What are the key takeaways from this Beyond Biotech episode?

Key insights include: **Shared Interaction Space Architecture:** Prophet converts drug discovery into a search problem by embedding proteins and small molecules into a single shared mathematical space — similar to a Google search engine — enabling rapid scoring of billions of molecules simultaneously. This approach bypasses the need for solved protein structures, making it viable for data-scarce, hard-to-drug targets.; **Sequence-Only Training Strategy:** Unlike most AI drug discovery tools that require 3D protein geometry or large complex datasets, Prophet trains exclusively on protein sequences. This design choice dramatically expands the range of targetable proteins and reduces dependency on experimentally derived structural data, which is scarce for many disease-relevant targets.; **Complementary Positioning vs. Existing AI Tools:** Current AI in small molecule discovery largely replicates physics-based simulations to model precise binding interactions — a slow, data-heavy process. Prophet operates upstream, rapidly narrowing the candidate pool before high-resolution methods are applied. Pharma teams can use Prophet to define where to look, then apply detailed simulation tools afterward.

What did Avital Sherry Ivry discuss on Beyond Biotech?

Avital Sharir, cofounder and CSO of Prophet, an Israeli AI drug discovery startup launched in late 2024 from ION Labs, explains how Prophet maps proteins and small molecules into a shared mathematical space to screen billions of molecules at scale, targeting previously undruggable proteins without requiring solved 3D structures. Key topics include: **Shared Interaction Space Architecture:** Prophet converts drug discovery into a search problem by embedding proteins and small molecules into a single shared mathematical space — similar to a Google search engine — enabling rapid scoring of billions of molecules simultaneously. This approach bypasses the need for solved protein structures, making it viable for data-scarce, hard-to-drug targets.; **Sequence-Only Training Strategy:** Unlike most AI drug discovery tools that require 3D protein geometry or large complex datasets, Prophet trains exclusively on protein sequences. This design choice dramatically expands the range of targetable proteins and reduces dependency on experimentally derived structural data, which is scarce for many disease-relevant targets..

How long is this episode of Beyond Biotech?

This episode is 32 minutes long. SignalCast provides an AI-generated summary so you can get the key insights in about 3 minutes.

Beyond Biotech

Using AI to crack undruggable drug targets

February 13, 2026

32 min episode · 2 min read

Avital Sherry Ivry

Episode

32 min

Read time

2 min

Topics

Relationships, Startups, Fundraising & VC

AI-Generated Summary

Published Mar 14, 2026

Key Takeaways

✓Shared Interaction Space Architecture: Prophet converts drug discovery into a search problem by embedding proteins and small molecules into a single shared mathematical space — similar to a Google search engine — enabling rapid scoring of billions of molecules simultaneously. This approach bypasses the need for solved protein structures, making it viable for data-scarce, hard-to-drug targets.
✓Sequence-Only Training Strategy: Unlike most AI drug discovery tools that require 3D protein geometry or large complex datasets, Prophet trains exclusively on protein sequences. This design choice dramatically expands the range of targetable proteins and reduces dependency on experimentally derived structural data, which is scarce for many disease-relevant targets.
✓Complementary Positioning vs. Existing AI Tools: Current AI in small molecule discovery largely replicates physics-based simulations to model precise binding interactions — a slow, data-heavy process. Prophet operates upstream, rapidly narrowing the candidate pool before high-resolution methods are applied. Pharma teams can use Prophet to define where to look, then apply detailed simulation tools afterward.
✓Pharma-First Development Model: Prophet validates models directly against data from Merck and AstraZeneca, both embedded partners through ION Labs' venture studio structure. This prevents building in a vacuum — a common failure mode for deep-tech startups. Teams building AI tools for regulated industries should prioritize real-world data access and continuous experimental feedback loops from day one.
✓Library Size Directly Correlates with Hit Rate: A 2024 paper cited by Sharir demonstrates that screening larger small molecule libraries consistently produces higher hit rates — meaning current drug discovery not only samples a fraction of chemical space, but likely samples a suboptimal fraction. AI-enabled scale screening is therefore not just faster but structurally more likely to find viable candidates.

What It Covers

Avital Sharir, cofounder and CSO of Prophet, an Israeli AI drug discovery startup launched in late 2024 from ION Labs, explains how Prophet maps proteins and small molecules into a shared mathematical space to screen billions of molecules at scale, targeting previously undruggable proteins without requiring solved 3D structures.

Key Questions Answered

•Shared Interaction Space Architecture: Prophet converts drug discovery into a search problem by embedding proteins and small molecules into a single shared mathematical space — similar to a Google search engine — enabling rapid scoring of billions of molecules simultaneously. This approach bypasses the need for solved protein structures, making it viable for data-scarce, hard-to-drug targets.
•Sequence-Only Training Strategy: Unlike most AI drug discovery tools that require 3D protein geometry or large complex datasets, Prophet trains exclusively on protein sequences. This design choice dramatically expands the range of targetable proteins and reduces dependency on experimentally derived structural data, which is scarce for many disease-relevant targets.
•Complementary Positioning vs. Existing AI Tools: Current AI in small molecule discovery largely replicates physics-based simulations to model precise binding interactions — a slow, data-heavy process. Prophet operates upstream, rapidly narrowing the candidate pool before high-resolution methods are applied. Pharma teams can use Prophet to define where to look, then apply detailed simulation tools afterward.
•Pharma-First Development Model: Prophet validates models directly against data from Merck and AstraZeneca, both embedded partners through ION Labs' venture studio structure. This prevents building in a vacuum — a common failure mode for deep-tech startups. Teams building AI tools for regulated industries should prioritize real-world data access and continuous experimental feedback loops from day one.
•Library Size Directly Correlates with Hit Rate: A 2024 paper cited by Sharir demonstrates that screening larger small molecule libraries consistently produces higher hit rates — meaning current drug discovery not only samples a fraction of chemical space, but likely samples a suboptimal fraction. AI-enabled scale screening is therefore not just faster but structurally more likely to find viable candidates.

Notable Moment

Sharir reframes the term "undruggable" entirely — arguing it simply reflects the limits of current tools and explored chemical space, not biological impossibility. Some targets previously deemed unreachable may have viable small molecule solutions that existing screening methods, constrained by scale, have never encountered.

Know someone who'd find this useful?