ConforNets: Latents-Based Conformational Control in OpenFold3

์ €์ž: | ๋‚ ์งœ: 2026-04-20 | URL: https://arxiv.org/abs/2604.18559 📄 PDF


Essence

Figure 1

Figure 1. ConforNets induce conformations by mixing and biasing the channels of AF3โ€™s pair latents to satisfy different

ConforNets๋Š” AlphaFold3์˜ Pairformer ์ด์ „ pair latent์— ์ฑ„๋„๋ณ„ affine ๋ณ€ํ™˜์„ ์ ์šฉํ•˜์—ฌ ๋‹จ๋ฐฑ์งˆ์˜ ๋Œ€์•ˆ conformation์„ ์ „์—ญ์ ์œผ๋กœ ์ œ์–ดํ•˜๋Š” ๋ฐฉ๋ฒ•์„ ์ œ์•ˆํ•œ๋‹ค. ์ด๋Š” unsupervised๋กœ ๋‹ค์–‘ํ•œ conformational state๋ฅผ ์ƒ์„ฑํ•˜๊ณ , supervised๋กœ ๋‹จ๋ฐฑ์งˆ ํŒจ๋ฐ€๋ฆฌ ๊ฐ„ conformational transfer๋ฅผ ๊ฐ€๋Šฅํ•˜๊ฒŒ ํ•œ๋‹ค.

Motivation

Achievement

Figure 1

Figure 1. ConforNets induce conformations by mixing and biasing the channels of AF3โ€™s pair latents to satisfy different

Unsupervised conformational generation: 104๊ฐœ ๋‹จ๋ฐฑ์งˆ ์Œ์˜ multi-state benchmark์—์„œ ๋ชจ๋“  ๊ธฐ์กด ๋ฐฉ๋ฒ•์„ ๋Šฅ๊ฐ€ํ•˜๋Š” ์„ฑ๋Šฅ ๋‹ฌ์„ฑ. 200-residue ๋‹จ๋ฐฑ์งˆ์— ๋Œ€ํ•ด 40 GPU ์ดˆ ์ด๋‚ด์˜ ๋น ๋ฅธ ์ตœ์ ํ™” ์†๋„ ๊ตฌํ˜„. Supervised conformational transfer: GPCR active state (24โ†’79%), kinase DFG-out (6โ†’23%), transporter outward-open (16โ†’57%) ๋“ฑ์—์„œ ํ‘œ์ค€ AF3 ๋Œ€๋น„ ํ˜„์ €ํ•œ ์„ฑ๋Šฅ ํ–ฅ์ƒ ๋‹ฌ์„ฑ. Architecture-agnostic design: Channel-wise transform์œผ๋กœ protein-specific ์ œ์•ฝ์„ ์ œ๊ฑฐํ•˜์—ฌ ๋‹ค์–‘ํ•œ ๋‹จ๋ฐฑ์งˆ๊ณผ AF3 ๊ธฐ๋ฐ˜ ๋ชจ๋ธ์— ์žฌ์‚ฌ์šฉ ๊ฐ€๋Šฅํ•˜๊ฒŒ ์„ค๊ณ„.

How

Figure 1

Figure 1. ConforNets induce conformations by mixing and biasing the channels of AF3โ€™s pair latents to satisfy different

Originality

Limitation & Further Study

Evaluation

Novelty: 4/5 Technical Soundness: 4/5 Significance: 4/5 Clarity: 4/5 Overall: 4/5

์ดํ‰: ConforNets๋Š” AlphaFold3 ๊ธฐ๋ฐ˜ ๋ชจ๋ธ์—์„œ conformational control์„ ๋‹ฌ์„ฑํ•˜๊ธฐ ์œ„ํ•œ ์‹ค์šฉ์ ์ด๊ณ  ํšจ์œจ์ ์ธ ๋ฐฉ๋ฒ•์„ ์ œ์‹œํ•œ๋‹ค. Pre-Pairformer latent์— ๋Œ€ํ•œ ์ฒด๊ณ„์ ์ธ ๋ถ„์„, channel-wise perturbation์˜ novelํ•œ ๋„์ž…, ๊ทธ๋ฆฌ๊ณ  conformational transfer๋ผ๋Š” ์ƒˆ๋กœ์šด supervised task์˜ ์ œ์‹œ๋Š” ๊ตฌ์กฐ ์ƒ๋ฌผํ•™ ๋ฐ ์•ฝ๋ฌผ ์„ค๊ณ„ ๋ถ„์•ผ์—์„œ ์˜๋ฏธ ์žˆ๋Š” ๊ธฐ์—ฌ๋ฅผ ํ•œ๋‹ค. ๋‹ค๋งŒ ๋ฌผ๋ฆฌ์  calibration ๋ถ€์žฌ์™€ ํ‰๊ฐ€ ๋ฒ”์œ„์˜ ์ œ์•ฝ์ด ์žˆ๋‹ค.

