined with structural priors as a promising approach for ill-posed inverse problems in MRI, CT, ultrasound, and spectroscopy.] used_reports:<nil> zoom_reauthentication_required:<nil>] Access this meeting report ( https://u25608997.ct.sendgrid.net/ls/click?upn=u001.CrUrehIev9dAOd9CS5LlcqT6... ) Account creation required. This is your personal link, please do not forward this email to others. View Report ( https://u25608997.ct.sendgrid.net/ls/click?upn=u001.CrUrehIev9dAOd9CS5LlcqT6... ) Recap Chapters & Topics 🔒 Upgrade to view chapters & topics ( https://u25608997.ct.sendgrid.net/ls/click?upn=u001.CrUrehIev9dAOd9CS5LlcqT6... ) Chapters & Topics 🔒 Upgrade to view chapters & topics ( https://u25608997.ct.sendgrid.net/ls/click?upn=u001.CrUrehIev9dAOd9CS5LlcqT6... ) Chapters & Topics • Opening and Speaker Introduction • map[action_items:[] billing_plan:<nil> chapters:[map[chapter:Opening and Speaker Introduction topics:[]] map[chapter:Seminar Topic and High-level Motivation topics:[Dr. Li’s seminar focused on using generative AI for 3D cryo-EM reconstruction. Structural biology relies on experimental techniques like cryo-EM to determine biomolecular structures, which are essential because structure determines function.]] map[chapter:Cryo-EM Impact and Adoption topics:[]] map[chapter:Experimental Workflow and Data Collection topics:[Cryo-EM output consists of many noisy 2D particle projections that can be used to reconstruct 3D density maps after estimating orientations and translations. A gap exists: limited AI models operate directly in density-map (voxel) space compared with atomic-coordinate models.]] map[chapter:Modeling and Unknowns in the Forward Model topics:[Reconstruction is an ill-posed inverse problem with unknown rotations, translations, imaging parameters, noise, and potential heterogeneity.]] map[chapter:Inverse Problems, Priors, and Bayesian Framing topics:[Priors and Bayesian approaches can guide reconstruction toward physically plausible solutions.]] map[chapter:Prior-FM Project, Data, and Model Architecture topics:[The project will train a foundation model (prior-FM) on ~40,000 public high-resolution cryo-EM densities curated from EMDB. The team selected flow matching as the generative framework and a 3D U-Net architecture for the voxel-space model.]] map[chapter:Flow-matching prior and unconditional sampling topics:[Cryo-FM is trained to map a Gaussian distribution to cryo-EM data distribution using flow matching. The model supports unconditional sampling by iteratively following a learned vector field from Gaussian noise to the data distribution.]] map[chapter:Bayesian formulation and flow posterior sampling topics:[Bayes’ theorem was re-expressed in the flow-matching context to combine a learned prior and a forward-operator-derived likelihood for posterior sampling. Posterior sampling can be performed with the pre-trained FlowFM without any fine-tuning or retraining of the prior.]] map[chapter:Synthetic missing-wedge test and quantitative result topics:[A synthetic missing-wedge posterior sampling test increased average FSC in the masked region from 0 to 0.42.]] map[chapter:Integration into EM iterative refinement and anisotropy correction topics:[Integrating FlowFM into iterative EM reconstruction resolved anisotropic deficits and recovered structural details that standard EM missed.]] map[chapter:Spatially non-uniform noise modeling and multi-dataset results topics:[FlowFM was extended to model spatially varying noise via bandpass/wavelet decomposition across frequency and real-space locations.]] map[chapter:Supervised fine-tuning for map modification and benchmarking topics:[Fine-tuning the prior as a supervised model for map modification produced better FSC0.5 results than EM-READY on the provided dataset. The model and code are open-source and accompanied by two publications and documentation.]] map[chapter:Related work, open source, and future directions topics:[Future work priorities include resolution-aware priors, faster/stabler posterior sampling, richer likelihoods, and autonomous end-to-end workflows.]] map[chapter:Presentation closing and acknowledgements topics:[The presenter concluded the talk and invited questions after thanking their postdoc lab and collaborators.]] map[chapter:Applicability to other nanoparticle systems topics:[Applying the reconstruction approach to large nanoparticles (e.g., 50–150 nm LNPs) should be straightforward for homogeneous samples and more challenging for heterogeneous ones.]] map[chapter:Complex compositions and model generality topics:[The model is trained on diverse EMDB density maps (including complexes with RNA/DNA) and focuses on local features, so it is not restricted to proteins.]] map[chapter:Rationale for density-map reconstruction and structural workflows topics:[Density maps are treated as a necessary validation step for cryo-EM and are part of the established workflow from data to