๊ฐ™์ด ๋ณด๋ฉด ์ข‹์€ ๋…ผ๋ฌธ

๊ธฐ๋ฐ˜ ์—ฐ๊ตฌ
์ˆœํ™˜์  ์ปจํฌ๋ฉ”์ด์…˜ ์ƒ˜ํ”Œ๋ง ๋ฐ ์ œ๋„ˆ๋ ˆ์ดํ‹ฐ๋ธŒ ๋ชจ๋ธ์˜ ์žฅ๊ธฐ ๋™์—ญํ•™ ์˜ˆ์ธก ๋Šฅ๋ ฅ ๊ด€๋ จ ์ตœ์‹  ๋ฐฉ๋ฒ•๋ก ์  ๋ฐฐ๊ฒฝ์„ ์ œ๊ณตํ•ฉ๋‹ˆ๋‹ค.
๊ธฐ๋ฐ˜ ์—ฐ๊ตฌ
AlphaFold ๊ตฌ์กฐ DB ๋ฐ ์ฟผํ„ฐ๋„ˆ๋ฆฌ ๋ชจ๋ธ์˜ ์„ค๋ช…๋ ฅ์„ ๋…ผ์˜ํ•˜๋Š” ๋…ผ๋ฌธ์ด ConforNets์˜ conformation ์ œ์–ด ๊ฐœ๋…์— ์ด๋ก ์  ํ† ๋Œ€๋ฅผ ์ œ๊ณตํ•œ๋‹ค.
๋‹ค๋ฅธ ์ ‘๊ทผ
์œ ๊ธฐ ๊ฒฐ์ •๊ตฌ์กฐ ์ƒ์„ฑ๊ณผ ๊ฐ™์ด ๋‹จ๋ฐฑ์งˆ ๊ตฌ์กฐ ๋Œ€์ฒด ๊ฒฝ๋กœ ์ƒ์„ฑ์„ ์œ„ํ•œ ์ƒ์„ฑ ๊ธฐ๋ฐ˜ ์ ‘๊ทผ์„ ๋‹ค๋ฅด๊ฒŒ ์ œ์•ˆํ•œ๋‹ค.
๋‹ค๋ฅธ ์ ‘๊ทผ
Zero-Shot ๋ฐฉ์‹์œผ๋กœ ๋‹จ๋ฐฑ์งˆ conformation ensemble์„ ์ƒ์„ฑํ•˜๋Š” ๋”ฅ๋Ÿฌ๋‹ ์ ‘๊ทผ๋ฒ• ๋น„๊ต๋กœ, ConforNets์˜ latent ์ œ์–ด ๋ฐฉ์‹๊ณผ zero-shot ๋ฐฉ๋ฒ•๋ก ์˜ ์ฐจ์ด๋ฅผ ๋ถ„์„ํ•  ์ˆ˜ ์žˆ๋‹ค.
ํ›„์† ์—ฐ๊ตฌ
ConforNets ๋…ผ๋ฌธ์€ ์˜คํ”ˆํด๋“œ latent ๊ธฐ๋ฐ˜ ๊ตฌ์กฐ ์ œ์–ด๋ผ๋Š” ์ ์—์„œ, AlphaFold latent flooding์˜ ์‹ค์ œ์  ์ ์šฉ ํ™•์žฅ์„ฑ์„ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค.
← ๋ชฉ๋ก์œผ๋กœ ๋Œ์•„๊ฐ€๊ธฐ

๐ŸŽง Audio Overview

์ด ๋…ผ๋ฌธ ๋ฆฌ๋ทฐ๋ฅผ ํŒŸ์บ์ŠคํŠธํ˜• ์˜ค๋””์˜ค๋กœ ์ƒ์„ฑํ•ฉ๋‹ˆ๋‹ค. (Gemini ยท ํ‚ค๋Š” ๋ธŒ๋ผ์šฐ์ €์—๋งŒ ์ €์žฅ ยท ์™„์„ฑ๋ณธ์€ ์ด๋ฉ”์ผ๋กœ๋„ ์ „์†ก)
โ–ธ ๊ณ ๊ธ‰: ๊ตฌ์„ฑ ๋ฐฉํ–ฅ(๋Œ€๋ณธ ์ž‘์„ฑ ์ง€์นจ) ์ง์ ‘ ์ˆ˜์ •