density to atomic model.]] map[chapter:Cryo-ET vs cryo-EM, orientations, and hallucination risk topics:[Cryo-ET provides known orientations and richer 3D information but distributes electron dose across tilts, whereas cryo-EM concentrates dose per particle, creating trade-offs between methods. With extremely limited or single-orientation inputs, the model can produce prior-driven reconstructions that may include hallucinated features, making denoising the primary realistic goal.]] map[chapter:Extensions to medical imaging and inverse problems topics:[The presenter identified forward-process modeling combined with structural priors as a promising approach for ill-posed inverse problems in MRI, CT, ultrasound, and spectroscopy.]] map[chapter:Meeting close, logistics, and administrative questions topics:[]]] content_tease_experiment_variant:<nil> distributor_names_string:Robin D Terwilliger email_access_invite_token:eyJhbGciOiJIUzI1NiIsInR5cCI6IkpXVCJ9.eyJraW5kIjoic2Vzc2lvbl9lbWFpbF9hY2Nlc3NfaW52aXRlIiwiZXhwaXJlc19hdCI6MTgwNDYxOTM2MjQwMywic2Vzc2lvbl9pZCI6IjAxS0s5VlFZRkhRVkpaMVlHOFhXMFZFWEhIIiwiZW1haWwiOiJibWVncmFkc3R1ZGVudHMtbGlzdEBlY24ucHVyZHVlLmVkdSJ9.cTUiqufzjLaHnvJnFzjJHpuBmIKLk40_OZe2yNGtlAs email_domain:<nil> email_verification_token:eyJhbGciOiJIUzI1NiIsInR5cCI6IkpXVCJ9.eyJraW5kIjoiZW1haWxfdmVyaWZpY2F0aW9uIiwiZXhwaXJlc19hdCI6MTgwNDYxOTM2MjQwNCwiZW1haWwiOiJibWVncmFkc3R1ZGVudHMtbGlzdEBlY24ucHVyZHVlLmVkdSJ9.JDKc1Nv3J-MNH5rvvDOLTfHIOrlPhvVyUAExniEQN8g end_time_string:<nil> engagement_score:%!s(float64=64) has_thumbnail:%!s(bool=false) is_meeting_owner:%!s(bool=false) is_read_user:%!s(bool=false) is_subject_line_emoji_experiment:<nil> is_upsell_experiment:<nil> is_upsell_preview:<nil> is_zoom_upsell:<nil> key_questions:[Could the methods used for protein reconstruction be applied to other nanoparticle-based systems such as LNPs or coil-like particles? How large are the particles discussed and what size range would be targeted for application? Can the model reconstruct structures composed of mixed biomolecules (lipids, carbohydrates, proteins) rather than proteins alone? Why reconstruct a density map first instead of going directly from images to atomic models? Does this approach offer advantages over cryo-ET, and would one still use cryo-ET if available? How does the model handle very limited orientations or scarce data, and what is the risk of hallucination? Could this framework be applied to medical imaging modalities (MRI, CT, ultrasound) and spectroscopy inverse problems like SAXS? What are the processes for handling conflicts of interest for co-op internships and working externally during the school year?] late_by_string:0s live_notes_url:<nil> meeting_date_string:March 9, 2026 meeting_platform:teams meeting_title:Yilai Li - Seminar Purdue AI‑Driven Biomolecular Design, Biotech & Computational Biology Position metrics_level:full num_domain_users:<nil> num_recommendations:<nil> on_time:%!s(bool=true) overall_score:%!s(float64=72) raw_meeting_end_time:2026-03-09T19:05:04 raw_meeting_start_time:2026-03-09T18:00:00 raw_session_data_expiration_date:Apr 8 recipient_has_meeting_recap_enabled:%!s(bool=false) remaining_reports:<nil> report_quota:<nil> report_url:https://u25608997.ct.sendgrid.net/ls/click?upn=u001.CrUrehIev9dAOd9CS5LlcqT6... rss_vanity_id:<nil> search_copilot_upsell_content:Could the methods used for protein reconstruction be applied to other nanoparticle-based systems such as LNPs or coil-like particles? search_copilot_upsell_link_placeholder:Could%20the%20methods%20used%20for%20protein%20reconstruction%20be%20applied%20to%20other%20nanoparticle-based%20systems%20such%20as%20LNPs%20or%20coil-like%20particles%3F search_copilot_upsell_variant:sc_upsell_5 send_recap_to_participants:<nil> sentiment_score:%!s(float64=79) session_id:01KK9VQYFHQVJZ1YG8XW0VEXHH sgParameters:map[replyTo:rterwill@purdue.edu templateID:d-27268193e8204d13bd4207d81aabe6c2 templateName:Participant Meeting Recap 3/6/2026 - Revert Action Items Section versionID:066ab4f9-a894-4a56-b6e5-44efc96f9e75] start_time_string:<nil> subject:<nil> summary:The meeting presented Dr. Yilai Li’s seminar on using generative AI to improve 3D reconstruction in cryo-electron microscopy (cryo-EM). Dr. Li reviewed the growth and importance of cryo-EM in structural biology, outlined the experimental workflow (plunge freezing, 2D projection imaging, particle picking, and 3D density reconstruction), and characterized reconstruction as... summary_paragraphs:[The meeting presented Dr. Yilai Li’s seminar on using generative AI to improve 3D reconstruction in cryo-electron microscopy (cryo-EM). Dr. Li reviewed the growth and importance of cryo-EM in structural biology, outlined the experimental workflow (plunge freezing, 2D projection imaging, particle picking, and 3D density reconstruction), and characterized reconstruction as...] talk_time:<nil> talking_pace:<nil> top_talkers:[map[name:Crystal D O'Neal rank:%!s(float64=1) talk_time_percentage:%!s(float64=100) talk_time_string:1h 5m 52s] map[name:Pengfei Wu rank:%!s(float64=2) talk_time_percentage:%!s(float64=0) talk_time_string:5s]] topics:[Dr. Li’s seminar focused on using generative AI for 3D cryo-EM reconstruction. Structural biology relies on experimental techniques like cryo-EM to determine biomolecular structures, which are essential because structure determines function. Cryo-EM output consists of many noisy 2D particle projections that can be used to reconstruct 3D density maps after estimating orientations and translations. A gap exists: limited AI models operate directly in density-map (voxel) space compared with atomic-coordinate models. Reconstruction is an ill-posed inverse problem with unknown rotations, translations, imaging parameters, noise, and potential heterogeneity. Priors and Bayesian approaches can guide reconstruction toward physically plausible solutions. The project will train a foundation model (prior-FM) on ~40,000 public high-resolution cryo-EM densities curated from EMDB. The team selected flow matching as the generative framework and a 3D U-Net architecture for the voxel-space model. Cryo-FM is trained to map a Gaussian distribution to cryo-EM data distribution using flow matching. The model supports unconditional sampling by iteratively following a learned vector field from Gaussian noise to the data distribution. Bayes’ theorem was re-expressed in the flow-matching context to combine a learned prior and a forward-operator-derived likelihood for posterior sampling. Posterior sampling can be performed with the pre-trained FlowFM without any fine-tuning or retraining of the prior. A synthetic missing-wedge posterior sampling test increased average FSC in the masked region from 0 to 0.42. Integrating FlowFM into iterative EM reconstruction resolved anisotropic deficits and recovered structural details that standard EM missed. FlowFM was extended to model spatially varying noise via bandpass/wavelet decomposition across frequency and real-space locations. Fine-tuning the prior as a supervised model for map modification produced better FSC0.5 results than EM-READY on the provided dataset. The model and code are open-source and accompanied by two publications and documentation. Future work priorities include resolution-aware priors, faster/stabler posterior sampling, richer likelihoods, and autonomous end-to-end workflows. The presenter concluded the talk and invited questions after thanking their postdoc lab and collaborators. Applying the reconstruction approach to large nanoparticles (e.g., 50–150 nm LNPs) should be straightforward for homogeneous samples and more challenging for heterogeneous ones. The model is trained on diverse EMDB density maps (including complexes with RNA/DNA) and focuses on local features, so it is not restricted to proteins. Density maps are treated as a necessary validation step for cryo-EM and are part of the established workflow from data to density to atomic model. Cryo-ET provides known orientations and richer 3D information but distributes electron dose across tilts, whereas cryo-EM concentrates dose per particle, creating trade-offs between methods. With extremely limited or single-orientation inputs, the model can produce prior-driven reconstructions that may include hallucinated features, making denoising the primary realistic goal. The presenter identified forward-process modeling combined with structural priors as a promising approach for ill-posed inverse problems in MRI, CT, ultrasound, and spectroscopy.] used_reports:<nil> zoom_reauthentication_required:<nil>] See all 21 chapters → ( https://u25608997.ct.sendgrid.net/ls/click?upn=u001.CrUrehIev9dAOd9CS5LlcqT6... ) Action Items No action items generated Key Questions • Could the methods used for protein reconstruction be applied to other nanoparticle-based systems such as LNPs or coil-like particles? • How large are the particles discussed and what size range would be targeted for application? See all 8 key questions → ( https://u25608997.ct.sendgrid.net/ls/click?upn=u001.CrUrehIev9dAOd9CS5LlcqT6... ) Review notes ( https://u25608997.ct.sendgrid.net/ls/click?upn=u001.CrUrehIev9dAOd9CS5LlcqT6... ) Work smarter, not harder Read makes meetings more effective and efficient with meeting recaps, transcripts, playback, personalized coaching, meeting recommendations, and more. Get started for free → ( https://u25608997.ct.sendgrid.net/ls/click?upn=u001.CrUrehIev9dAOd9CS5LlcqT6... ) Need help? Visit the Support Center ( https://u25608997.ct.sendgrid.net/ls/click?upn=u001.CrUrehIev9dAOd9CS5Llcqxn... ) or contact us ( support@read.ai ). Unsubscribe ( https://u25608997.ct.sendgrid.net/asm/unsubscribe/?user_id=25608997&data=Am6... ) • Email Preferences ( https://u25608997.ct.sendgrid.net/ls/click?upn=u001.CrUrehIev9dAOd9CS5LlcqT6... ) Team at Read AI • 999 3rd Ave, Suite 3300, Seattle, WA 98104